Abstract: Aiming at the problem that a single model cannot balance prediction accuracy and stability in the short-term power prediction of photovoltaic power stations, an adaptive fusion method combining LR-RF with error compensation is proposed. Firstly, data preprocessing including outlier elimination, missing value filling, feature screening, lag feature construction and standardization is carried out on historical operational and meteorological data. Then, linear regression (LR) and random forest (RF) are constructed as base models. On this basis, weights are dynamically assigned according to recent prediction performance and meteorological scenario characteristics. Meanwhile, sliding window deviation correction and scenario-based error compensation are introduced. The experimental results show that the proposed method outperforms single models and fusion models without compensation in evaluation indicators including RMSE, MAE, MAPE and R2. This method possesses superior prediction accuracy and stability, and has practical engineering application value.