Short-term electricity load forecasting based on large language models and weighted external factor optimization

Accurate short-term electricity load forecasting is essential for the stable and efficient operation of modern power systems. This paper proposes a novel forecasting framework that integrates a time-decay weighted average of historical loads with external factor adjustments derived from structured data (weather and calendar) and unstructured textual sources (news). A key innovation lies in the dual role played by the Generative Pre-trained Transformer (GPT), which is used both to extract semantic features from news texts and to perform adaptive self-tuning mechanism that dynamically refines external factor scores. The model is validated through a case study using real-world data from New York State. Results show that the optimized prediction significantly improves forecasting accuracy, reducing the mean absolute error (MAE) by 90.7% and the root mean square error (RMSE) by 88.5% compared to the initial estimate. This work demonstrates the practical potential of large language models in energy forecasting and offers a scalable approach for integrating natural language understanding into quantitative prediction systems