ElecLoad Puts Events at the Center of Electricity Demand Forecasting

A heatwave can raise electricity use just as a demand-response program asks customers to cut it. A new preprint explores whether forecasting systems can better navigate such overlaps by using event information not just as an input, but as a guide to the forecasting process itself.

Electricity demand forecasts help power-system operators plan generation, procurement and other resources. Getting the shape of demand right matters as well as its overall level: an unexpected peak or a poorly timed decline can complicate planning.

Regular historical patterns are only part of the picture. Weather alerts, operational notices and price schedules can signal changes in consumption, but their effects may overlap, persist or fade at different times. A heatwave might increase cooling demand while a program to reduce consumption temporarily offsets some of that increase. Demand could then recover after the program ends.

In a preprint, Huajian Zhang and colleagues propose ElecLoad, a system designed to use such information to organize its own forecasting workflow. The authors report that it performs well against a range of comparison methods, while their experiments also show that the best approach is not identical for every dataset.

From event descriptions to a forecast plan

ElecLoad takes a history of hourly electricity load along with available contextual information, such as calendar features, numerical covariates and event descriptions. The events in the experiments are available at the time the forecast is made. Their stated dates help locate them in time, but the system distinguishes those dates from the potentially longer or delayed period during which demand might be affected.

A main agent first assesses each event’s relevance, likely direction and timing of impact, and possible interactions with other events. It then chooses between two paths: forecast the entire future period at once, or divide it into segments where the expected effects change.

On the segmented path, separate sub-agents produce initial numerical forecasts for their assigned periods. Each receives the original load history and context, plus the main agent’s analysis, but not the other sub-agents’ initial forecasts. The main agent reviews the assembled result and can ask for selected segments to be revised. A final check can retain or adjust values near segment boundaries. The agents are different roles in the workflow, not independently trained models: they share the same underlying language-model parameters.

The researchers trained the model in stages. Initial training covered load prediction and event understanding; subsequent training targeted the agents’ forecasting and control decisions. Training used examples reviewed with the help of known load values, followed by reinforcement learning that accounted for forecast error and the token cost of the calls.

What the tests found

The evaluation covered household, regional and system-level electricity load, with day-ahead and multi-day horizons. It compared ElecLoad with statistical, machine-learning, time-series and language-model-based methods. The main error measure, weighted absolute percentage error, expresses the total size of forecast deviations relative to observed load; a second measure gives more weight to large deviations.

Across 22 dataset-and-metric comparisons, ElecLoad ranked first in 13 and second in six. On the GEFCom2014 multi-day evaluation, its weighted absolute percentage error was 6.68%, compared with 8.18% for the strongest baseline on that measure. It was not always best: on the held-out Australian and Texas power-system evaluations, it ranked between second and fourth across the four reported metric comparisons.

The authors also compared ElecLoad with versions of itself. Removing event text or other covariates increased errors across the three datasets used for that ablation. Always forecasting directly, always segmenting, using equal-length segments, and removing selective refinement also produced higher average errors than the full system in those tests. On GEFCom2014, adaptive routing used fewer tokens than always segmenting, but more than forecasting directly.

For PJM cases grouped as having multiple relevant events, ElecLoad’s weighted error was 3.69%, versus 4.31% for a version that used the same model and inputs but always forecast the full horizon directly. That comparison tests the combined adaptive workflow, rather than isolating any one component.

Where the evidence stops

To examine interactions under controlled conditions, the researchers also created synthetic events from GEFCom2014 load sequences, modifying future demand according to specified event effects. They compared forecasts with no added event, each event separately and both together. ElecLoad was more accurate than its direct-only variant at estimating the specified interactions in the reported controlled comparison.

Those synthetic results show how the method behaves when event effects are constructed by the researchers; they do not establish that it can identify the true causal effects of real-world events. The real-data evaluations, meanwhile, measure forecast accuracy under the available inputs and selected test cases. For household tests, the validation and test households had also appeared in training, so those results concern later dates for known households.

Taken together, the preprint offers evidence that event-guided planning and selective review can improve forecasts in the evaluated settings. Its broader proposal is that interpreting when events matter can help determine not only what a forecast should say, but how the forecast is produced.