Investigation on Photovoltaic Power Forecasting Model Using Sensitivity Analysis

Abstract

Photovoltaic (PV) power forecasting is challenged by high-dimensional input features that cause redundancy and overfitting. To address this, we propose a short-term PV forecasting model integrating a CNN-LSTM architecture with active subspace-based global sensitivity analysis, which quantitatively evaluates feature influence via gradient information. Using operational data from a PV plant, we compare three feature selection schemes across CNN-LSTM, CNN, and LSTM models. For CNN-LSTM, the baseline yields an average R2 of 0.8769 (RMSE: 134.2 kW), while Pearson-based selection slightly reduces performance (R2 = 0.8646, RMSE: 140.7 kW). In contrast, the active subspace method achieves an average R2 of 0.9341 (RMSE: 83.6 kW), representing improvements of 6.52% and 8.04% over the no-selection and Pearson schemes, respectively. Overfitting is mitigated and convergence accelerates. The method also shows consistent improvements on CNN and LSTM models, confirming its model-agnostic effectiveness. These results demonstrate that the active subspace method is an efficient and interpretable feature engineering tool for PV forecasting.