ElecLoad lets shifting event effects set the forecast segments: PJM WAPE falls from 4.31% to 3.69% when events pile up
Synopsis
ElecLoad reads weather alerts, grid notices and news event descriptions alongside historical load, judging each event's relevance to the target load, the timing of its effect, and whether events reinforce or offset one another, then decides whether to forecast the whole horizon at once or split it where event effects change. On the segmented path, sub-agents draft each segment independently and the main agent reviews them and sends feedback to selected segments only. Across hourly load data including GEFCom2014, PJM and LCL it ranks near the top in most evaluation scenarios; under multiple events on PJM its WAPE is 3.69% versus 4.31% for a direct-only variant using the same model and inputs.
Interpretation
Electricity load forecasting can now let how event effects change over time steer the forecasting process itself, rather than only feeding event text in as extra input. Earlier load-forecasting methods encode news or calendar text and fuse it with load sequences, and some analyze news relevance, effect direction and duration, but how to use changes in multiple events' effects to guide the forecasting process remained to be explored. On GEFCom2014, ElecLoad lowers WAPE and NRMSE by 18.3% and 18.6% relative to the strongest baseline, MemCast; across 22 metric comparisons it ranks first in 13 and second in six, placing top-two in all 18 in-domain comparisons.
The mechanism has a main agent read the events and choose between whole-horizon forecasting and segmentation, after which sub-agents draft segments and the main agent picks segments for feedback. The main agent infers each event's effect direction and impact window from the activities, regions and timing in its description, and the window may include delayed response and recovery after the event ends. In GEFCom2014 groups built for single, multiple and irrelevant events, the full workflow reaches 7.61% WAPE under multiple events versus 9.50% for the direct-only variant; on PJM the pair is 3.69% and 4.31%.
Refinement is invoked only when the main agent judges a forecast inconsistent with the event analysis, and it is capped by a budget. Refinement without feedback and random refinement both raise error, so the gain comes from where the feedback points, not merely from calling the model again. On GEFCom2014 the full workflow reaches 6.68% WAPE, refinement without feedback 7.06% and random refinement 6.94%, at comparable token cost of 18.7k versus 18.6k tokens.
Perspective
Evaluation covers hourly household, regional and system load from GEFCom2014, Nord, PJM, SPP, LCL and iFlex, plus AEMO and ERCOT as out-of-domain systems used only for testing; event announcements are available at the corresponding forecast origins. Synthetic event cases are built from GEFCom2014 original sequences by modifying future targets according to specified magnitudes and timing, to test single-event, multi-event and irrelevant-event conditions.
Across the four out-of-domain metric comparisons on AEMO and ERCOT, ElecLoad ranks second to fourth among 15 methods rather than first. Which path wins varies by dataset: segmented-only beats direct-only on GEFCom2014, while the reverse holds on LCL. On LCL the 95% interval of the paired WAPE difference against direct-only includes zero.