Llama3-QLoRA-GeoWeather: A Spatiotemporal Feature Fusion and Two-Stage Fine-Tuning Framework for Power Load Forecasting

Power load forecasting is essential for power system security and energy dispatch. With the increasing renewable integration, load patterns have become more volatile and uncertain, difficult for traditional forecasting methods to maintain high adaptability. To address this challenge, this study proposes the Llama3-QLoRA-GeoWeather framework, a novel power load forecasting approach based on the open-source large language model Llama 3.3 70B. The framework introduces a two-stage progressive fine-tuning strategy based on QLoRA, significantly reducing computational costs and allowing adaptation on constrained hardware. Moreover, geographic features from the OpenStreetMap ecosystem and meteorological data from OpenWeatherMap API are integrated to further enhance the forecasting performance. A comprehensive Llama3-PowerFrame enhancement framework for future power systems is also designed. Experimental results demonstrate that Llama3-QLoRA-GeoWeather achieves the best forecasting performance (MAPE = 1.16%), outperforming the state-of-the-art baselines. This corresponds to a reduction in MAE, RMSE, and MAPE by approximately 42.7%, 67.8%, and 42.3% respectively, providing a viable technical pathway for building the next-generation intelligent load forecasting system across multiple scenarios with high credibility and strong adaptability.