TCN-LSTM-AM Short-Term Photovoltaic Power Forecasting Model Based on Improved Feature Selection and APO.

The inherent volatility and intermittency of solar power generation pose significant challenges to the stability of power systems. Consequently, high-precision power forecasting is critical for mitigating these impacts and ensuring reliable operation. This paper proposes a framework for photovoltaic (PV) power forecasting that integrates refined feature engineering with deep learning models in a two-stage approach. In the feature engineering stage, a KNN-PCC-SHAP method is constructed. This method is initiated with the KNN algorithm, which is used to identify anomalous samples and perform data interpolation. PCC is then used to screen linearly correlated features. Finally, the SHAP value is used to quantitatively analyze the nonlinear contributions and interaction effects of each feature, thereby forming an optimal feature subset with higher information density. In the modeling stage, a TCN-LSTM-AM combined forecasting model is constructed to collaboratively capture the local details, long-term dependencies, and key timing features of the PV power sequence. The APO algorithm is utilized for the adaptive optimization of the crucial configuration parameters within the model. Experiments based on real PV power plants and public data show that the framework outperforms multiple comparison models in terms of key indicators such as RMSE (2.1098 kW), MAE (1.1073 kW), and R2 (0.9775), verifying that the deep integration of refined feature engineering and deep learning models is an effective way to improve the accuracy of PV power prediction.