Enhanced hourly temperature prediction using advanced ensemble neural networks for energy system efficiency optimization in hyper-arid regions

This paper presents advanced techniques for air temperature forecasting based on the relationship between meteorological variables and feedforward neural network models aimed at improving prediction accuracy. An optimal single neural network model was developed, and optimization through dataset partitioning resulted in high predictive preformance, with all correlation coefficients exceeding 0.97 in the test phase and a root mean square error (RMSE) of 1.5587. Additionally, a bootstrap-aggregated neural network (BANN) model comprising 30 networks in a stacked ensemble was implemented, yielding robust results; both validation and testing correlation coefficients were close to the ideal value. Sensitivity analysis further revealed that true solar time, extraterrestrial radiation, and wind speed were the most influential predictors—an outcome that is consistent with their physical relevance to temperature variation. Moreover, the model’s validity was confirmed through applicability domain analysis using the Williams plot, which showed that over 97% of the predictions fell within acceptable error limits. A comparative evaluation against other existing models demonstrated the superior predictive performance of the proposed approach, reinforcing the potential of the BANN model for accurate air temperature forecasting under diverse meteorological conditions.