The accurate short-term forecast of photovoltaic power generation is crucial for meeting the dispatching needs of the power system and maintaining grid safety. This paper proposes a fusion forecasting model based on data feature engineering to improve precision under different weather conditions. The model integrates multiple forecasting methods to capture the nonlinear relationships between weather variables and power generation. Abnormal data points are filtered using normal distribution analysis and irradiance thresholds. Meteorological features are selected via an improved filtering and wrapping technique, and the dataset is clustered using the K-means algorithm. The final model fuses eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and MultiLayer Perceptron (MLP) through a dual-hidden layer perceptron, which highlights the nonlinear relationships between features. We conducted extensive experiments on real-world datasets under various weather conditions. The results indicate that the fusion model consistently outperforms individual models across all environments, with a particularly notable 0.23 reduction in mean squared error (MSE) under rainy conditions.