Socially Aware Load Forecasting Utilizing Large Language Models

Accurate residential load forecasting is crucial for power system management, and incorporating socio-demographic information can significantly enhance prediction accuracy. However, traditional methods struggle to leverage the unstructured and textual socio-demographic factors, which limits their effectiveness in accurately predicting energy consumption influenced by social factors. To address this gap, we propose LLMSA, a socially-aware load forecasting framework that leverages the large language model (LLM) to understand those unstructured and textual factors and integrate them into the load forecasting process. LLMSA incorporates a time series input transformation to learn the temporal features of historical load data and then utilizes a socio-related adapter with a socially aware reprogramming layer to fuse socio-demographic information into the forecasting process effectively. Simulation studies conducted based on the Irish Commission for Energy Regulation dataset have shown that LLMSA outperforms existing models, including PatchTST, iTransformer, and TimeLLM, particularly in predicting peak energy usage periods. Compared to baseline models, LLMSA achieves up to a 4.96% reduction in mean squared error. Our research shows that by incorporating textual social context, LLMSA provides more accurate and reliable load predictions, demonstrating its potential to enhance energy management strategies.