Enhancing Solar Radiation Prediction for Computational-Constrained Environments Using Hybrid Artificial Neural Networks

Abstract

This paper presents a hybrid predictive model optimized for resource-constrained environments, focusing on renewable energy systems in remote areas. The model integrates various Artificial Neural Network (ANN) architectures—Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Bidirectional Convolutional Neural Networks (BiCNN)—to balance accuracy and computational efficiency. Implemented on a Raspberry Pi 4, the model combines the efficiency of MLP, the long-term dependency capabilities of LSTM, and the low memory consumption of BiCNN. It incorporates preprocessing stages to address the challenges of solar radiation prediction. Comparative analysis shows a 90.74% reduction in Root Mean Square Error (RMSE), an 89.85% reduction in Mean Absolute Error (MAE), and a 3.51% increase in the coefficient of determination (R2). The model significantly reduces parameter count with minimal impact on execution time and memory consumption, making it highly suitable for deployment in resource-constrained environments. This approach advances distributed generation systems and offers potential benefits for renewable energy solutions in underserved communities.

Keywords

• Renewable energy
• predictive modeling
• artificial neural networks (ANN)
• resource-constrained environments
• distributed generation
• solar radiation prediction