Transfer-learning-based prediction of abrupt regional grid load changes under cold-wave weather

Abstract

Cold-wave weather produces rapid heating demand growth, abnormal industrial operating schedules, and larger reserve- pressure in regional power grids. Conventional short-term load forecasting models trained on normal winter samples often under-estimate the magnitude and direction of abrupt load changes because cold-wave samples are sparse and distribution- shifted. This paper proposes a cold-wave transfer-learning Transformer, named CW-TLFormer, to predict both the 24-h load trajectory and the probability of a load mutation. The model first learns seasonally stable temporal-weather representations from multi-winter source data and then adapts them to rare cold-wave target episodes through maximum mean discrepancy alignment, cold-wave gating, and a ramp-weighted loss. Hourly load, meteorological, calendar, grid- state, and optical-sensing weather features are jointly encoded, so the method can support automated dispatch in energy facilities, electrified transport hubs, and EV charging clusters during extreme cold conditions. A rolling-origin experiment on an anonymized regional-grid dataset from six winter seasons shows that CW-TLFormer reduces cold-wave RMSE from 150.8 MW for a standard Transformer to 120.5 MW and improves ramp-event F1 from 0.68 to 0.81. Ablation tests further confirm that domain alignment, ramp weighting, and the cold-wave gate are all necessary for stable mutation prediction.