A comparison of neural networks for prediction of generation of thermal energy of Flat Plate Vacuum solar thermal collectors

Solar thermal collectors are essential for sustainable energy production, yet accurately predicting their energy output remains challenging. This research compares the precision of three time series neural networks—NARX (Nonlinear Autoregressive with Exogenous Inputs), NAR (Nonlinear Autoregressive), and Input-Output models—for forecasting thermal energy generation y(t) from flat plate vacuum solar collectors based on time series data x(t). The objective was to determine which neural network architecture provides the highest reliability for energy output prediction, enabling more effective system management. Using a correlational design, each neural network was constructed with historical thermal energy and solar radiation data, then subjected to training, validation, and testing phases. Predictive accuracy was evaluated through linear regression analysis between network outputs and corresponding targets, quantifying how well each model could generalize to new data. Results revealed that the NARX model demonstrated superior overall performance with consistent high correlation coefficients across all phases. The Input-Output model showed exceptional accuracy particularly during the testing phase, suggesting strong practical reliability. The NAR model, while effective overall, exhibited reduced accuracy in the testing phase, indicating limitations in generalizing to unknown data. This study concludes that the NARX architecture provides the most stable and accurate framework for predicting thermal energy generation in flat plate vacuum solar collectors. These findings contribute to more effective planning of solar energy systems, optimization of resources, and improved maintenance scheduling, ultimately reducing operational costs and extending system component lifespans.