Short-Term Photovoltaic Power Forecasting Based on Feature Selection and GA-BP Neural Network Model

This study proposes a hybrid prediction model integrating feature dimensionality reduction and genetic algorithm (GA)-based parameter optimization for backpropagation neural networks to enhance the accuracy of short-term photovoltaic power forecasting. Employing the Random Forest (RF) algorithm to assess the correlation between multiple meteorological variables and output power. High-importance features (top 6 ranked by RF) are selected to eliminate redundant inputs, reducing computational complexity while maintaining predictive relevance. Secondly, to mitigate local minima issues in BP neural network training, GA is introduced to optimize initial weights and thresholds. The weight matrix is encoded as genetic chromosomes, with the evolutionary process guided through crossover and mutation operations, yielding globally optimal network configurations This approach effectively mitigates the inherent limitation of traditional BP networks in converging to local minima. The simulation experiment is carried out with the measured data of a photovoltaic power station, and different prediction models are established to compare. The experiment shows that the proposed model has excellent prediction accuracy, and the method provides an effective new way for short-term photovoltaic power prediction.