Optimizing Solar Radiation Forecasting for Renewable Energy Systems:

A Comparative Analysis of Machine Learning and Feature Engineering Techniques

Abstract

Accurate solar radiation prediction is pivotal for optimizing solar energy systems, as it allows for better energy storage, grid integration, and renewable energy planning. This study compares the predictive accuracy of three machine learning models—Random Forest, XGBoost, and Multilayer Perceptron (MLP)- in forecasting solar radiation based on a meteorological and temporal features dataset. The dataset, encompassing Temperature, humidity, wind speed, and sunrise/sunset times, was preprocessed through transformations (Box-Cox, logarithmic scaling) and feature selection methods (SelectKBest, Extra Trees Classifier) to enhance model performance. XGBoost demonstrated superior performance, achieving an R² of 0.93 and RMSE of 81.87, effectively capturing complex nonlinear relationships within the data. MLP, while slightly lower in R², yielded the lowest mean absolute error (MAE = 41.74), underscoring its precision in individual predictions. SelectKBest identified set Hour (sunset hour), Month, and Wind Direction as critical features, while Extra Trees prioritized Wind Direction, Minute, and Humidity, reflecting model-specific feature importance. Collectively, these models illustrate the benefits of integrating feature engineering with advanced machine learning for renewable energy optimization, with XGBoost and MLP demonstrating particular efficacy for accurate solar radiation forecasting. This study underscores the potential of machine learning in enhancing solar energy management, facilitating a more efficient transition to sustainable energy sources.