Trustworthy, Uncertainty-Aware Short-Term Electricity Load Forecasting

The crisis

• Electricity load forecasts drive generator scheduling and reserve procurement; when the forecast is wrong and its uncertainty is unstated, the grid either over-provisions (wasting fuel, money, and carbon) or under-provisions (risking shortfalls and outages).
• A point forecast reports a single number and hides the one thing an operator's decision actually depends on — how uncertain that number is, and under which conditions it becomes unreliable.
• Variable renewables, electrified heating and transport, and more frequent extreme-weather regimes are making demand more volatile, so forecast uncertainty is rising precisely as the cost of mis-estimating it rises.
• Most uncertainty methods report a wider band when unsure but cannot act to reduce that uncertainty; and most produce intervals whose stated confidence is not the confidence you actually get under distribution shift.
• As AI systems increasingly mediate grid operations, honest, coverage-guaranteed, auditable uncertainty — not just accuracy — is what decides whether a forecast can be trusted in the control room.

About this research

Short-term electricity load forecasting is treated here as a problem of honest uncertainty rather than point accuracy alone: a dispatch decision depends not only on the predicted demand but on how trustworthy that prediction is, especially when demand enters an unusual regime such as a heat wave, holiday, or cold snap. This thread investigates how a forecasting system can produce calibrated, well-behaved uncertainty that means what it says under shifting conditions, and how such a system can act to reduce that uncertainty rather than only report it, while keeping the whole process auditable. It is framed throughout as operator decision support that drafts a calibrated recommendation; a human keeps final say on dispatch. The work draws on probabilistic forecasting, uncertainty quantification, agentic LLM systems, and rigorous evaluation on public energy data. Faculty-advised.

References

• Hong et al. (2016) · Probabilistic energy forecasting: GEFCom2014 and beyond (Int. J. Forecasting)
• EIA · Hourly Electric Grid Monitor (Form EIA-930) — open hourly demand data