Materials & Energy

Before Funding Energy AI, Ask It to Beat Last Week

2026-09-10

A sophisticated energy-demand model should first outperform a clearly defined seasonal baseline on untouched future periods. The comparison must use the same forecast horizon and information available at the time. Existing load and weather records can reveal whether the proposed improvement survives calendar effects, timestamp issues and ordinary seasonal structure before a team invests in a more elaborate AI system.

A simple baseline is a commercial question

If demand at a particular hour often resembles demand at the same hour last week, repeating last week's value is a useful hurdle. Forecasting: Principles and Practice describes seasonal naive methods that carry forward observations from the corresponding season. Such a baseline is not a claim that energy systems are simple. It measures how much a new model adds beyond an obvious pattern.

The research question should be phrased in terms of the decision. Does the proposed model improve day-ahead planning enough to justify its data requirements and maintenance? An intuition about a hidden weather interaction can become a candidate feature. But the intuitive explanation earns attention only if its added value remains visible after the calendar baseline is accounted for.

Sources: Forecasting: Principles and Practice, simple forecasting methods.

Audit the clock before the architecture

Open Power System Data provides documented time-series packages that include electricity-related series and metadata. A public-data demonstration can begin there, with the exact release and geographic series frozen. The inputs for a specific analysis still need inspection: units, sampling interval, missing periods, revisions and whether a column is actual or forecast load.

Align time zones and daylight-saving transitions explicitly. Decide how repeated and absent local hours are represented. Check publication lag for every explanatory variable. A weather measurement recorded after the forecast deadline cannot legitimately improve a forecast issued before that time. These mundane details can determine whether a reported improvement is real or simply access to future information.

Sources: Open Power System Data: Time series.

A hypothetical improvement with a smaller practical meaning

Imagine a weekly seasonal baseline with a mean absolute error of 100 megawatts and a complex candidate at 95 on a fixed historical holdout. That is a five-percent reduction in this metric, not a five-percent energy saving. If the candidate has a much larger error during the specific peak hours relevant to the buyer, it may still be less useful for the stated decision.

Now suppose a calendar-adjusted linear model scores 96 under the same setup. The complex model's advantage over a credible alternative is one megawatt, not five. These invented numbers illustrate why the comparator matters. The report should show incremental value over both a minimal baseline and a reasonable compact model rather than celebrating an isolated score.

Run the smallest comparison that can change the decision

Choose one forecast horizon and reserve later time blocks before feature exploration. Compare weekly seasonal naive, a transparent calendar model and the proposed candidate. Fit missing-value treatment and transformations inside each training period. Evaluate the same timestamps for all models and record cases where a method cannot produce a prediction.

Break down errors by weekday, weekend, season and predefined peak periods. Add an ablation that removes the intuitively proposed feature while keeping the rest of the pipeline constant. If the claimed mechanism is a special weather interaction, this comparison asks whether the interaction contributes beyond simpler weather terms. It does not prove causality, but it gives the explanation a concrete predictive burden.

Decide what counts as enough

A useful stop condition is failure to beat the compact baseline by a pre-agreed, decision-relevant margin across the required periods. Another is an improvement that relies on information unavailable by the forecast deadline. Do not keep changing the test interval until a favorable season appears. Record uncertainty and the number of independent periods supporting the comparison.

A successful backtest does not establish future savings, trading profitability, grid safety or robustness to structural changes. It supports a narrower statement about historical predictive performance. If the business value depends on a downstream decision rule, that rule needs its own offline evaluation rather than assuming that every decrease in forecasting error has equal economic value.

What to ask for in the first research proposal

The first proposal should specify the horizon, timestamp rules, baseline family, evaluation periods and the decision margin that would justify further work. Deliverables should include a clean data manifest, reproducible predictions and an error report with adverse slices visible. A model checkpoint without these materials is difficult to assess and expensive to trust.

This kind of computational challenge is particularly useful for teams that already have an appealing AI narrative. It can distinguish a genuine unresolved modeling opportunity from a calendar pattern dressed in expensive software. Faster progress means finding that distinction early, while the project can still change direction cheaply and without claiming a result that has not been demonstrated.

Questions this raises

Does a five-percent forecasting improvement mean five-percent savings?

No. Economic effects depend on the downstream decision, costs and which errors changed.

Can this start without private energy data?

Yes, with a public-data demonstration. Relevance to a particular operation must be assessed separately.

Sources and their limits

Prepared with AI assistance. The linked sources support the specified technical points; they do not validate applied psionics as a whole or guarantee a result for a client.

Read the editorial and evidence standard.

Continue reading

Explore Scientific Oracle consultingfor a scoped review of an existing-data research decision. Start with a non-confidential outline of the question, available evidence and the decision it needs to inform.