AI in Energy: Separating Forecasting, Decisions, and Control

As of July 30, 2026, forecasting tomorrow’s demand and deciding how to charge a battery are not the same problem. Forecasting produces a distribution over an uncertain future; optimization chooses an action under constraints; control corrects when reality departs from the plan. Blending them into one promise of “AI for energy” prevents evaluation. The useful skill is following that chain and identifying the cost of each error.

Start with the decision and work backward

An operator may need minutes to balance frequency, hours to schedule generation, or years to plan capacity. Each horizon changes inputs, resolution, and tolerable error. Geography matters too: a home, substation, and country aggregate different behavior.

First write who will act, when, and against which alternative. If a solar forecast sets reserve capacity, underprediction may waste energy while overprediction may leave a deficit. The loss function should reflect that asymmetry, or evaluation should at least separate it.

The seasonal baseline is the first gate

Electricity demand contains hour, working day, season, and weather. Before an LSTM or boosting model, compare “same as yesterday,” “same as last week,” and a regression using calendar and temperature. The open textbook Forecasting: Principles and Practice explains baselines, temporal validation, intervals, and residual diagnostics.

A complex model that cannot beat a seasonal rule under the same cutoff has not demonstrated additional capability. A baseline also detects experimental failures: an enormous improvement may indicate that the model saw future data or that the reference was artificially weak.

Split in time; do not shuffle

Shuffling a series mixes past and future. Evaluation should train up to one date and predict the next block. In rolling-origin validation, the cutoff moves and the exercise repeats, exposing seasons, tariff changes, and extreme events across periods.

Every transformation follows the same clock: means, scalers, imputation, and selection are fitted only on the available past. Weather variables must distinguish observation from forecast. Using the weather that eventually occurred to evaluate a day-ahead prediction leaks information.

A bidirectional LSTM may use earlier and later context inside an already observed sequence, but it cannot consult an operational future that does not exist. If an experiment permits that view while predicting beyond the cutoff, the gain is leakage rather than temporal sensitivity.

Demand and renewables need traceable inputs

Demand inputs include calendar, temperature, activity, and lagged values. Solar uses irradiance, cloud, orientation, and equipment state; wind uses speed and direction at relevant heights plus the physical power curve. NREL’s NSRDB provides documented solar and meteorological data; the EIA open API publishes United States energy series. Version, time zone, resolution, and revisions must be preserved.

More variables do not guarantee a better forecast. A sensor may arrive late, an observation may be revised days later, and an aggregate may change definition. Each feature record needs availability time, unit, and missing-value policy.

A point model is one part of the map

Gradient boosting can capture nonlinear relationships in calendar-weather tables; recurrent or convolutional networks can learn sequence patterns; a Transformer can relate horizons and variables. DeepAR modeled probabilistic distributions across related series. Temporal Fusion Transformers combined static inputs, known future inputs, and observed inputs for multi-horizon forecasting.

Names do not replace an experiment. Fix the same history, horizon, variables, and budget, then compare error by season, hour, asset, and demand level. An architecture wins only inside that perimeter.

Uncertainty is an output, not a footnote

One value hides that two hours with the same prediction may carry different risk. Intervals or quantiles support reserve planning. Pinball loss evaluates quantiles; coverage checks how often reality falls inside an interval. Width matters too: an enormous interval can cover everything without helping a decision.

Calibration should be checked by horizon and condition. During a storm or heat wave, intervals should recognize increased uncertainty. If the model remains confident precisely when data leaves distribution, optimization receives false precision.

Metrics must preserve units and costs

MAE expresses average absolute error in megawatts or kilowatts; RMSE penalizes large errors more heavily. MAPE divides by the actual value and becomes unstable near zero, common in nighttime solar generation. An aggregate percentage can hide a critical failure at dawn or during peaks.

Report bias—overprediction versus underprediction—error quantiles, and extreme episodes alongside averages. Translate error into reserve, curtailed energy, violation, cost, or emissions. The best metric is the one that clarifies consequences, not the one producing the smallest number.

Optimization means choosing under physical constraints

The forecast enters a different problem. An objective may minimize cost, emissions, or deviation, but it must obey supply-demand balance, line limits, generator ramps, state of charge, battery power, and reserve requirements. Without constraints, an algorithm can recommend an impossible action.

PyPSA documents an open framework for optimizing and simulating power systems with generation, storage, and networks. Its educational value is making variables and constraints explicit. An ML model may supply demand, renewable availability, or approximations, while the energy problem retains structure.

Robust and scenario optimization

Planning only with the mean ignores tails. Scenarios sample plausible trajectories of demand and renewables; stochastic optimization considers their distribution. Robust optimization seeks decisions valid across an uncertainty set. Both expose the price of protection.

A probabilistic forecast is wasted when optimization discards its distribution. End-to-end evaluation compares cost and violations under realized future data: baseline forecast plus optimizer, new model plus the same optimizer, and perhaps a perfect-information plan as an unattainable bound.

Reinforcement: a policy is not permission

Reinforcement learning can learn a charging or control policy from reward. But DQN, introduced in the original video-game work, does not provide power-grid guarantees. State, action, reward, simulator, and constraints require specification.

If low price is rewarded without battery degradation or comfort, the agent will exploit the omission. Actions should be projected into a safe set, a fallback controller should exist, and tests should begin in simulation and shadow mode. Online learning changes the deployed policy; every update needs a gate and rollback.

Explanation does not repair a poor forecast

LIME or SHAP can display local sensitivity or attribution under assumptions, but they do not prove causality or stability. Energy systems also need response curves, residual analysis, and physically plausible counterfactual tests. If increased temperature reduces demand in a condition where it should raise it, data and model need investigation.

Operational transparency includes something simpler: training date, sources, horizon, available variables, valid range, intervals, version, and fallback. An operator needs to know when not to use a model, not only which feature received the highest attribution.

The loop connecting prediction and operation

At every cutoff: collect only available data; run baseline and model; produce value and uncertainty; solve optimization with constraints; approve or limit actions; observe the outcome; attribute cost to forecast error, physical model, or execution; monitor drift; promote a new version only after temporal evaluation.

The durable skill is drawing three boxes—forecast, optimization, and control—and demanding a measurable interface between them. The first does not decide, the second does not predict, and the third cannot ignore the world. An intelligent energy system is not the one with the newest algorithm; it is the one that acknowledges uncertainty and keeps every action inside a physically possible grid.

Finally, audit the clock itself. Record when every observation was measured, when it arrived, when the forecast ran, and when an action became effective. A technically accurate forecast delivered after the dispatch deadline has zero operational value. Latency and revision history therefore belong beside accuracy in the final scorecard.