Decision-first analytics: choose the method before the model

1. The problem this lesson solves

In practice, carbon and energy teams rarely lack software options. They lack a shared answer to a prior question: which authorised decision will change if we buy or build this tool? Vendors often arrive with a platform that promises invoice extraction, purchase classification, missing-meter imputation, anomaly detection, forecasting and production scheduling at once. The pitch sounds decisive. The decision file is usually empty.

This lesson trains a professional habit that Unit 1 already used for policy: name the constraint before you name the instrument. Here the “instrument” is analytics or AI. You will learn to separate observation, calculation, inference, recommendation and decision; to climb a method ladder from formula to machine learning only when simpler methods fail; and to read forecasting and AI-energy evidence without mistaking a published model for a plant procedure.

Real industrial cases show that AI can cut energy when the decision is narrow and measured — for example paintshop ventilation optimisation reported in the Stellantis / etalytics case below. Grey analysis from the International Energy Agency likewise argues that industrial AI can save energy at scale while AI infrastructure itself raises electricity demand. Those two facts belong in the same brief.

2. Start with the decision, not the technology

Carbon work collapses when teams confuse layers that must stay separate. An observation is what was recorded (a meter reading). A calculation is what follows from an agreed formula (activity data × approved factor). An inference estimates an unknown state or future (tomorrow’s demand). A recommendation proposes an action. A decision is what an authorised person or system will actually do. An outcome is what changed afterward.

A decision statement should be specific enough to test. You can use this template: [Decision owner] must choose [action] for [unit and time horizon] to improve [objective], subject to [constraints], using [available evidence], before [deadline], with [fallback and review]. Before procurement, you should complete a short decision canvas covering owner, action, objective, constraints, unit, horizon, evidence, error costs, authority, fallback and learning loop.

3. A method ladder for carbon decisions

You should use the least complex method that is adequate and defensible. Human structured review fits novel, low-volume or value-laden choices — still with criteria and documentation. Manual does not mean uncontrolled.

A deterministic formula returns the same output for the same inputs and version. Prescribed Scope 1 combustion accounting almost always belongs here: activity data × approved factor, with unit conversion and GWP where required. AI may help collect or quality-check inputs; it should not silently replace the method. Business rules encode IF–THEN logic when criteria are stable and contestable. Descriptive and diagnostic analytics answer what happened and why — often without prediction.

When the future is genuinely unknown, forecasting and statistical models become candidates — and they should beat simple baselines such as yesterday, seasonal average or last year. Reviews by Yildiz, Bilbao and Sproul (2017) and Ahmad, Zhang and Yan (2020) map how electricity-load and renewable-requirement forecasting methods perform under different data regimes; Akhter et al. (2019) do the same for photovoltaic generation. Closer to the plant clock, Tarmanini et al. (2023) compare ARIMA and neural approaches for short-term load forecasting, while Oreshkin et al. (2021) show that specialised deep architectures can help mid-term load planning — if the decision, baseline and monitoring path are defined.

Optimization chooses actions under constraints (for example cost plus a chosen emissions weight). The weight is a governance choice, not a mathematical truth. Machine learning estimates patterns from examples when relationships are too numerous or nonlinear to encode economically. Hybrid systems are normal: extract → validate → calculate → display → human approve.

4. Enough AI literacy to choose wisely

You do not need to train neural networks to govern them. You do need a few distinctions. Artificial intelligence is the broad field; machine learning learns behaviour from data; a model is the fitted representation that turns inputs into outputs. Training fits parameters; inference uses a deployed model; monitoring watches for drift and misuse after go-live. Good performance on training data is not the goal — relevant new cases are.

Leakage is a special trap: if evaluation uses information unavailable at decision time, or related records cross train and test splits, the score is fiction. Symbolic rules are transparent and brittle; learned models are flexible and data-dependent. Dependable carbon systems usually combine both.

Documentation matters as much as accuracy. Mitchell et al. (2019) propose model cards — structured statements of intended use, limits and evaluation conditions — that a carbon team can demand from vendors. NIST’s AI Risk Management Framework frames trustworthiness as contextual and risk-managed, not as a badge on a slide.

Meanwhile, Turkish industry is already meeting AI in carbon workflows: closed-corpus assistants such as KarbonBot for SKDM-facing guidance, and OCR-assisted invoice capture feeding GHG calculations in the Brisa–Enerjisa collaboration. You should treat those as decision-support layers, not as substitutes for method, authority or audit trail.

5. Failure modes that matter on the plant floor

These patterns rarely appear as dramatic headlines. They appear as confident dashboards that cannot survive a verifier’s question.

A proxy target optimises spend or a convenient KPI instead of physical emissions. Historical bias teaches yesterday’s inconsistent manual labels as if they were truth. Class imbalance can produce 99% accuracy while missing rare refrigerant leaks. Leakage inflates tests; drift arrives when tariffs, product mix, meters or processes change after go-live. Automation bias lets staff rubber-stamp high-confidence outputs. Objective gaming improves the metric while absolute energy or displaced emissions worsen. Hidden integration errors — MWh read as kWh, wrong timestamps, silent unit conversion — defeat even a correct model. Generative fabrication invents factor sources or policy clauses in fluent prose.

A related risk is carbon leakage through optimization: shifting production may cut location-based electricity emissions while raising transport or supplier emissions. You should state the boundary explicitly and track material displaced effects. Turkey-facing emission outlook work such as Bakay and Ağbulut (2021) also reminds us that electricity-mix forecasts are themselves uncertain — so a schedule optimiser that treats tomorrow’s grid factor as a known constant is making a modelling choice, not discovering a fact.

Market-price models such as Lu et al. (2020) on carbon allowance prices belong in a different decision class again: they may inform trading or budgeting, not Scope 1 combustion arithmetic.

6. Architecture of a governed decision-support system

The model is only one layer. A reliable system separates source evidence (meters, invoices, factors with owner, timestamp, unit and quality flag), processing (versioned transformations and reconciliation), prescribed calculation, the model itself, decision policy (abstention, escalation, constraints), the human interface, the action interface (authorisation, logging, rollback) and monitoring.

When those layers blur, responsibility dissolves. When they stay distinct, a plant can use OCR, forecasting or optimisation without pretending that fluency equals assurance. That architecture is what lets summer-school workshops and MOOC readings talk to the same plant file.

7. What this means for sequencing (now → later)

Now. You should write decision statements for your two highest-priority carbon or energy choices. You should map each proposed digital function to formula, rule, analytics, forecast, optimisation, ML or human policy, and refuse any function that lacks owner, baseline, fallback and monitoring.

Next. Lesson 4.2 asks whether you can reproduce the number — lineage, meters, missing data and integrity. Lesson 4.3 asks whether digital support is trustworthy enough for rights, oversight and procurement.

Later (Unit 3). The same evidence chain becomes a cyber-physical asset: sensors, historians and claim integrity under attack and recovery.

8. Applied case: LumaWorks Materials

LumaWorks Materials is the continuing workplace case for Units 4 and 5. It is a fictional manufacturer of engineered insulation panels with two natural-gas process ovens, purchased electricity, rooftop solar, PLC-controlled lines, SCADA, an OT historian, an energy-management application, ERP and a cloud carbon-accounting platform.

Management wants monthly emissions estimates one day after month-end, and lower energy use without harming product quality or safety. A vendor proposes an “AI sustainability copilot” covering invoice extraction, purchase classification, missing-meter imputation, anomaly detection, energy forecasting and production scheduling — one platform, one dashboard, rapid decarbonisation. The proposal does not define which decisions will change, who remains accountable, what simpler alternatives were tested, or how errors will be detected.

9. References

• International Energy Agency. Energy and AI; Key Questions on Energy and AI. https://www.iea.org/reports/energy-and-ai
• NIST. Artificial Intelligence Risk Management Framework (AI RMF 1.0). https://www.nist.gov/itl/ai-risk-management-framework
• Mitchell, M., et al. (2019). Model Cards for Model Reporting. https://doi.org/10.1145/3287560.3287596
• Yildiz, B., Bilbao, J. I., & Sproul, A. B. (2017). A review and analysis of regression and machine learning models on commercial building electricity load forecasting. Renewable and Sustainable Energy Reviews. https://doi.org/10.1016/j.rser.2017.02.023
• Ahmad, T., Zhang, H., & Yan, B. (2020). A review on renewable energy and electricity requirement forecasting models for smart grid and buildings. Sustainable Cities and Society. https://doi.org/10.1016/j.scs.2020.102052
• Akhter, M. N., et al. (2019). Review on forecasting of photovoltaic power generation based on machine learning and metaheuristic techniques. IET Renewable Power Generation. https://doi.org/10.1049/iet-rpg.2018.5649
• Tarmanini, C., et al. (2023). Short term load forecasting based on ARIMA and ANN approaches. Energy Reports. https://doi.org/10.1016/j.egyr.2023.01.060
• Oreshkin, B. N., et al. (2021). N-BEATS neural network for mid-term electricity load forecasting. Applied Energy. https://doi.org/10.1016/j.apenergy.2021.116918
• Bakay, M. S., & Ağbulut, Ü. (2021). Electricity production based forecasting of greenhouse gas emissions in Turkey with deep learning, SVM and ANN. Journal of Cleaner Production. https://doi.org/10.1016/j.jclepro.2020.125324
• Lu, H., et al. (2020). Carbon trading volume and price forecasting in China using multiple machine learning models. Journal of Cleaner Production. https://doi.org/10.1016/j.jclepro.2019.119386
• Cao, Z., et al. (2022). Toward a Systematic Survey for Carbon Neutral Data Centers. IEEE Communications Surveys & Tutorials. https://doi.org/10.1109/comst.2022.3161275