Objective With the advancement of the "dual carbon" strategy and the construction of new power systems, electricity load forecasting plays an increasingly important role in ensuring grid security and optimizing resource allocation. As a typical high-altitude region, Xizang experiences electricity load influenced by multiple factors including climate, terrain, and economic structure, making it difficult for traditional forecasting models to meet dispatching requirements.

Method Taking Shigatse City as an example, this study utilized hourly electricity load data and concurrent meteorological observations from December 2021 to November 2024 to construct a Random Forest regression-based short-term load forecasting model. The temporal load characteristics and their multi-scale correlations with seven meteorological factors were systematically analyzed.

Result The results show that Shigatse's load exhibits a "high in winter, low in summer" seasonal pattern, with winter peaks reaching 329 MW and summer lows around 240 MW, and a distinct morning-evening "double peak" diurnal profile. Relative humidity (−0.33) and dew point temperature (−0.27) show the strongest negative correlations with load. After removing the seasonal cycle, the correlation structure changes significantly: the temperature-load correlation strengthens from −0.10 to −0.16, while relative humidity and dew point temperature reverse to positive correlations. Season-specific analysis reveals that temperature is positively correlated with load in winter (+0.13), spring (+0.24), and summer (+0.33), but weakly negative in autumn (−0.12). The Random Forest model achieves an overall MAE of 9.78 MW, RMSE of 14.39 MW, MAPE of 4.74%, and R² of 0.80 on the test set. Prediction accuracy is highest during low-load periods (MAE=5.67 MW, MAPE=3.95%) and relatively lower during high-load periods (MAE=16.31 MW, MAPE=6.58%). Monthly errors are lowest in February (MAPE=2.76%), and the overall MAPE decreases to approximately 4.17% after excluding December 2023, which experienced an anomalous load surge.

Conclusion This study demonstrates the applicability of the Random Forest model for short-term load forecasting in high-altitude regions, providing methodological support for plateau power grid dispatching. Future work may incorporate load trend indicators and extreme weather features to further enhance model performance.