Hybrid Convolutional Neural Network-Multilayer Perceptron Model for Solar Radiation Prediction

Urgent transition from the dependence on fossil fuels towards renewable energies requires more solar photovoltaic power to be connected to the electricity grids, with reliable supply through accurate solar radiation forecasting systems. This study proposes an innovative hybrid method that integrates convolutional neural network (CNN) with multi-layer perceptron (MLP) to generate global solar radiation (GSR) forecasts. The CMLP model first extracts optimal topological and structural features embedded in predictive variables through a CNN-based feature extraction stage followed by an MLP-based predictive model to generate the GSR forecasts. Predictive variables from observed data and global climate models (GCM) are used to predict GSR at six solar farms in Queensland, Australia. A hybrid-wrapper feature selection method using a random forest-recursive feature elimination (RF-RFE) scheme is used to eradicate redundant predictor features to improve the proposed CMLP model efficiency. The CMLP model has been compared and bench-marked against seven artificial intelligence-based and seven temperature-based deterministic models, showing excellent performance at all solar energy study sites tested over daily, monthly, and seasonal scales. The proposed hybrid CMLP model should be explored as a viable modelling tool for solar energy monitoring and forecasting in real-time energy management systems.