2026-11-16
Energy Monitoring and Sustainability Reporting for Factories on Azure

Show notes
You cannot reduce what you cannot measure. In this episode of the iFAKT series, we look at energy monitoring and sustainability reporting in manufacturing: metering data, IoT ingestion, allocating energy to products and orders, and linking it to production data with Power BI and Microsoft Fabric. You will learn how to build energy reporting that stands up to reporting requirements and actually helps plants save cost. Who it is for: Plant managers, sustainability leads, data engineers and Azure architects in manufacturing. Become a supporter of this podcast: https://www.spreaker.com/podcast/m365-fm-microsoft-365-azure-cloud-power-platform-ai--6704921/support. This episode includes AI-generated content.
Host: Mirko Peters
Related articles
- Why Point-to-Point Manufacturing Integrations Fail (And How to Fix Them)
- How to Perform a Comprehensive Power Platform Governance Assessment
- What Is Environmental, Social, and Governance (ESG)? Simple Explanation and Practical Guide for Microsoft Admins
- What Is Environmental, Social, and Governance (ESG) Meaning and How Does It Apply to Microsoft-Focused Organizations?
- What Is Social Governance and How Does It Shape ESG in Business and IT?
- What Is Environmental, Social, and Governance? A Practical Guide for IT Leaders and Admins
- Industrial IoT Message Routing: Avoiding the Trap of Event-Driven Overload
- Why Treating High-Volume Industrial Telemetry as Event Grid Messages Breaks Cloud Architectures
Transcript
Read the full transcript
iFAKT - Energy Monitoring and Sustainability Reporting for Factories on Azure 00:00:00 Speaker: Your energy bill tells you what the factory used. But when energy spending rises and a customer asks about emissions for their order, you've got meter totals and production records with no trusted link between them. So which parts can you measure and which need assumptions? That's the question we'll work through. Tracing energy from the meter to the production order, then into a report you can defend when someone asks how you calculated it, I'd start with the decision, not the data platform. Before you connect anything. Ask who needs this information, what they're going to change because of it and how quickly they need an answer they can trust. Understand our energy use sounds reasonable, but it's too broad a maintenance lead looking for equipment that draws power overnight needs a different answer from a finance controller. Checking whether a product still earns its margin, and a customer asking for emissions evidence creates another task entirely. You might use some of the same source records, but the detail review process and people responsible for approving the answer can differ quite a lot. Take idle load as the first example. If your question is whether an area keeps drawing power after production stops, you need information soon enough for someone to investigate while the conditions still exist. Speed matters here, within reason. A monthly total might confirm there's a problem, but it won't help the shift team check what's running during the quiet period. They're responsible for. The owner might be facilities maintenance or the production team. I'd agree that responsibility before building the report because an alert without someone able to investigate it usually becomes another message that everyone learns to ignore. Now consider product cost. Finance might need a stable view across a closed accounting period, with enough production detail to compare products without confusing a busy month with an efficient one. Here, immediate updates might matter less than repeatable results. You need an agreed definition of the cost you're checking. A clear period and someone responsible for resolving differences before people use the result to change prices. That changes what good enough means. A quick estimate could help identify which products deserve attention. While a number used in a commercial negotiation needs a stronger basis in a clear review. Customer emissions requests bring a different kind of pressure. The person receiving the request needs to know exactly what the customer has asked for, what evidence they expect, and who inside your company can approve the response? An answer can arrive quickly and still miss the request. I'd rather clarify the question early than spend weeks building a calculation that the customer can't use, because it answers something slightly different from what they asked. So write down the decision in plain words. We want the maintenance team to investigate electricity use outside. Scheduled production in this area gives the project a much firmer starting point than we want. Sustainability insights. Then choose a small enough boundary to test that decision. Pick one production area, one energy source, and one business question so you can check whether the information actually helps before extending the work across the plant. That's a test boundary, not the limit of your ambition. Starting with electricity in one area lets you discover what the chosen question demands without asking every department to settle every possible energy question at once. Agree on success before you build. If the result helps the named owner take the intended action at the right speed and with enough evidence, you've learned something you can use for the next step, we'll need a shared example to test those choices. Let's use a hypothetical factory throughout this discussion. So each choice about detailed timing and ownership has to answer a production question rather than just sound sensible in an architecture meeting. Picture a plant that produces metal parts for industrial equipment. Its machining cells cut the parts. A heat treatment furnace handles batches from those cells, and a shared compressed air system supplies equipment across the production area. We'll keep this factory hypothetical. The point is to follow a familiar production setting, where several teams can describe the same working week correctly and still disagree about whether energy performance improved. The plant runs two product families through much of the same equipment. One family consists of small parts with short machining cycles, while the other needs longer cycles and usually moves through production in smaller batches. Their production patterns differ even when both families keep people busy. A shift can finish plenty of small parts, while another shift spends most of its time processing fewer large parts through the same machining cells. The furnace brings those differences into the same weekly discussion. Depending on which orders arrive from machining, the team might process batches with different part counts, even though the furnace remains part of the same production route. Compressed air doesn't belong neatly to either product. Family. Both families use equipment connected to the shared supply, so the factory's electricity total includes a service that supports several activities happening at once. Now put yourself in the weekly production meeting facilities brings the utility bill and the site consumption records, while the planner brings the orders completed, the work still waiting, and the commitments for the next week. Finance brings a cost view of those same operations. The controller wants to understand the energy expense carried by the products the plant cells, especially when the balance between the two product families changes. Nobody needs to misunderstand the factory for this meeting to become difficult. Each person starts from records that support their own job, but those records describe different slices of what the plant has been doing. Say the plant uses less electricity this month than last month. Facilities can point to that reduction, while production records show fewer finished parts and a larger share of orders from the product family. With longer machining cycles, electricity per good part could rise. At the same time, the total can fall, while the energy used for each accepted part increases because the number of good parts falls faster than electricity consumption. Both statements can be correct. Facilities reports that the site used less electricity, while the production team reports that each good part carried more electricity than it did in the previous month. That leaves the controller with a question. Neither statement settles did the factory run less efficiently, or did it simply spend more time producing parts that normally require longer cycles and arrive in smaller batches? The planner has another concern if the month included more of the slower product family, comparing the combined part count with last month's total might tell you very little about how well production performed. Even a good month needs some explanation here. Facilities might mean lower total consumption. Finance might mean lower energy costs for a product, and production might mean completing the required order mix without adding delays. I'd keep those claims separate during the discussion. Otherwise, the meeting can turn into an argument about whose number deserves trust. When the real disagreement concerns what each number actually describes and which comparison makes sense. The factory doesn't need everyone to use one number. It needs people to name the measure, the period and the production activity behind it before they treat a change as evidence that something improved or deteriorated. That starts with the words on the report. Before we ask whether energy performance improved, we need to agree on what each energy number means, because a lower side total and a higher figure per good part can describe the same month. Power tells you how fast equipment is using energy when the furnace draws electricity. Its power reading describes the rate of use at that moment, rather than the electricity consumed across the whole heating cycle. You'll usually hear electrical power expressed in kilowatts. Energy expressed in kilowatt hours adds time to that picture, so a steady load running for longer consumes more energy, even when its power reading doesn't change. Here's a simple example with deliberately steady loads, a machine drawing ten kilowatts for two hours uses twenty kilowatt hours. Assuming that power stays constant throughout the period rather than rising and falling during production. Actual equipment rarely holds that steady throughout a job with a furnace, the power draw can rise during heating and fall when the controlled cycle. So energy use comes from adding up consumption over time. A single power reading can't describe that whole cycle. If someone captures a reading during the hottest part of the process and treats it as typical of the entire run, the resulting energy estimate can overstate consumption. The distinction changes how you interpret a sudden spike. A brief rise in power might add little to total energy use, yet still deserve attention. If it affects the demand measure used in your electricity contract. Demand charges need their own explanation. Some tariffs include a charge based on the highest average power recorded across a defined interval, rather than charging only for the total kilowatt hours consumed during the billing period. The contract defines the interval and the charging rule. A very brief spike might contribute little to that average, while several large loads running together for longer could push the build demand higher for that period. Don't assume every peak carries the same financial consequence. You need to know whether the charge depends on an interval average when that interval falls, and whether the contract includes other conditions that affect build demand for our furnace. Timing can change cost without changing energy. If its heating cycle overlaps with other large loads, that overlap could increase build demand under the applicable tariff. Even though the furnace consumes the same kilowatt hours, cost needs more than an energy total. Your electricity bill can combine consumption charges with demand charges, fixed fees and other contract terms, so multiplying kilowatt hours by one rate might answer only part of the question. That simple calculation can still have a clear purpose. If you're estimating the consumption charge for a defined period, using the applicable energy rate might be suitable, provided you name the costs. That calculation leaves out time dependent prices. Add another consideration where your contract charges different rates at different times. Identical electricity consumption can carry different costs depending on when the equipment runs and which rate applies during that period. Lower cost doesn't automatically mean lower emissions. A cheaper operating period could have a higher emissions factor than a more expensive period, depending on the electricity supply and the accounting method used for the calculation. Carbon introduces a different kind of calculation for an energy related emissions estimate. You take an activity quantity such as electricity consumed, and multiply it by an emissions factor that matches the activity and the intended reporting basis. That factor describes emissions per unit of activity. The electricity meter supplies an energy quantity, while the factor supplies the emissions relationship. So the carbon result depends on information beyond the meter itself and its readings. Keep the units consistent before you multiply. If consumption uses kilowatt hours and the factor uses emissions per megawatt hour, you need the unit conversion or the calculation can produce a badly wrong result despite using valid inputs. I'd label these measures separately from the start a power peak, an energy total, an electricity expense, and an emissions estimate. Each answer a different question even when they refer to the same furnace run. Their changes don't have to move together. A change in the electricity price can alter cost without changing consumption, while a change in the applicable emissions factor can alter calculated emissions without changing the machine's operation. That gives the project a more precise starting point before choosing meters or building reports, decide whether the question calls for power over a billing interval, energy across a process, contract based cost, or calculated emissions. The choice determines what information you'll need. Buying a faster meter won't explain a tariff, and adding more decimal places to consumption won't resolve a carbon calculation that uses the wrong emissions basis. There's still a question to settle what belongs inside the calculation. You can choose the right measure and still produce a misleading result. If the calculation includes equipment, buildings or activities that don't belong to the question you're answering. I'd write the boundary down in plain language. For a hypothetical plant, that could mean electricity entering the machining building, rather than every energy source used anywhere on the site during the reporting period. A physical boundary describes the area or equipment you're counting. It might follow the side fence one building a production line, a single machine, or a shared utility system that supplies several parts of the factory. Those boundaries don't always follow department names. The machining department might occupy one building while its compressed air comes from another. So the department's physical location doesn't describe every energy input needed to run its equipment. Consider the furnace within our example. Counting energy entering the furnace gives you a different boundary from counting the heat treatment area, which might also include ventilation and other equipment that supports the process. Neither boundary wins by default. The right choice depends on whether you're assessing the furnace itself or the wider production activity, because the supporting equipment belongs in one question, but might sit outside the other. Shared utilities make this harder. If you define the boundary as the compressed air system, you're following a service across the plant, rather than following the walls around a building or the machines assigned to one supervisor. Then there's the reporting boundary. This describes which operations and activities belong in the reported result, and it can differ from the physical area where someone collects the energy readings for the project. A company inventory asks about the organization's emissions. Its boundary follows the chosen accounting approach for the operations the company owns or controls, which needs agreement beyond the production team responsible for a particular factory A production process study follows a different question. You might examine the energy associated with heat treatment. While a product footprint follows the activities included for the product, rather than stopping automatically at the factory gate. Don't let a convenient data source decide that boundary. If the intended study covers a process that depends on several energy supplies, a complete electricity feed still leaves the other supplies outside your records until you address them. Electricity might only be part of the picture. A plant could burn natural gas in a furnace by steam from a neighboring facility, or receive heating and cooling through services that don't appear on its electricity account. List those energy inputs before you count them. For each input, establish whether it enters the boundary from outside or comes from equipment already inside it, because that distinction affects how the totals fit together. Internal conversion deserves particular care. If a boiler inside your boundary burns gas to produce steam. Counting both the gas input and that steam as separate external energy inputs would count the same supply twice. Emissions reporting adds another set of boundaries under the greenhouse gas protocol. Scope one covers direct emissions from sources the organization owns or controls, such as fuel burned in a boiler within its organizational boundary. Scope one extends beyond energy use. Depending on the operation. Direct emissions can also include releases from industrial processes or leaking refrigerants, so an energy project alone won't necessarily capture the entire category. Scope two covers emissions from generating purchased or acquired energy that includes electricity, steam, heat, and cooling consumed by the organization, with the emissions generally occurring where that purchased energy gets generated rather than where the factory uses it. Scope three covers other value chain emissions outside those direct and purchased energy categories. Depending on the reporting boundary, these can include emissions associated with purchased materials and transport, among other upstream and downstream activities connected to the business. Keep the reporting perspective attached to those labels. The same physical emissions can fall into different scopes for different companies because the categories describe each reporting organization's relationship to the activity that causes them. I'd settle this with operations and the reporting owner together. Write down what's included, what's excluded, which energy supplies cross the boundary, and whether the result describes a site, a process, or an organization's inventory that gives the meter plan a clear job instead of collecting whatever readings happen to be available. You can ask where measurement needs to follow the agreed boundary and where that boundary still has gaps. Start by finding out what the existing meters actually cover before buying anything. Ask the electrical team to trace the supplies and confirm which loads sit behind each meter, including equipment that the circuit labels might no longer describe correctly. The main supply meter gives you the broadest measurement point where it covers the site's incoming electricity. It provides a site total, but that total doesn't reveal how much electricity reaches each production area or shared service. Submeters give you those smaller splits. A sub meter measures a branch of the electrical supply, so its usefulness depends on what that branch feeds and whether separating those loads helps someone answer the chosen question. I'd place meters where the answer could change a decision if knowing the furnaces consumption could support a different operating choice. Separating that load deserves more attention than dividing office lighting into ever smaller groups. Our hypothetical furnace needs a careful inspection of its electrical feeds. The heating elements might use one supply, while fans and handling equipment use another, so a meter on the heater feed alone wouldn't measure everything people might call furnace electricity. Give that measurement a precise name. Furnace Heater supply tells the maintenance team what the meter covers. While furnace energy could imply a wider measurement than the installation can actually support or explain, the compressor raises a similar placement question. A meter at the compressor package might capture its electrical consumption, but a separate dryer or cooling unit could sit elsewhere and remain outside that reading unless the installation includes it. Check that arrangement before choosing the meter position. If the question concerns the electricity needed to provide compressed air, the electrical team needs to identify the supporting loads rather than assume the compressor feed contains them. All high load machines deserve attention, but load alone isn't the deciding factor. I'd also consider whether consumption varies enough to matter, whether someone can act on the result, and whether an existing meter already provides a useful measurement. Sometimes an area level meter is the sensible first step. If several machines always operate together and the initial question concerns that whole area, measuring each machine separately might add installation work without changing the immediate decision. Shared circuits can limit what you learn. When a machining cell shares its supply with extraction equipment and nearby services, the meter records their combined electricity rather than a clean measurement of the machining cell alone. You have several choices at that point. You can keep the combined measurement and label it honestly, add meters on suitable downstream branches, or ask whether changing the electrical arrangement would justify the cost and disruption. Unmetered branches need an explicit place in the meter plan. Leave them identified as uncovered supplies, rather than allowing a collection of well instrumented machines to create the impression that every relevant load now has a measurement machine, controller values can help you investigate those gaps. Some controllers expose values from electrical measurement hardware, while others calculate consumption from drive information. Operating states or assumptions that might exclude equipment outside the controller's own reach. Ask where the number comes from. A displayed energy value needs documentation of its source and coverage before you treat it as equivalent to a dedicated meter installed on the machine's full electrical supply. Installation belongs with qualified electrical personnel. They need to choose suitable equipment, assess the circuit and working conditions, and plan safe isolation or other approved procedures without treating the monitoring project as permission to bypass plant safety rules. Even a read only project involves physical electrical work. Adding a meter or current transformers can require access to electrical panels, so production, access permits and any required shutdown need agreement before the installation date. Accuracy depends on the whole measurement arrangement. The meter specification matters, but so do current transformer selection, installation configuration and the checks that confirm the complete setup measures the intended circuit within the required tolerance. 00:17:40 Speaker: Keep calibration and verification records with the meter's installation details. Record the meter's identity, the circuit it measures commissioning checks, and any later inspection or calibration results so someone can assess the readings basis rather than trust the label. Once installed, the meter still needs to deliver a usable reading. The next question concerns what that reading actually represents, because software receiving the value can't repair a measurement whose meaning never left the electrical panel. A number arriving from a meter needs a clear description. If the furnace sends a value of forty. The receiving system needs to know the unit, the reading type, and the time that value describes before it can interpret anything. Give each meter a stable identity in the data. A friendly name can help people recognize it, but the record should carry an identifier. That won't change when someone renames the furnace or edits a maintenance label. Keep the unit attached to the value even when a source normally sends the same unit, its documented data format needs to make that unit explicit, so a later configuration change doesn't silently change the meaning. Time needs the same care record when the measurement occurred at the source with enough information to identify the time zone or offset, rather than relying on a timestamp added somewhere downstream. Some devices don't timestamp their own readings. In that case, record when the collecting equipment read the value and document that this represents collection time rather than a precise timestamp supplied by the meter itself. Include the readings quality status as well. If the device flags a value as invalid, uncertain or stale, carry that status with the value instead of stripping it away because the number still looks plausible. A stale value can be particularly misleading. The device might keep returning its last successful measurement, even though it can no longer update it, so an unchanged number doesn't always mean consumption has remained steady. The reading type tells you how to use the number. A cumulative energy register is a running total, while an interval energy reading describes consumption across a stated period, and sampled power describes a rate at a particular time for a cumulative register. Consumption comes from the difference between readings. You need both endpoints and their timestamps, because the difference represents energy accumulated between those moments, rather than energy consumed only when the second reading arrived. An interval reading already describes energy across a period. Its record needs the interval boundaries, or a clear convention that identifies them. Because a timestamp alone might mean the start or the end of that period. Sampled power requires a different treatment. Estimating energy from those samples means accounting for elapsed time, and making an assumption about the power between samples, which should remain part of the documented calculation method. Cumulative registers bring their own complications. A reset, a rollover at the registers maximum value, or a replacement meter can interrupt the running total. So the next reading can't always be compared directly with the previous one. Those events need records of their own. When maintenance replaces a meter preserve, the old meter's closing reading where available, the new meter's opening reading, and the time the replacement entered service on that supply. Don't hide a reset by forcing the difference to zero. That would discard the discontinuity without explaining how much energy belongs to the affected period, leaving a clean looking record that no longer describes what the source actually measured. Scale factors also need a history. If a meter sends raw register counts that require multiplier. Retain the raw value and the applicable multiplier, including when a configuration change took effect for that source. Check where scaling already happens a meter, a collecting device, and a receiving application might each have scaling settings. So the data contract needs to state which value is raw and which value already includes the conversion. Then choose a collection rate that suits the question. Monthly consumption reporting may need much less frequent collection than an investigation into short machine cycles, provided the readings still support the required periods and checks for continuity. Faster collection doesn't guarantee finer measurement. Polling every second won't reveal second by second changes. If the meter updates its register less often, and repeatedly collecting the same value adds traffic without adding detail about the process. For short cycles, check the meters update behavior before choosing polling frequency, where the available readings can't separate those cycles. Record that limit rather than presenting the resulting calculation as though each cycle received its own measurement, I'd settle this reading format before connecting the plant to anything else. Once identity units, timestamps, reading types, and quality status travel with the values. The next task is moving them out without putting production at risk. The connection needs an operating agreement before it needs a cable information technology, or it manages enterprise systems and data services. While operational technology or OT runs and monitors the physical processes that keep the plant producing the maintenance windows can look very different. An enterprise team might schedule a service restart overnight, while the furnace team needs that same period to complete a heating cycle that nobody should interrupt for a monitoring change. Both teams have legitimate responsibilities to connect the dots between IT and OT, agree who approves access, who supports the connection, and who can stop collection if it affects equipment for our factory, consider the furnaces existing meter connection before introducing another application that pulls it. Check whether a controller or plant monitoring system already collects the required readings through an interface approved for other systems to use. An existing source can reduce the work at equipment level, but its owner still needs to confirm that another consumer won't exceed connection limits, interfere with current polling, or create an unsupported dependency in the production network. An industrial gateway can provide a controlled collection point. It reads approved plant sources and passes their data to what enterprise services, using hardware and software suited to the installation and supported by the people maintaining it keep its job narrow a gateway collecting meter readings doesn't need broad access to machine programs or unrelated controllers, even if granting that access would make the first connection easier during commissioning. The protocol depends on what the equipment supports. Modbus is common in metering, but reading it correctly requires the device's register documentation, including how values occupy registers and which requests the device. Accepts plan polling with the equipment owner asking every meter for every available register as fast as possible can load a shared communications link, especially when that link already supports production equipment with its own timing needs. Another option is Open Platform Communications Unified Architecture, or OPC. UA, where a supported server exists, it can provide name data items and configured security controls with access permissions that still need deliberate setup and testing. Support doesn't mean secure by default. Check which security modes the server supports, how certificates get trusted and renewed, and whether the collecting account can access only the data required for this energy project. Read only collection should be an enforced permission where a protocol or device can't restrict access by function. Use gateway controls and network rules to constrain requests and have the OT team review the remaining exposure network zones help contain that exposure. The meta network, production control network, and enterprise network shouldn't become one unrestricted space just because the energy team needs a path between a source and a reporting service. Use controlled boundaries between those zones, depending on plant policy. A collection service might pass data through an industrial demilitarized zone, a separate network area where approved exchanges can occur without granting direct enterprise access to controllers. Prefer outbound connections where the design allows them. A gateway can initiate an approved connection to what. An enterprise endpoint, which avoids opening a general inbound route from outside the plant to the equipment collecting reading's outbound traffic still needs limits, restrict destinations and service accounts, protect credentials, and agree how updates reach the gateway. Because a connection's direction alone doesn't prevent a compromised application from misusing its access. Give remote support its own approved path. Engineers maintaining the gateway should use the plant's controlled support process, rather than leave an extra remote access tool running because it proved convenient during the installation, work commissioned the connection under real operating conditions check communications load and equipment alarms with the OT team, then verify that stopping or restarting collection leaves the furnace and its existing control system operating as intended. That test needs an agreed rollback procedure. If polling introduces faults or unexpected delays, the team should know how to disconnect the new collector without improvising changes to equipment that still needs to complete production support. Responsibilities continue after commissioning, someone needs to own gateway health certificate, expiry access reviews and planned updates, with a clear route for operators to report trouble. When a monitoring change coincides with a production problem, then pull the network connection during an approved test. The furnace should keep operating while the collection team checks what happens to readings that can't leave the plant until the connection becomes available again. During that test, watch the collector rather than the reporting service. The useful question is whether new readings reach local storage while the connection stays down, and whether those stored readings survive a restart of the gateway itself. A queue held only in memory won't pass that second test. For readings you need to preserve the gateway needs durable storage with a confirmed right before it treats a reading as safely held for later delivery. This is store and forward behavior. The collector stores readings locally when delivery fails, then forwards the stored records when the connection returns, while continuing to collect new readings from the equipment. Make that behavior part of the acceptance test. Disconnect the approved network path, let the collector accumulate records and restart it under controlled conditions to check that pending records remain available. After it boots again, check the timestamps on those pending records. A reading collected during the outage must retain its original measurement or collection time, rather than receiving a replacement timestamp that makes it appear to describe the recovery period. Recovery time belongs in a separate field. If an hour of furnace readings arrives together after connectivity returns, the receiving service needs to recognize an hour of past activity rather than a sudden burst of current consumption. The gateway's clock needs monitoring to where the collector supplies timestamps. A clock that drifts during a long outage can move readings away from the production activity they actually describe, even though collection never stopped. Agree how the gateway maintains time while disconnected record clock synchronization status and any detected time corrections, so the support team can identify periods where timestamp accuracy needs review before those readings support later calculations. Delivery also needs a clear acknowledgement rule. The gateway should retain a pending record until the agreed receiving endpoint confirms acceptance with a documented definition of what that confirmation actually guarantees about storage downstream. A lost acknowledgement creates an awkward but normal case, the receiving endpoint might accept the record while the gateway misses the response, leaving the gateway uncertain whether it should remove that record from its pending queue. Retrying should be safe in that situation. Give each record a persistent identity before its first send, and keep that identity unchanged on retries, so a repeated delivery doesn't become another quantity of energy. Replaying a record mustn't create another consumption record. Test that explicitly by interrupting delivery around the acknowledgment step, then checking that retrying the pending record produces one accepted energy record rather than two. Recovery traffic needs limits as well. A gateway clearing several hours of stored readings should pace that traffic, so the backlog doesn't exhaust its resources or prevent it from collecting the readings arriving now, local storage has a limit and someone needs to size it, use the expected record volume and the outage duration the plant intends to tolerate, then leave room for normal variation logs and recovery work on the gateway. Monitor remaining capacity before the buffer fills an alarm needs to reach someone who can respond through an available route, which might require local notification when the same outage also prevents cloud based health messages from leaving. Choose an overflow policy before you need one. If storage fills, the collector might discard the oldest records or stop accepting new ones. But either choice creates a gap that must remain visible to the support team. Silent deletion isn't an acceptable default document. What gets lost? How the loss gets recorded, and who receives the warning. Rather than letting a full disk turn missing history into an apparently quiet production period, plant operation must remain independent of this backlog. A full buffer or failed upload shouldn't block a furnace sequence, hold a machine operation open or require an operator to restore cloud access before production can continue. That independence needs a physical test, not just agreement with the equipment owner. Present. Verify the monitoring systems failure behavior and confirm that restoring collection doesn't require changes to the local controls. Keeping production running. Only then would I choose the Azure Ingestion path. The choice should follow how the gateway connects, what delivery confirmation it needs, and how stored records replay. Rather than starting with the preferred service and hoping the plant can adapt, Azure IoT hub fits when you're managing connected devices. If the energy project needs a registered identity for each gateway control device access and a way to manage those connections, that device oriented model deserves consideration. The registered device might be the gateway, not the meter. In a hypothetical plant, one gateway could collect readings from the furnace meter and several machining cell meters, then send those streams through its own authenticated connection. Those identities answer different questions as your needs to establish which connected gateway can send messages, while your energy application needs to establish which physical meter each message describes and which gateway supplied it. Keep both identities in the receiving records. Authenticating the gateway doesn't independently prove which meter produced a reading, because that relationship still depends on the collection configuration and the controls around who can change it. Device registration also creates an ongoing support task. Whoever operates this connection needs to provision gateway credentials, replace or revoke them when required, and confirm that a retired gateway can no longer send readings into the production service. IoT hub isn't mandatory for every meter stream. If an existing industrial platform already manages equipment connections and publishes collected records, Azure might only need to receive that platform's event feed rather than manage another device layer. Azure Event Hubs fits that event ingestion role. A gateway or existing platform can publish events into an event hub, and downstream applications can read those events independently without each application needing its own connection back to the plant. Treat that as a different responsibility. Event hubs receives event streams from authorized senders, while the source platform remains responsible for the equipment, connections and any meter level management that happens before publication for our plant, either path could be reasonable. A separately managed gateway fleet could point toward IoT hub, while a supported energy management platform with an existing publishing interface could point toward event hubs instead. You don't need both. By default, choose the path that meets the connection and operating requirements, and add another service only when you can name the responsibility it takes on and who will support it. Work through the receiving contract before deployment, the source team and Azure team need to agree on the message format, expected traffic, permitted message sizes, and the response when the receiving service temporarily limits incoming traffic. Capacity includes recovery traffic, not just normal production. When several gateways reconnect together, the service configuration needs to accommodate their backlog alongside current messages within the limits and pricing of the selected service tier. After ingestion, a small processing step might help. Azure functions can run code in response to incoming events. Check whether a message matches the agreed format and route it toward the appropriate downstream destination. Keep each function's task bounded, checking required fields or selecting a destination suits a small handler while packing every later energy calculation into that same handler makes its failures harder to isolate and repair. Failure behavior belongs in that design. If a function writes a record successfully but fails before its invocation completes, a retry can execute the write again unless the receiving operation recognizes that it already succeeded. That makes retry safe processing a deliberate implementation choice. Use the persistent record identity established at the source, with a destination side rule that prevents repeated processing from creating repeated results when the same event reaches the function again. The trigger doesn't settle every delivery question check the selected triggers, retry and checkpoint behavior, then define how failed records remain recoverable rather than assuming that any failed invocation automatically receives the handling your application needs. Separate service ownership from calculation ownership the Azure Operations Team can keep ingestion available and monitor processing failures, while the Energy data owner defines which records the application should accept for further use. An accepted message proves something fairly narrow. It tells you that the receiving service accepted the transmission under its connection and access rules, rather than proving that the enclosed energy reading can support an order calculation. The furnace gateway could authenticate correctly and send unusable content. Neither IoT hub nor event hubs knows your Meta's intended measurement, meaning simply because the message arrived through an authorized connection and entered the configured service. Successful ingestion is the start of another check. Before that message becomes an energy record people use. The receiving application needs to establish whether its contents meet the agreed rules for the measurement they claim to describe. Put incoming readings through checks that match the source's documented behavior. I'd separate records that pass from records needing investigation while keeping rejected content available, so someone can work out whether the source or the checking rule needs correction. Begin with the agreed schema, meaning the structure of the record check that required fields exist. Values have the expected data types, and timestamps can be interpreted rather than accepting any message that happens to contain a number. Unit checks should test meaning not just spelling a field labeled as interval energy shouldn't accept a power unit, even if the value falls within a range that looks perfectly reasonable for the furnace. Conversions need controlled rules. If a source changes from what hours to kilowatt hours, accept that change through an approved conversion or configuration update, rather than letting the receiving application guess from the size of the value for cumulative readings, calculate differences only across valid pairs. A negative difference should send the affected period for review against known meter events, instead of passing through as negative consumption. That reduces the factory's reported total. A positive difference can also deserve investigation. Compare the implied average power with a documented limit for the measured supply, allowing for measurement tolerance and the equipment actually connected during that period before declaring the reading impossible. Use those limits as checks, not automatic corrections. If a reading exceeds the expected range, preserve it with a reason for review, because clipping it to a chosen maximum replaces evidence with an undocumented estimate. Duplicate handling needs a similarly careful rule where the source supplies a stable event identifier. Use that identifier together with its source identity to recognize repeat deliveries without accidentally merging records from different meters. Some older sources won't provide an event identifier. You might then use a documented combination of meter identity reading type, source timestamp, and sequence information, provided that combination really distinguishes the records the source can generate. 00:35:34 Speaker: Equal values alone. Don't establish a duplicate. Two valid observations can contain the same number, so removing every repeated value could erase legitimate records and leave gaps that you're processing. Then mistakes for collection failures. Conflicting records need a separate path. If two messages claim the same event identity but contain different values, flag the conflict for investigation instead of silently keeping whichever version happens to arrive last time. Checks bring another decision into the pipeline. Processing should place readings according to the time they describe, while retaining arrival time to measure delivery delay and help explain why a reporting period changed after its first calculation. Out of order readings can change which values belong together. For a cumulative register, a late reading between two existing readings can refine the interval breakdown, so the processing needs to revisit affected differences rather than simply append another result. Agree how long a period remains provisional. A defined allowance for late arrivals lets the team balance timely information against completeness, with readings arriving beyond that allowance. Entering a controlled correction process rather than disappearing Missing intervals should remain visible as missing intervals. Check expected arrivals against the collection schedule and distinguish a delayed record from an invalid record. Because those conditions can need different responses from the support team. A gap doesn't always erase the total energy information. Valid cumulative readings on either side may establish consumption across the wider gap while leaving its distribution within that period unknown, so completeness needs to describe both coverage and timing. Estimation can help when the intended use permits it. For example, an agreed method might use comparable operating periods to fill a gap, but the resulting quantity must carry an estimated status and a reference to that method. Keep measured and estimated quantities distinguishable throughout processing. If the missing records arrive later, replace or revise the affected estimate through a recorded calculation change so the estimate and the recovered measurement don't both contribute to consumption. Make these rules testable with deliberately faulty records, send a wrong unit, a conflicting duplicate and a late reading through the pipeline. Then check the accepted results and review records rather than relying only on successful message counts. Each record now needs a disposition and a reason passing records can support calculations, while flagged records rejected content and estimation decisions need durable retention, so an investigation or recalculation can recover what actually entered the pipeline. Storage choices should follow what you'll need to do with the records. I separate the retained source material from the tables people query, because those two jobs place different demands on storage, processing, and day to day support. The retained source material gives you a way back. If a transformation contains an error, you need the original payload and its source details available without asking. The plant gateway to supply months of history again, Azure Blob Storage can hold that retained material. Azure Data Lake Storage builds on blob storage with a hierarchical namespace, which can suit analytical workloads that organize files into directories and need directory level access controls. Choose that arrangement deliberately, rather than treating it as temporary staging. Organize records by sensible source and time boundaries. Preserve the received content, and set retention rules that match the period during which investigation or recalculation might be required. Raw storage shouldn't depend on a later calculation succeeding. A retained record needs to remain recoverable, even when the processing job that turns it into a usable analytical row fails and needs repair. Retention also needs a deletion policy. Keeping everything forever can create avoidable costs and access exposure, while deleting too early can remove the evidence needed to reproduce a result that someone still relies on. Microsoft Fabric takes on a different job here. One lake provides fabrics, shared logical data lake and a lake house can hold files alongside managed analytical tables that multiple supported processing and reporting tools can use. Those tables contain data prepared for defined analytical use. For our furnace. That could mean a table of accepted energy intervals with consistent units. Rather than asking every consumer to interpret the original gateway messages independently. A lake house table isn't just another copy without a purpose. It gives consumers a documented structure and lets processing produce repeatable data sets, while the retained source material remains available if that preparation needs to run again, you don't necessarily need another physical copy of every file where supported one leg shortcuts can reference data in external storage, including Azure Data Lake storage. So assess whether referencing retained files meets your access and performance needs. A shortcut doesn't transfer every responsibility to fabric. The external storage still needs its own permissions, availability, and retention settings, and deleting the underlying files removes the data that the shortcut points toward. Streaming analysis can justify another processing path. Fabric event streams can receive and route supported event sources with supported transformations. When the requirement calls for handling incoming events without waiting for a scheduled batch job, check the actual source and destination support before committing connection. Methods, authentication options and limits can differ, so validate the path from your selected Azure ingestion service rather than assuming every fabric item connects directly to everything else. Event House addresses time based analytical queries over event data. It's kusto based databases can suit investigations across frequent readings, such as asking how furnace demand changed during a particular operating period without scanning a large archive of original messages. That doesn't make Eventghost compulsory for energy monitoring. If the business question needs a daily consumption data set, a scheduled transformation into lakehouse tables might meet it with less operational work than maintaining a separate streaming analysis path. Latency needs a business reason. Ask how much sooner someone can act if processing finishes within seconds rather than hours, then compare that benefit with the extra capacity and support the faster path requires. Volume changes the calculation to frequent readings from many sources. Place different demands on ingestion, file organization, and query performance than a modest collection of daily totals that people review once each week. Small files deserve attention during testing. Writing every message as a separate file can burden later analytical reads, so use supported batching or consolidation methods that preserve the required detail without creating needless file handling work cost depends on more than stored bytes include processing capacity, retention, query frequency, and data movement where applicable, then test representative workloads instead of estimating the whole system from the cheapest storage price alone. Team skills should influence the choice. A design your support team can diagnose and restore usually beats a more elaborate design whose behavior nobody understands. When an overnight processing run leaves the morning data set incomplete, the next obstacle sits beyond storage. Your meter records can now be retained and queried by connecting them to production still requires agreement between the identities used by the energy sources and those used by the production systems. Try asking each system for the Furnace's identity. Maintenance might return an asset number, production might return a work center code, and the energy application might return a label created by whoever commissioned the meter connection. Those identifiers can all be legitimate. Your factory model needs to record which identifiers describe the same equipment, which describes something broader and which only looks similar because people chose familiar names. A work center doesn't always mean one physical machine in an enterprise resource planning system or ERP, it might represent a group of resources used for planning, while maintenance tracks each individual asset separately. That distinction affects what you can connect with confidence. If the ERP code represents the whole heat treatment area, matching it directly to one furnace would give your data a level of detail the business record doesn't actually contain. I'd create an identity register before building the energy dashboard for every relevant meter, equipment, asset line, site, product and order. Record the stable identifier and the source responsible for maintaining that identity. Keep source identifiers alongside the shared identifier. When someone investigates a disputed order, they should be able to find the original record in its source system without reverse engineering. How the analytical platform renamed it. Names still help people navigate the information, but matching records by names alone creates trouble when abbreviations change. Two sites use the same label or an operator types the equipment name slightly differently. Site identity needs particular care across several plants. A machine called cell four might exist at every location, so its local name needs a site context before anyone treats it as a unique equipment reference. Order numbers can have the same problem. An order reference might only be unique within a company, plant or source application, which means the shared identity must preserve that context rather than keeping the visible number alone. Products need a similarly deliberate definition. If your systems distinguish product revisions or variants, decide which distinctions the energy data set must preserve and don't merge them simply because their descriptions happen to match. The ERP provides the business context for these identities. Depending on your implementation, it can supply production orders and master data describing products, organizational structures, and the resources used for business planning and transactions. Confirm those responsibilities in your own environment. A source of record label should describe who actually maintains and approves the information, rather than reflect an architectural preference that nobody follows during daily production work. The manufacturing execution system, or Mez, contributes recorded shop floor execution. It can identify which equipment the production team records against an order or operation using the execution detail available in that particular implementation. Its identifiers won't necessarily match the ERP identifiers. One business order might appear through several execution records, so your integration needs the documented references connecting those records, rather than an assumption that matching text proves identity. Ask production to validate a few real examples. Choose records that people recognize. Find them in both systems and check whether the proposed mapping describes what the plant actually recorded against each resource. Some records will remain unresolved at first, give them an explicit review status and a named owner, rather than attaching them to the nearest plausible asset just to increase the percentage of successfully matched records. Ownership is the part that keeps this model usable. Someone must approve new equipment identities, someone must maintain product references, and someone must resolve disagreements. When the production and maintenance records describe an asset differently, that doesn't require one person to own every definition. It requires a clear change process that brings the relevant owners together. When a correction crosses system boundaries, instead of leaving the integration developer to decide what the factory means. Equipment moves provide a useful test of that process. If a machining cell transfers to another line, who updates its location? Who checks the production system reference, and who confirms whether the meter assignment still describes the installed arrangement. The physical asset should retain its identity through that move. Its location and assigned references can change, while the register preserves enough change history for someone to understand which records describe the old arrangement and which describe the new one. Now, the factory model can support a more specific question. Once everyone agrees which assets and production records they're talking about, we can describe where electricity flows and which production activities depend on the equipment receiving it. Knowing which asset a record describes gives us a starting point. Allocation needs the connections between assets because a meter measures an electrical supply, while a production order refers to work performed using equipment connected to that supply. Follow the furnace meter through those connections. The model should describe which circuit it measures, which equipment that circuit supplies, and which process uses that equipment without treating those relationships as interchangeable links. Each connection answers a different question. Measures describes the meters relationship to the circuit, while supplies describes the circuits relationship to equipment and performs connects that equipment to a production process. Those verbs carry practical meaning. If every connection simply says associated with software can't distinguish equipment receiving electricity from equipment merely located nearby, and someone must reconstruct that meaning whenever they build a calculation. Consider the furnace and its handling equipment. They might participate in the same heat treatment process while receiving electricity through separate circuits, so the process, relationship and the electrical relationship describe different arrangements within the same production area. Keep both arrangements in the model. Following the electrical connections establishes what a reading covers while following the process. Connections. Establishes which resources participate in the activity you're trying to understand. For a product that brings product roots into the discussion, product process resource relationships connect what you're producing to the operations it requires, and the equipment or other resources capable of carrying out those operations under the specified conditions. The root describes more than a list of machines for our hypothetical plant, a product might require machining followed by heat treatment, with each process linked to the resources approved to perform it for that product. Capability and actual use need separate relationships. A furnace approved for a product describes an available option, while a furnace recorded against an executed operation describes the resource production actually used for that particular work. Confusing those relationships would spread energy across the wrong resources and allocation query needs to follow the recorded resource assignment rather than treat every machine capable of processing the product as though it participated in the order. A knowledge graph makes these relationships explicit and queryable. It represents entities such as products, processes, circuits, and equipment, then gives the connections defined meanings that applications can follow when answering a particular question. For example, start with a product's heat treatment process. Follow its resource relationship to the furnace, then follow the electrical supply relationships to identify the meter records relevant to that equipment, subject to the coverage those relationships describe. The graph doesn't decide the allocation method. Its contribution is making the relevant connections available for inspection and calculation, so the method can use documented relationships instead of embedding hidden assumptions in separate reporting queries. A digital twin adds a maintained representation of relevant assets and state in this setting that might describe the furnace, its connected services, and its current operating condition, with updates from the systems responsible for those facts. The twin needs a defined purpose. You choose which relationships and states to maintain because they support the energy question, rather than trying to reproduce every physical detail of the factory in software. Neither concept requires starting with a graph database. Relational tables can describe meters, circuits, equipment, processes, and their connections quite well, particularly when the questions follow a modest set of clearly understood relationships. I'd start with the simplest structure. The team can maintain a table linking circuits to equipment with explicit relationship types, and controlled references can provide a sound basis without introducing another platform merely to use its name. Relationship depth can change that choice. When many questions follow different paths across shared services, alternative resources and product routes. Graph tools may make those paths easier to query and inspect than repeated chains of table joins. Maintained asset state can justify twin tools for a different reason. Assess the effort needed to keep that state current and check whether a twin platform addresses that task better than the systems and tables already available. Whichever structure you choose, test its answers with production and electrical owners. Ask whether a path from product to process to equipment to meter describes the installed arrangement, and require unclear connections to remain unresolved until someone confirms them. There's a time condition attached to every confirmed path. A circuit can supply one arrangement today and another after a plant change, so the model must distinguish relationships valid for past readings from relationships describing the factory. Now, the next join needs to follow production as it actually occurred. A valid equipment relationship tells you where to look, but the operations recorded start and stop determine which period of activity you're trying to connect to the energy record. A scheduled start won't establish that period. Suppose our hypothetical furnace batch starts later than planned because parts remain in machining, while the furnace completes work belonging to an earlier batch during the scheduled slot, using the schedule would connect readings to the wrong work. Keep plan times for comparing intentions with execution and use recorded execution times when establishing when an order actually occupied the resource involved in the energy calculation. Even recorded times need a clear definition. An operator pressing start might mean loading begins, the heating program begins, or an execution record opens depending on how the plant configured that event. Ask production what those events mean on this furnace. If the record spans loading through unloading, preserve that meaning instead of treating the whole recorded duration as though it describes only the heating program, running an operation can need more than one time window, where the production system records distinct stages. Retain those stages so a later calculation can select the period relevant to its stated purpose without inventing boundaries. Put execution events and energy intervals on a shared time basis. I'd normally store their comparable timestamps in coordinated Universal Time or UTC while retaining the source's original time information for investigation and local interpretation. Shift calendars still belong to local plant operations. A night shift crossing midnight needs an agreed production date rule, so its operations don't split between reporting days, simply because a calendar date changed during the shift. Daylight saving changes make local timestamps particularly awkward when clocks move backward. The same local time occurs twice, and an execution event needs its offset or another unambiguous reference before you can place it correctly. Test the spring clock change as well. A local time subtraction can suggest an extra or missing hour, so calculate elapsed duration from unambiguous timestamps and apply local shift labels separately through the plant's calendar. Then check whether the execution windows can co-exist. Two orders recorded against one resource during overlapping periods might reflect concurrent work. Separate positions within the equipment, or a recording mistake that production needs to resolve. Overlap isn't automatically a data error. Our furnace can process several orders within one batch, so the model needs one batch execution window linked to its participating orders, rather than pretending each order occupied the furnace alone. The order link needs a defined meaning to confirm whether every linked order remained present throughout the batch window, or whether the records describe different participation periods that the matching process needs to preserve. Setup needs its own time identity. If operators prepare the resource between orders, retain that setup window separately where the source supports it without deciding during this join which order should eventually carry its energy. Interrupted runs need more detail than one long duration, an operation that pauses and resumes should retain those events, so the matching process can distinguish its recorded segments rather than assume uninterrupted activity between the first and last timestamps. Some systems only provide a start and a completion event. In that case, label the result as an elapsed operation window with the known limitation that unrecorded pauses remain inside it rather than being silently removed. Corrections can arrive after the shift has ended. An operator might fix a completion time or move a batch reference to the correct order, changing which execution windows intersect the energy intervals already processed for that resource. 00:53:41 Speaker: That change should trigger a bounded recalculation. Identify the affected resource and period. Rebuild the relevant matches and record which production correction caused the revised result. Instead of refreshing everything without a traceable reason. Keep provisional and approved results distinct during that correction. The team reviewing an order needs to know whether its energy match reflects the current execution record or an earlier version still awaiting controlled replacement in reporting. The output here is a time match, not an allocation. It establishes which energy intervals intersect, which recorded operations, including ambiguous overlaps, before any calculation decides how much of an interval belongs to an individual order. Direct attribution becomes possible where those boundaries genuinely match with a confirmed meter boundary and a confirmed operation window. We can next examine what the measurement supports without claiming more detail than the timing allows. Take a dedicated machining cell meter and one confirmed operation window. If no other order occupies that cell during the window, you can attribute the electricity measured across that period to the operation within the meters. Documented coverage. The cleanest case has readings at both operation boundaries. Subtracting the opening cumulative reading from the closing reading gives the electricity recorded during that window, without needing to divide a larger interval between different jobs. That remains an operation level measurement, not a measurement of each part. If the operation produces several parts, dividing its electricity among those parts introduces another calculation. Even though the operation total comes directly from the readings at its boundaries. Most installations won't give you perfectly placed boundary readings. The operation might begin between two meter observations and end between another pair, leaving some intervals wholly inside the operation and others crossing its start or finish. Treat those two situations differently. Energy recorded wholly within the exclusive operation window, can enter its attributed total directly, while energy in a crossing interval needs an explicit rule for determining the operation's share. A time based split is one possible rule. It assigns the operation a fraction of interval energy equal to the fraction of that interval covered by the operation, assuming consumption spreads uniformly across the interval. Let's put numbers on that assumption in a hypothetical ten minute interval containing six kilowatt hours. An operation occupying five minutes receives three kilowatt hours under an equal by time split, regardless of how the load actually varied. The arithmetic is exact. The distribution is estimated if the cell draws most of that electricity during the other five minutes. The calculated share won't describe the operation's actual consumption, despite adding up correctly across the interval. That distinction should shape the results description. Call it electricity attributed using a time based boundary split rather than electricity measured for the operation because only the interval total received its own measurement. In this example, short jobs make this limitation more noticeable when several jobs finish inside one meter interval. Splitting the total by their durations assumes each job draws the same average power while it occupies the measured equipment. Different cutting programs can break that assumption. A short operation with a demanding cut could consume more electricity than a longer, lighter operation, so duration alone might reverse their actual ranking for energy use. You can still use the split for a defined purpose. I'd make the estimated contribution visible and check whether the intended comparison remains useful. Rather than reject every imperfect result or present every calculated share as equally certain, better supported splitting needs evidence about the changing load. If suitable, higher frequency power measurements exist, their relative pattern can help distribute the measured interval energy provided the calculation documents their coverage and how the weighting works. Without that evidence, retain the simpler rules stated limits don't invent a detailed consumption profile from operation names or nominal machine ratings, merely because the resulting order totals look more specific than a time based estimate. The records should preserve that distinction structurally. Keep the measured interval quantity in its own record, then create separate attribution records linking portions of that quantity to the relevant operations and their production orders. Each attributed share needs a method reference store, the source interval, the overlap duration, the rule version, and the resulting share so another person can reproduce the allocation without relying on whoever originally wrote the calculation. Check the accounting across every split interval. The attributed shares and any remainder must sum to the measured interval energy within the chosen rounding tolerance and the original measurement must remain unchanged by the splitting process round for presentation. After doing the calculation, repeatedly rounding small shares before adding them can create apparent differences. While displaying many decimal places can imply confidence that the underlying time based division simply doesn't support an order. Total can contain different kinds of evidence. Separate the contribution from intervals wholly inside its operations from the contribution estimated at crossing boundaries, so users can understand how much depends on the splitting rule. There's also electricity outside those operation windows preserve that remainder rather than stretching an orders window to absorb it. Because energy consumed before, between, or after recorded production needs an attribution decision of its own, the furnace can consume electricity before any batch enters it in a hypothetical plant. The team brings the furnace to its required operating temperature before loading the first batch, creating consumption that no loaded batch window captures. Warm up supports production, but deciding who carries it requires a rule. Charging it all to the first batch could make that batch appear unusually energy intensive, even when later batches benefit from the temperature established before the first load. A campaign based rule gives you another option you could collect warm up electricity separately and distribute it across the batches in a defined furnace campaign, using a documented basis appropriate to the process and intended calculation. Define when that campaign starts and ends. If the furnace cools down and needs another warm up, the calculation needs to recognize a new preparation event rather than spread that consumption across unrelated batches from earlier production machines. Standby raises a different ownership question between orders. A machining cell might keep its controls and auxiliary equipment running, drawing electricity while the next order waits for material or an available operator. That consumption belongs in a named operating category. Production might distinguish a short planned gap from a prolonged wait, because those conditions give the team different opportunities to reduce consumption without disrupting the next operation. Don't automatically charge every gap to the following order. Doing that can make a product look worse simply because it followed a delay, even though neither its machining requirements nor its processing conditions caused the weight use. Named rules for setup idle time and unassigned consumption, a setup rule might charge preparation to the incoming order, while an idle rule might retain waiting consumption at area level, provided those choices fit the calculations. Purpose. Unassigned consumption needs an owner even before it receives an order. Keeping it in a separate category lets the production team investigate why it remains outside the agreed rules, without forcing a convenient assignment that hides an unresolved question. Now consider parts that don't pass inspection. The energy used to process them remains part of the factory's consumption, even when those parts never become saleable output and therefore don't appear in the good part. Count. Keep processed quantity and good quantity. Separate energy per processed part describes consumption across the parts receiving the operation, while energy per good part relates that consumption to the accepted output under the stated quality definition. Here's a hypothetical batch with a simple quality loss. If ten processed parts consume one hundred kilowatt hours and only eight pass inspection. The batch averages ten kilowatt hours per processed part and twelve point five per good part. The difference comes from yield, not a change in measured consumption. Keeping both measures helps production distinguish a process using more electricity during operation from a process consuming the same electricity but delivering fewer accepted parts afterward. Scrap also needs its place in the route. A part rejected after machining shouldn't receive heat treatment electricity for a stage it never reached, while a part rejected after heat treatment has already passed through that energy consuming operation. Record the rejection stage alongside the affected quantity for an order level calculation. An agreed rule might retain scrap related energy against the order, while the good part measure spreads that burden across its accepted output rather than deleting it, rework adds another pass through production. If a part returns for another machining operation, link that execution and its energy to the original part, batch or order through whatever traceable reference the production system supports. Rework isn't another newly produced part. It's extra processing belongs in the energy total, but counting the same physical part again as new output would dilute the reported energy per good part and conceal the cost of correction check when acceptance becomes final. For the measure, you're publishing, a provisional pass at one stage and final acceptance at another can produce different good part counts. So the quality status needs a clear definition and reporting point. These rules can affect comparisons between products and shifts before interpreting a higher figure as poor machine performance. Check whether it contains more warm up, longer waits, lower yield or additional processing that the comparison period didn't require. There's still consumption that supports several resources at once. After classifying preparation, weighting and quality losses. The next difficulty is a shared service whose electricity can't be assigned simply by finding the order occupying one machine. Compressed air gives us a useful case to work through. The compressor supplies several machining cells, but those cells can demand very different amounts of air while spending the same amount of time on an order. Runtime alone won't capture that difference. One operation might use air in brief bursts, while another keeps an air driven tool running for much of its cycle. So equal operating durations don't establish equal demand. Cooling creates a similar allocation problem. Two machines connected to the same cooling system can reject different amounts of heat, and their share of the cooling systems. Electricity won't necessarily follow their production counts. I choose the allocation driver with the utility owner. The driver is the quantity used to divide shared consumption, and it should have a supported relationship to how the connected equipment uses that particular service. Measured airflow can provide that relationship with suitable branch flow meters. You can establish the volume delivered to each production area over the allocation period using consistent reference conditions. When comparing those volumes, flow still doesn't directly measure compressor electricity for an order dividing electricity according to delivered air volume assumes that volume is a suitable basis for sharing the system's consumption across the users included in that calculation. Pressure requirements can complicate that assumption. If one process requires a higher supply pressure, its effect on compressor electricity might extend beyond its volume share, particularly when the whole system operates at that pressure. Cooling needs its own supported driver. Depending on the installation, measured coolant flow, together with the temperature difference across a user, can help estimate heat removed, which can support sharing electricity associated with providing that cooling. The service quantity and electricity aren't interchangeable. Cooling equipment's efficiency can vary with operating conditions, so a thermal demand allocation remains a method for distributing electricity rather than a direct electrical measurement at each connected machine. Not every plant can measure those service quantities immediately. Runtime can provide a first allocation basis where equipment demands are reasonably comparable, or where the intended use accepts the limitations, and those limitations remain explicit. Machine demand estimates can refine that starting point. You might combine recorded time in particular operating states with documented air demand estimates, provided engineering confirms that those estimates reasonably describe the equipment and conditions involved. Treat rated demand with care. A tools published air requirement might describe operation under specified conditions rather than its average use throughout an order, so multiplying it by the entire elapsed order duration could overstate demand. Separate background consumption before dividing the production related share. A compressed air system can consume electricity, while connected machines aren't requesting useful air, including consumption associated with leaks or the compressors own unloaded operating behavior. Background load needs an agreed treatment. You might retain it as a utility overhead or distribute it using a stated reporting rule, while using measured demand to divide the portion that changes with production activity. Don't assume the background stays constant. Pressure settings, equipment availability, and system condition can change it, so the utility owner needs to confirm whether the chosen estimate still represents the period being allocated. Consider a deliberately simplified hypothetical allocation. Suppose engineering has identified one hundred kilowatt hours of production related compressor electricity for a period after treating background consumption separately under the plant's agreed method for that calculation, two orders each occupy their machines for one hour. A runtime based rule assigns fifty kilowatt hours to each order because their durations match without considering whether their machines actually request similar quantities of compressed air during those hours. Now, suppose branch measurements show unequal air use. If one order accounts for three quarters of the measured volume and the other accounts for one quarter, a volume based rule assigns seventy five and twenty five kilowatt hours, respectively. The compressor total hasn't changed. The assigned shares have. That difference tells you how strongly the order comparison depends on the driver, and it gives engineering a specific assumption to assess before finance relies on the result. Document why the chosen rule fits the system. Record the service being divided, the driver, the treatment of background consumption, and any connected users whose demand isn't represented well enough by the available measurements. Keep the method visible alongside the assigned quantity. An order supported by branch flow measurements and an order supported by runtime estimates shouldn't appear equally certain merely because both results display the same number of decimal places. Uncertainty belongs in the result, not just the calculation notes where a different, reasonable driver could materially change an orders share. Preserve that limitation explicitly rather than letting precise arithmetic imply a precise understanding of consumption. Before trusting those order shares, check the electricity balance for the site. Add the relevant submeter totals and compare them with the main meter over exactly the same period, keeping any difference visible rather than distributing it automatically across production. That comparison needs a defined meter hierarchy. If an area meter includes the furnace supply, adding both readings to the site subtotal would count the furnace twice, so select measurement points that don't overlap electrically. For our hypothetical plant, build a reconciliation for one closed week. Compare incoming electricity with the non-overlapping measured branches. Then list the supplies outside those branches. So the difference has a physical context that the electrical team can investigate. Matching dates alone won't guarantee matching periods. A main meter total ending at midnight and a submeter total ending at the morning shift change. Describe different windows even when both records carry the same reporting date. Use common opening and closing boundaries for the comparison. Where readings don't land on those boundaries, record how you derived the period totals and how much of the reconciliation depends on estimated boundary values rather than matching observations. Next, compare the incoming electricity record with the utility invoice. Check the invoices, actual reading dates, meter reference and whether its quantities come from measured readings or supplier estimates. Before treating a difference as a fault in your pipeline, a billing month might not follow your calendar month if finance closes its report on the last evening of the month, but the supplier reads the meter several days later. Reconcile the overlapping period rather than compare unmatched totals. Separate the invoice quantity check from the financial. Check your establishing whether the billed electricity quantity agrees with the corresponding meter record. Not trying to explain every charge on the invoice through a kilowatt hour balance. The remaining difference needs named categories and supporting evidence. Unmetered building services, electrical distribution losses, measurement error, and mismatched boundaries can all contribute, but their presence shouldn't become a convenient explanation for whatever quantity remains after subtraction. Some differences have a physical basis. Electricity measured before a transformer can exceed electricity measured downstream because of transformer losses, so meter placement determines whether those losses fall inside the balance you're attempting to close. Other differences require investigation rather than a loss allowance. If submeter totals exceed the incoming total, check overlapping coverage, configuration period. Boundaries and measurement uncertainty. Instead of recording a negative unexplained balance as though it represented physical consumption, keep unallocated energy separate from unexplained energy. You might know that a measured quantity supplies general building services without assigning it to products, whereas an unexplained residual means the available evidence doesn't yet establish where that quantity belongs. Give each unresolved difference an owner and review date. The reconciliation record should distinguish confirmed explanations from working hypotheses with links to the checks supporting them. So next month's reviewer doesn't inherit guesses that gradually acquire the status of facts. Set review thresholds with the measurement team. A reasonable threshold depends on the installed meters, expected losses, reading resolution and boundary estimates, rather than a universal percentage copied from another factory with a different electrical arrangement. Confidence labels should describe evidence, not personal comfort. A label such as reconciled with an estimated branch tells users more than high confidence, especially when the estimate contributes enough electricity to affect a product comparison or a reported site result. Test the assumptions that carry the largest estimated shares. Recalculate using plausible alternative values supported by engineering, then check whether the result changes enough to alter the decision someone intends to make from those figures. That test can direct the next measurement investment. If a modest change in an estimated branch reverses which product appears more energy intensive, resolving that branch could matter more than improving an already well measured furnace supply. Sometimes the decision survives the uncertainty. If the plausible range leaves the same conclusion intact, record that range and its bases while keeping the estimate visible, rather than delaying every useful action until every branch receives a meter shows the balance with exceptions not forced agreement. The approved record should state the measured incoming quantity, accounted for consumption and remaining difference, with explanations strong enough that another reviewer can follow the result without inventing missing context. Only then is the energy balance ready for emissions calculations. Multiplying consumption by an emissions factor won't resolve an unexplained energy gap. It will carry that gap into another unit and leave the underlying question unanswered. An emissions factor needs its own managed record. Once the energy quantity is ready, the next task is selecting a factor whose coverage matches the activity, the reporting period and the calculation you're preparing. I'd assign that selection to a named reporting owner. Engineering can confirm what the factory consumed, while the person responsible for the emissions method checks whether the chosen factor represents the emissions that this particular calculation needs. Start with what the factor actually includes. An electricity factor might describe emissions from generation, while another data set includes upstream activities as well. So similar labels don't guarantee that two factors cover the same sources. 01:11:46 Speaker: Read the documentation behind the number. A factor covering fuel combustion alone answers a different question from one that also includes extracting, processing and transporting the fuel before it reaches the factory. That distinction belongs in the factor record store. The stated coverage in terms someone reviewing the calculation can understand, rather than leaving them to infer it from an abbreviated data set name or a downloaded file name. Greenhouse gases also need a common reporting unit. Carbon dioxide equivalent expresses the warming effect of greenhouse gases relative to carbon dioxide, using global warming potentials defined over a stated time horizon. You'll often hear that shortened to CO2 equivalent if the source already supplies a combined factor in that unit, don't apply another warming multiplier because the source has already performed that conversion under its documented method. Check which gases the factor covers. A factor labeled only for carbon dioxide might exclude methane and nitrous oxide, whereas a factor expressed as carbon dioxide equivalent can include those gases within its stated coverage. Preserve the conversion basis when the source provides it. Different assessment versions can use different global warming potentials, so retain that information rather than assuming every carbon dioxide equivalent figure uses exactly the same underlying convention. The fact is, geography needs a precise description to a national factor. A regional factor, and a factor intended for another electricity system aren't interchangeable simply because each comes from a recognized publishing body with clear documentation. Match geography to the activity and required method if the available data set doesn't cover the intended location. Document the permitted substitute and its limitation instead of silently treating the nearest available country as an exact match. Separate the factor year from its publication date. A data set published this year could describe energy supplied during an earlier year, and that distinction affects whether the selection fits the reporting period you're calculating. Revised publications need identifiable versions. Keep the publisher data set, title, release date and specific factor reference with a retained copy or permitted reference that lets another reviewer recover the entry you actually used. A web address alone can be fragile evidence. Publishers can update files at the same address. So your calculation should reference a fixed internal factor version rather than fetch whatever number happens to occupy that address later. Fuel calculations require another compatibility check. A gas quantity recorded as volume may need an approved calorific value to express its energy content, and the appropriate value can depend on the supplied gas and stated measurement conditions. Check whether the volume is already standardised. A supplier's quantity expressed at reference temperature and pressure shouldn't receive the same correction again, just because a separate conversion routine normally applies it to untreated meter volumes. Then confirm the fuel energy basis. Net calorific value excludes the heat recovered by condensing water vapour, while gross calorific value includes it, so consumption and the emissions factor must use compatible bases. The labels can vary between data sets. Net and gross calorific value are also called lower and higher heating value, which makes checking definitions safer than assuming two different labels necessarily describe two different quantities. Keep conversions as explicit calculation steps. Record the starting quantity, each conversion source and the resulting quantity used with the factor, so a reviewer can check the route without reconstructing hidden spreadsheet formulas or application settings. The output needs more than an emissions total attached. The activity reference factor, identifier, factor version, and calculation method to each result, allowing a later factor correction to identify exactly which results require review. Factor selection now becomes a controlled part of the calculation for purchased electricity. That leads to a further choice about reporting method, because the same consumption record can support distinct emissions results that need to remain separately identified. Purchased electricity can carry two separately named scope two results. Location based reporting describes consumption using applicable grid average factors, while market based reporting uses qualifying contractual information about electricity purchases under the guidance governing your inventory. Neither result replaces the electricity quantity recorded at the factory where applicable. Guidance requires. Both methods retain both results with their method labels, rather than choosing whichever number looks better in the annual disclosure or customer response. Location based accounting follows the electricity system serving the consumption for our hypothetical factory. That means applying the appropriate grid average basis to purchase electricity, regardless of whether procurement negotiated a particular supplier contract for that reporting period. Market based accounting follows qualifying contractual claims associated with that purchase, depending on the applicable rules and available market information. The calculation can use supplier specific factors or instruments that convey Electricity Generation attributes to the reporting organization. The contractual claim needs evidence behind it. A supplier describing a tariff as renewable does not by itself establish which emissions factor your organization can use, or whether the arrangement satisfies the requirements of your reporting method. Procurement needs to supply more than the contract's marketing description. The reporting owner needs documentation of the covered quantity, the generation attributes the relevant period and market, and how the organization holds the right to claim those attributes. Energy attribute instruments need particular care depending on the market. These can include certificates such as guarantees of origin with rules governing ownership, transfer and cancellation or retirement, so the same attributes aren't claimed repeatedly. Buying an instrument isn't the final accounting step. Check that it meets the applicable quality criteria, covers the relevant consumption, and receives the required cancellation or retirement treatment for your organization under the reporting arrangement being used. Supplier specific factors also need their supporting method. Establish whether the factor applies to your purchased supply and accounts appropriately for attributes sold or claimed elsewhere. Rather than assuming a supplier's published average automatically qualifies for your calculation. Uncovered consumption still needs a market based treatment where applicable. Residual mix exists. It accounts for electricity attributes remaining after claimed attributes are removed, helping avoid assigning those same attributes again to consumers. Without those claims, a residual mix isn't simply another name for the grid average. If suitable residual mix information isn't available, follow the fallback hierarchy in the applicable guidance and disclose that choice rather than quietly treating the methods as interchangeable. Consider what happens when our factory changes its electricity contract. A qualifying purchase could change its market based emissions result while leaving its location based result unchanged. Provided consumption and the applicable grid average factor remain the same for the comparison. That difference describes the accounting basis, not contradictory. Meter readings keep separate, calculated measures in the analytical model, so power BI can present each approved result without adding them together or letting one silently overwrite the other. Onsite generation creates another boundary to track if the factory installs solar panels, distinguish electricity generated electricity consumed on site, and electricity exported because the generation total alone doesn't establish how much purchased electricity production, displaced imports and exports need separate records, a net figure can conceal electricity purchased at one time and exported at another, leaving too little information to apply the intended accounting treatment to each flow across the site. Connection attribute ownership matters for on site electricity to. If the factory sells the attributes associated with its solar generation, consuming that electricity physically doesn't automatically entitle it to make the same renewable electricity claim as the buyer of those attributes. Exports don't automatically become a deduction from scope two. Emissions apply the relevant method to purchase electricity consumed and treat any separate claim about avoided grid emissions under its own stated basis, rather than subtracting it without explanation. Corporate electricity accounting still doesn't settle a product carbon claim. A company might substantiate a market based scope two result, but that doesn't establish that every product manufactured during the period can carry the same contractual treatment. The product method has to permit that treatment before passing a corporate electricity factor into an auto calculation. Check the applicable product rules and customer specification, including how they handle electricity attributes and their assignment to products. A supported company claim can still be insufficient for a product claim. That leaves a separate question before anyone sends an order level carbon figure to a customer, what exactly does the product calculation include and which rules govern it? I'd name the first deliverable quite narrowly, the product's factory energy contribution. That gives you a useful result without implying that the calculation includes the material entering the plant or everything that happens after the finished part leaves a product carbon footprint. Needs a stated system boundary. That boundary determines which activities belong in the assessment, so its coverage should come from the assessments, purpose and applicable method, rather than the availability of meter records for our metal parts. Purchase material could introduce another emissions contribution. If the selected boundary includes material production. The calculation needs information about that material and its supply route, not just the electricity used to machine it. Material quantity needs a clear definition. A calculation based only on finished part mass could miss material removed during machining depending on how the method treats material inputs, production losses and any recovered scrap, outsourced work needs the same attention. Suppose a coating supplier treats the part after it leaves our factory. A boundary covering manufacturer through that coating stage needs to address the supplier's contribution. Even though our meters never encounter it, transport can connect those production stages where the boundary requires it. Moving material to the plant or sending parts to the coating supplier introduces activities that need suitable quantities and emissions information in their own right. The chosen endpoint controls how far you follow the product. A cradle to gate assessment generally follows inputs from resource extraction through the defined production gate, while a cradle to grave assessment also addresses later stages such as use and end of life. Name the gate precisely for a part sent elsewhere for finishing factory gate could mean our dispatch point or the finishing supplier's dispatch point, and those endpoints leave different activities inside the calculation. The product quantity also needs a definition. Emissions per part leaves questions about the parts specification, condition and completion stage, especially when the same order contains variants with different dimensions or processing requirements. A declared unit states the quantity being assessed depending on the method, that might be one kilogram of a specified finished component at dispatch, with its included processes and condition defined so another person can interpret the result. A functional unit goes further by describing the service delivered. For a comparison between alternative components. That could mean providing a specified load bearing function for a defined service life under stated operating conditions, rather than comparing equal masses alone. That distinction affects whether comparisons make sense. Two parts can weigh the same, yet deliver different service lives. So a lower footprint per kilogram doesn't establish a lower footprint for the function the customer actually needs. Allocation choices belong in that product definition to the assessment needs to state how it divides burdens where processes support several outputs, including whether the applicable method requires avoiding allocation where practicable before applying a specified allocation basis, follow the chosen method consistently across included stages. A carefully supported factory energy split won't resolve a material data set that uses an incompatible boundary or allocation approach, so review those assumptions before combining the contributions into one product result. Set the reporting period and its intended meaning. An assessment representing a year's typical production differs from an order specific assessment, particularly if material suppliers process routes or operating conditions change during the period being represented. Background data may describe a different year, record that difference and assess whether the data still represents the technology, geography and period required, rather than labeling the entire footprint current because the production order carries today's date. Published guidance gives this work a defined structure. The Greenhouse Gas Protocol product lifecycle accounting and reporting standard addresses product level greenhouse gas inventories, while International Organization for standardization standards provide related assessment principles and requirements. You'll hear that organization shortened to ISO, ISO fourteen forty and ISO fourteen oh four four address life Cycle assessment and ISO fourteen oh six seven addresses product carbon footprints, but citing a standard doesn't by itself establish that your assessment meets its requirements. Check for applicable product category rules as well. These can narrow choices about boundaries, units, data and allocation, making a product assessment more consistent with others prepared under the same rules. Rather than merely sharing a standards name. That leaves a practical question for whoever receives the number. Is the customer asking for this defined product assessment? A manufacturing stage contribution or evidence for a separate disclosure requirement that needs a different response from the same underlying records. The request itself needs an owner before anyone starts preparing the response. Ask that person to obtain the customer's instructions, identify the intended use and resolve ambiguous wording with the requester, rather than letting the data team interpret commercial obligations. A questionnaire can contain several different requests in one document. Procurement might ask about your company's emissions, the electricity used at a named plant, and a footprint for a purchased component, with separate definitions hidden in accompanying guidance. Treat each requested measure as a separate reporting task. Give it an identifier and record the wording, the recipient, the deadline and any instructions that determine what an acceptable response actually contains. Then check what the customer will accept. They might permit an estimate with a stated method, require a specified calculation template, or expect independent verification, and those conditions change the work required before you can submit anything. Company disclosures follow their own requirements. Their reporting owner needs to establish which entities and reporting periods the obligation covers, along with the required disclosures, rather than assuming the customer's questionnaire defines the company's formal reporting duties. European requirements provide a useful example with a caution. The European Union's Corporate Sustainability Reporting Directive, often shortened to CSR and the European Sustainability Reporting Standards, or SRS, need an applicability assessment for the organization concerned. A factory's location alone won't settle that assessment. The responsible specialists need to check current legal scope, the company's position within its group, applicable national implementation, and any changes affecting when reporting obligations begin. Don't build a compliance schedule from an old presentation before committing delivery dates or system changes. Ask legal and sustainability specialists to confirm the current requirements and document the basis on which the organization considers them applicable. Reporting dates need to become project dates. If an obligation requires information for a particular financial year, the team needs time before submission to collect missing inputs. resolve exceptions, review calculations and obtain the required approvals. Assurance expectations affect that timetable. To ask the reporting owner and assurance provider where relevant, what procedures and supporting information they expect so the engineering team doesn't discover those expectations during the final review. A product request can impose a separate timetable and review route. Commercial teams need to confirm whether the customer expects a standard submission, a contract specific declaration or an independently checked assessment before promising that an existing company disclosure will satisfy it. Bring those requirements into a method register. This is a controlled record connecting each reported measure to its definition, its boundary, its source information, its calculation method, and the person accountable for approving its use. Keep the entries specific enough to resolve disagreements. Instead of listing energy data as the source, identify the approved data set and responsible source owner so the person preparing the response knows exactly where the required input comes from. The calculation reference needs the same precision point to the approved method for that measure, including any permitted estimates or exclusions, rather than describing it as a formula that the reporting analyst can adjust when a field proves difficult. Ownership should include the authority to accept limitations. An analyst can identify that a requested figure isn't supported by available information, but the designated reporting owner needs to decide whether disclosure of that limitation permits submission. Give every requirement a clear completion condition for a questionnaire. That might mean the customer accepts the stated method. For a formal disclosure, it means satisfying the applicable reporting and approval requirements, not merely filling every field. Conflicting requests should produce an explicit decision, not an improvised compromise. If two recipients require different treatments, maintain separately named measures with their corresponding methods and explain why the approved answers differ. Despite drawing on some common inputs, this register also gives the platform team a bounded assignment. They can prepare the required data sets and calculations against named definitions, while the reporting specialists remain responsible for interpreting obligations and confirming whether the resulting information meets them. Software can't take over that accountability. Azure and fabric can support preparation and delivery, but selecting a service or buying reporting software doesn't establish that a company meets a legal obligation or a customer's contractual requirement. Eventually, someone will question a submitted figure. After the deadline has passed, the agreed method Register tells them what the organization intended to report. Answering how that particular number reached the submission requires the supporting records to remain available. Pick one submitted order figure and ask someone else to reproduce it. Give them access to the approved records, but don't let them ask the original developer which settings applied or which file contains the final calculation. That exercise tests whether the evidence trail actually works. The reviewer needs to move backward from the reported emissions quantity through the selected factor. Allocated energy production matches and transformations until they reach the retained source readings. Each calculation run needs a persistent identifier. Associate that identifier with the input dataset version's processing code, version, configuration, and execution time. So rerunning the same job tomorrow doesn't erase which conditions produced yesterday's approved result. Preserve relationships between calculation stages, not just processing logs. A log confirming that a job finished won't tell a reviewer which source intervals contributed to an order, particularly when the job processed thousands of orders together. Store those contribution links in Queryable records for an individual order. The reviewer should be able to retrieve its contributing intervals and allocation entries without searching every application log from the month when production occurred. Versions need to cover rules as well as code if the business changes. How it distributes furnace preparation energy. Assign the revised method an identifier and approval record rather than editing the existing methods. Description in place record when the change applies. A method approved in October might apply from November, while a correction approved in October might repair results dating from August. So approval date and effective date answer different questions. Equipment mappings need that same distinction. Suppose the electrical team discovers that a circuit assignment entered last month describes an installation change completed several weeks earlier. The model needs both the corrected effective period and the corrections recording date. 01:30:09 Speaker: Keep the superseded mapping available. Otherwise, someone reviewing the earlier report can't determine whether its result followed the information approved at the time, or came from a processing error against that information. Closed reports need a reproducible snapshot. Preserve the approved reporting data set alongside its calculation, references, unresolved exceptions and approval record so later source corrections don't silently change a figure already sent outside the organization. A snapshot doesn't have to duplicate every underlying file. It does need fixed, recoverable references to the relevant versions, with retention controls that prevent routine cleanup from removing inputs required to reproduce the approved submission. Corrections should enter a review workflow when a late production amendment changes an order's result, identify the affected submissions and calculate the difference before deciding whether the change requires an internal adjustment or an external restatement. Let the reporting owner authorize that decision. Keep the original submission, the corrected result, the reason and any replacement communication connected, rather than overwriting the original and leaving the recipient holding a number your system no longer recognizes. Access controls protect this history from accidental or unauthorized changes. Separate the ability to submit plant corrections, maintain calculation methods, run processing, and approve reporting results with permissions reflecting those responsibilities across Azure and Fabric. Use Microsoft Entra identities and supported workload identities where appropriate. Assign narrowly scoped permissions through the relevant Azure and Fabric controls. Because permission to manage a workspace or resource doesn't automatically describe the business authority someone should hold. Plant and business roles need different right permissions. A maintenance engineer might document a meter replacement without being allowed to approve an emissions factor, while a reporting specialist might approve a disclosure without editing the retained meter payloads. Test those boundaries using actual role assignments, fabric workspace permissions, item access, and access to underlying storage need review. Together, since restricting one reporting route won't help if the same person can modify its source through another record, privilege changes and review them. Administrative access might remain necessary for recovery, but its use should leave evidence that identifies who changed protected records or settings, and supports investigation when an approved result becomes difficult to reproduce. Power BI should consume those approved definitions rather than recreate them. Carry the reporting version, approval status and calculation references into the dataset used for analysis, giving each published measure a route back to the evidence supporting it. That makes a correction an explicit change instead of a silent refresh. With the approved definitions available to power BI. Production can ask questions of the numbers without each report introducing another interpretation of how those numbers were calculated. I'd begin with a question of production manager can act on which completed orders used more electricity than comparable work, gives the power BI model a clearer purpose than providing another running total for the whole factory. The shared semantic model defines how those questions get calculated. In practical terms, it connects prepared data sets and approved measures, so reports for production, finance and sustainability use consistent definitions while answering different questions about the same work. Keep the measurement levels distinct within that model. Site Side consumption belongs to a site and period, while attributed order consumption belongs to production work, so selecting an order shouldn't silently turn the site total into an order measurement. The relationships need deliberate testing. If an order has several operations, joining its energy records to every operation row can multiply quantities, producing a convincing report whose totals grow simply because someone requests more production detail. Cost and emissions need separately named measures as well. A production user should be able to ask about electricity quantity without inheriting a finance filter that restricts the answer to priced records, excluding valid consumption that hasn't received its cost calculation. Missing information shouldn't behave like a measured zero. If an order has energy but lacks an approved emissions result, the model should expose that incompleteness rather than imply the order produced no energy related emissions during manufacture. Ratios deserve particular attention because totals can appear reasonable while comparisons fail for energy per good part, calculate the relevant energy total divided by the corresponding good part total within the selected population instead of averaging individual order ratios. Here's a different hypothetical example. One order uses twenty kilowatt hours for ten good parts, while another uses ninety kilowatt hours for ninety good parts, giving individual ratios of two and one kilowatt hours per good part. A simple average gives one and a half across both orders, though one hundred and ten kilowatt hours divided by one hundred good parts gives one point one because the larger order contributes far more output, that isn't a special power by problem. The measure needs to divide summed quantities in the current selection, and it also needs to reject combinations where adding the part counts doesn't support a meaningful production comparison. Product mix belongs in that selection. I'd let production compare a product family with its own history before combining families with root or variant detail available, where those differences materially affect the work performed. The comparison period needs equal care. A completed week and a partly completed week don't offer equivalent evidence, especially when energy has already accumulated for operations whose accepted output won't be recorded until the next reporting period. Operating state gives the comparison another useful dimension separating processing, preparation and waiting lets the production team ask whether a change comes from the active operation or from how the resource spent the rest of its available time. Keep those categories consistent between the periods being compared. Otherwise, an apparent reduction in processing energy might come from a recording change that moved consumption into another state without any corresponding change in equipment behavior. I'd also make the comparisons coverage explicit if users select orders completed this week, clarify whether their energy includes work performed before this week, because that order based population differs from electricity consumed during the week itself. Freshness needs more detail than the last report. Refresh power BI might refresh successfully, while the production feed remains several hours behind the meter feed, leaving current consumption paired with incomplete execution or quantity. Records give users the latest usable time for each contributing source. They need a plain statement that production matching is complete through a particular point, alongside any provisional period. Still waiting for records or checks to finish? Quality status should travel with the answer they requested for a selected order or period, identify missing coverage and estimated contributions so users can judge that result. Rather than rely on a general assurance about the whole data set. This is where a report becomes useful in the production meeting. Someone can isolate a comparable population, check its coverage, and identify which operating conditions accompany the change without treating every movement in a ratio as proof of waste. Finding that pattern still leaves a decision unresolved. Power BI can help establish where consumption differs and which work is involved. But choosing what production should change requires examining the consequences beyond the measure itself. Return to the furnace with a proposed change. Wait for fuller batches. Suppose the production team identifies compatible work arriving later that day, and considers holding the next load so those parts can enter the furnace together. The energy argument sounds sensible. Combining compatible work could reduce the number of heating cycles needed for those orders, but waiting also consumes time that the rest of the production plan might already depend on. One order might need to leave for coating that afternoon. Missing its collection could delay delivery, even if the furnace completes the combined batch before the end of the shift and its own electricity result improves, the decision needs a wider production window. Ask what happens after heat treatment? Which commitments depend on completion, and whether the proposed weight pushes another operation beyond its available capacity or working calendar. Batch compatibility also sets limits. Parts sharing a furnace must meet the relevant process requirements, including approved treatment conditions and loading restrictions, rather than simply fitting inside the available physical space. At the same time, a fuller load isn't automatically an acceptable load. Quality requirements can constrain the arrangement and quantity of parts, so any proposed consolidation must respect the validated process rather than treat maximum occupancy as the operating target. This is a scheduling problem with an energy objective. Advanced planning and scheduling, or APS technology can construct and compare production schedules against modelled resource availability, operation sequences, and other constraints relevant to that particular implementation. The constraints determine whether its answer deserves consideration. If the model omits the coating collection or a furnace loading restriction, a mathematically feasible schedule can still fail. The practical test that the production planner applies. I'd ask the planner to compare specific alternatives, run the next compatible batch now wait for the additional parts, or use another approved resource. Then examine how each option affects energy completion times and downstream work. Each alternative needs the same business commitments. The ERP supplies, the order context, and recorded delivery requirements, while its transactions remain the controlled business record. Rather than becoming something the energy calculation independently rewrites the mess, closes the loop with execution feedback. When machining finishes later than expected or a furnace becomes unavailable. The scheduling process needs that updated condition before its next recommendation can represent the work still achievable. Those exchanges require integration, not just shared access to storage. Someone must define which order changes reach the scheduler, how accepted schedules return to execution systems, and which status changes require the planner to reconsider the plan. Microsoft Services can support those information exchanges. Fabric can supply prepared analytical inputs, while Azure can host integration or specialist applications. But the scheduling logic and manufacturing constraints still need appropriate software configuration and ownership. Tariff aware scheduling introduces another objective if prices vary during the available production window, a scheduler can evaluate moving eligible work into cheaper periods while respecting the capacity and delivery limits already represented in the model. Cheaper periods might attract several large loads. The proposed schedule therefore needs to evaluate applicable demand charge consequences as well as consumption prices. Otherwise, moving each machine independently could increase the cost of their combined operation. Carbon aware scheduling needs a separately defined signal where suitable time resolved emissions information is available. A planning model can compare alternative operating periods while keeping forecast operational signals distinct from the factors approved for formal reporting. The cheapest schedule might differ from the lowest emissions schedule. Management needs to specify how those objectives compete with delivery performance, because the scheduler can't infer which compromise the organization will accept when the objectives point toward different choices. Set boundaries on the trade offs before selecting an option. A planner could require all firm delivery commitments to remain protected, then compare costs or emissions among the remaining feasible schedules, rather than quietly trading late orders for better energy figures. Simulation helps test consequences. Before committing the plant, a model can examine queue buildup, downstream congestion and the effects of uncertain completion times, exposing weaknesses that a schedule built from fixed durations might overlook. Specialist optimization can then search the permitted alternatives more systematically. Its recommendations still depend on the modeled constraints and objectives, so review the proposed sequence with production before issuing changes to the systems that coordinate actual work. A revised furnace sequence changes more than a reported ratio. It changes when material moves, when people need to be ready and which commitments remain achievable. so the selected option needs a checked path into execution. AI can help evaluate the conditions surrounding that execution decision. Its role needs a clear boundary, though, especially when an explanation about energy use might influence a change that reaches equipment or alters an operator's instructions. Load forecasting is a useful place to apply it. A predictive model can estimate expected electricity demand from upcoming work and operating conditions, giving the energy team an early indication of periods that might need attention. The forecast should include uncertainty, not just one predicted number. If the upcoming work differs substantially from the model's training examples, flag that limitation so the team knows when the prediction needs closer engineering. Review before use. Unusual consumption detection addresses a different question. A model can identify readings that depart from expected behavior, helping maintenance focus its investigation without claiming that the departure proves a particular fault or justifies an immediate intervention. Consider unexpected compressor consumption during an otherwise quiet shift. A leak is one possibility, but the production context might reveal additional equipment preparing for work, an extended furnace warm up or rework that wasn't present in comparable periods. Those conditions change what counts as unusual. A detection system needs relevant operating context to distinguish unexplained consumption from consumption that follows a changed product mix or a legitimate activity outside the usual shift pattern. The alert should express what the evidence actually supports. Consumption exceeds the expected range during this operating state. Gives maintenance something to investigate. While compressed air leak detected requires additional evidence identifying that cause rather than merely observing excess electricity. Generative AI can help people question those approved records. Someone could ask why an order's attributed energy increased, and an assistant could retrieve the relevant operating events, quality records and calculation references before drafting an explanation. Retrieval needs permissions and a defined data boundary. An assistant used by a production supervisor should answer from authorized records with the same plant and business restrictions that apply when that supervisor accesses the underlying information directly. Ask it to identify evidence separately from interpretation. If records show additional processing, the answer can cite those records. If they don't, establish why consumption rose. The assistant should state that the cause remains unresolved. A fluent answer still needs verification. References should lead to records that support the particular claim, and the application should avoid presenting a generated explanation as approved merely because its inputs came from approved data sets. Recommendation and execution require another boundary. Generative AI can help express a question or explain a candidate action, while a constrained optimization engine checks proposed choices against the explicitly modeled limits governing that decision. Neither component should invent missing operating permissions if an action requires engineering authorization. An unavailable measurement or confirmation from the shift lead, the workflow should stop at that requirement rather than infer permission from a plausible explanation. Human approval needs enough detail to be meaningful. Present the proposed action, its supporting evidence, expected consequences, and unresolved assumptions, then route it to someone with authority and time to assess the physical change being requested. Approval isn't a substitute for local protection equipment interlocks, validated control sequences and safety functions remain responsible for enforcing their designed limits. Regardless of what an analytical service recommends or a business workflow approves. Azure hosted analysis can remain outside the direct control path. A supported integration can deliver an approved instruction through the plant's established process with acknowledgement and failure handling, rather than letting generated text become an equipment command, automation can still remove routine administrative work. It can assemble an investigation record, notify the responsible engineer or prepare an approval request, provided each task has defined inputs and a clear response when supporting evidence is missing. I test that boundary in a limited first rollout. Choose a narrowly defined advisory use where the team can inspect the evidence and reject the recommendation before considering any connection to actions that affect production. For the first release, choose an area with people available to verify the result. Use known meters, one production feed and one allocation rule. Then give the team a specific order calculation to accept or reject before extending coverage. Write acceptance criteria before the trial starts for a completed order, the reviewer should be able to confirm its energy share, identify any estimate and recover the supporting records without help from the developer. Make the test pack part of the deliverable, include representative production records and deliberately introduced faults with expected outcomes agreed by the meter owner, production owner and reporting owner. Before anyone runs the tests, an outage test needs a checked result after recovery. Compare the recovered dataset with the records retained locally, then confirm that the resulting consumption totals match the expected quantities without omissions or additions caused by replay. Counter resets need a different acceptance condition? Introduce a documented reset into a test stream and check that processing. Identifies the discontinuity, isolates the affected period, and avoids publishing an apparently valid order total across it. Test missing intervals through the reporting layer to the person reviewing the order should encounter the agreed incomplete or estimated status, rather than needing to inspect pipeline logs to discover that it's supporting measurements aren't complete. A corrected order should change only the intended results. Amend one execution record in the test environment, identify the expected affected calculations beforehand, and compare that list with what the recalculation actually revises in the reporting data set. Check unaffected orders as carefully as the corrected order. If changing one completion time alters unrelated results. Investigate the dependencies before accepting the release. Even when the corrected order now displays exactly the number everyone expected. Put operating costs through a trial as well. Measure normal collection backlog recovery, processing and reporting activity using the proposed configuration. Then turn those observations into a monthly estimate with the assumptions stated beside it. Device costs extend beyond the purchase price include installation, replacement communications charges where applicable, and the effort needed to maintain the gateway. Because a low cloud bill won't compensate for an installation that repeatedly requires specialist visits. Count the message flow. The system actually produces polling, batching, retries and recovery affect the workload. So compare observed traffic with the selected service limits and charging bases. Instead of budgeting from meter count alone, retention and fabric capacity need representative workloads? Run the planned queries and reporting refreshes while processing continues, then assess storage growth and capacity use under those combined conditions rather than testing each task separately in isolation. Support effort needs a place in the budget. Ask the intended support team to diagnose a failed feed and restore a failed processing run using the supplied instructions, recording where they need undocumented knowledge or developer assistance. The trial also needs a fair performance baseline. Agree the comparison method before changing operations using production volume and product mix to distinguish the effect of the intervention from changes in what the area actually produces. Lower consumption alone won't establish savings from the project. Compare observed consumption with the baseline expectation for the work performed and state how outages, missing data or unusual production conditions limit the strength of that comparison. Keep savings claims separate from technical acceptance. The architecture might pass every data test before the plant has operated long enough under comparable conditions, to support a defensible claim about reduced energy use or expense. Expansion should follow an acceptance decision with named owners before adding another area. Confirm who handles meter changes, failed feeds, allocation reviews and calculation corrections, including what happens when the usual contact isn't available during a shift. Approve the next area because the first one supports a trusted decision. Carry its test pack and operating instructions forward, but require fresh confirmation of local conditions rather than assuming another line behaves like the one you've just proved. Every order level carbon number carries a measurement boundary and an allocation choice. If you can't explain those choices, a customer can challenge the product claim even when every cloud service runs correctly and every calculation completes without an error. Where does that defensible path break in your plant today? I'd like to compare approaches on the hardest gap you're tackling, whether that's meter coverage, production links, or allocation rules.