Abstract
This paper presents a comprehensive investigation into enhancing photovoltaic (PV) power forecasting by systematically integrating feature selection techniques with artificial neural networks. Addressing the growing demand for reliable renewable energy forecasting, the study employs several feature selection methods, including ReliefF, minimum correlation, Chi-square test, and others, to identify the most relevant predictors for PV output prediction. Two predictive models, the multilayer perceptron (MLP) and long short-term memory (LSTM) networks, are developed and tested on a real-world dataset from southern Algeria. The results demonstrate that applying feature selection significantly improves forecasting accuracy. For instance, integrating ReliefF with MLP reduced the normalized mean absolute error (nMAE) to 9.21% with an R(2) of 0.9608, while the best LSTM configuration achieved an nMAE of 9.29% and an R(2) of 0.946 when using Chi-square selected features. These findings confirm that careful feature selection enhances model performance, reduces complexity, and ensures better generalization, offering valuable insights for more efficient solar energy management and grid stability.