Gaussian noise injection and EMD-based feature selection for photovoltaic power prediction: dynamic importance evaluation and multi-dimensional accuracy comparison with pearson correlation methods

In the prediction of photovoltaic (PV) power generation, accurate and efficient feature selection is particularly critical for precise power prediction and the stable operation of the power grid, especially under limited data conditions.Traditional feature selection in the prediction of PV power generation primarily involves analyzing the Pearson correlation coefficient between multi-dimensional input feature time series and the prediction target.This paper proposes a novel empirical mode decomposition (EMD) based method for Photovoltaic Power Prediction by using Gaussian noise injection for feature importance calculation.Among the modal components of photovoltaic power data, high-frequency components typically represent short-term rapid fluctuations in power, while low-frequency components usually characterize long-term trend variations in power. In this research, specific combinations of modal components, when used as inputs for power prediction, achieve better prediction accuracy than when all modal components are used as inputs for power prediction. Perturbation of each feature is achieved by injecting Gaussian noise. In each scenario of feature combinations, a weight value is assigned to each feature. Then, the weight values of a specific feature across all scenarios where it is included are aggregated. Finally, the importance value of this feature is obtained. The proposed feature importance calculation method is used to assess the contribution of each modal component. Based on the feature importance evaluation results, the feature combination that contributes most to prediction accuracy is selected to construct the final prediction model. To demonstrate the superior performance of the proposed method, a comparison is made between the feature selection based on feature importance and the feature selection based on the Pearson correlation coefficient. The same predictive model is trained using the selected features for each case, and the prediction accuracy is validated. The proposed method demonstrates higher accuracy across all dimensions of feature selection.