Abstract: Short-term load forecasting is a key technology for ensuring the safety and economic operation of power grids. Addressing the issue that actual load data is prone to strong uncertainties due to multi-source disturbances such as meteorological conditions, day types, and renewable energy output. This paper proposes a refined short-term load forecasting model designed to handle multiple types of disturbances. Built upon the LightGBM framework, the model employs feature engineering to identify the coupling effects of various influencing factors and utilizes a time-series-aware hyperparameter optimization technique to enhance generalization performance. Experiments conducted on actual load data from a regional power grid demonstrate that the proposed model improves forecasting accuracy by over 5% compared to baseline models such as ARIMA and linear regression, offering higher accuracy and stability.