Leveraging Pretrained Large Language Models for Power System Source-Load Forecasting

With the increasing penetration of renewable energy in power systems, high-precision and adaptive power forecasting is crucial for system stability. Traditional forecasting methods face challenges in capturing nonlinear relationships and adapting to dynamic conditions. This paper explores the use of large language models (LLM) for source-load forecasting, focusing on the cross-modal transferability of LLM weights and their effectiveness in forecasting applications. The proposed framework employs a time-series-driven fine-tuning strategy, which freezes attention weights, adjusts layer normalization, and uses data window slicing to enhance the model's ability to capture nonlinear patterns. A modular covariate fusion architecture is introduced to better integrate meteorological data. Through experimental validation, the framework demonstrates strong cross-modal transferability and practical forecasting performance across diverse scenarios, highlighting the potential of LLMs in improving source-load forecasting in power systems.