Large Language Model-Assisted Load Forecasting Methods for Integrated Energy Systems

To address the problems of complex feature coupling and redundant information affecting prediction accuracy in multi-energy load forecasting for integrated energy systems, this paper proposes a load forecasting method based on a pretrained large language model. Based on Time-LLM, the proposed method freezes the main parameters of the large language model and designs a channel-time adaptive feature selector, a multi-criterion correlation-redundancy filter, and a multi-channel prompt prefix module to achieve local dynamic feature weighting, global correlation-redundancy filtering, and statistical prompt enhancement, respectively. Experimental results on the Arizona State University integrated energy dataset show that, compared with Time-LLM, the proposed method reduces the MAE by 8.06%, 6.73%, and 8.93%, and the RMSE by 5.98%, 7.10%, and 5.31% for electricity, cooling, and heating load forecasting, respectively. The results indicate that the proposed method can improve the accuracy of multi-energy load forecasting and exhibits good stability.