Electricity load forecasting using large language models

With the increasing global energy demand and the widespread deployment of renewable energy sources, the operation and management of power systems face unprecedented challenges. Electricity load forecasting is a critical component of power system planning and operation, playing a significant role in ensuring the reliability of power supply, optimizing resource allocation, and enhancing economic benefits. This study proposes a deep learning-based electricity load forecasting framework that integrates two large language models, Qwen2.5-14B and Lag-Llama, to improve the accuracy and robustness of predictive models through meticulously designed prompt engineering, LoRA fine-tuning strategies, and data fusion tactics. The experiment utilized hourly electricity consumption data from May 1, 2004, to April 30, 2009, to forecast hourly electricity consumption from May 1 to October 31, 2009, and included the Shanghai Composite Index, CSI 300, CSI 500, and CSI 1000 as auxiliary financial data for prediction. The results demonstrate that our approach shows certain advantages in predictive accuracy, especially in handling multi-source data and considering spatiotemporal correlations. This provides valuable insights for power operators to optimize power generation plans and scheduling decisions, thereby improving economic efficiency and reducing environmental impact.