Collaborative short-term load forecasting with large and lightweight models under abrupt meteorological change scenarios

Abstract

Dynamic meteorological fluctuations are a major source of forecasting bias and robustness degradation in short-term power load forecasting. In periods with frequent weather variability, abrupt temperature rises or drops can trigger local load mutations that are difficult for conventional models to capture accurately. To address this problem, this study develops a collaborative short-term load forecasting framework that integrates multi-meteorological features, a large time-series model, and a lightweight specialized residual model. First, differentiated electricity-consumption scenarios during weather-sensitive seasons are analyzed, and the response mechanisms of loads under abrupt warming, abrupt cooling, gradual warming, and gradual cooling are quantified. A lightweight model combining gradient-boosting decision trees and an attention-gating mechanism is constructed to sense local load deviations caused by meteorological mutations with low training cost and rapid online adaptability. Second, multidimensional meteorological features, including temperature, humidity, wind speed, precipitation, and solar radiation, are fused into a large time-series forecasting model to learn the global nonlinear evolution of load sequences. Finally, a collaborative forecasting framework is designed in which the large time-series model captures global temporal patterns and the lightweight model corrects local real-time residuals. Offline pretraining and online adaptive tuning are combined to achieve complementary advantages. Case-study results show that the proposed framework improves forecasting accuracy and stability under complex weather conditions and provides technical support for refined grid dispatching and accurate load management.