Explainable Deep Learning for Hyperparameter Optimization-Based Prediction of Solar Radiation Intensity Classification

Abstract

Due to the intermittent solar presence, forecasting Solar radiation, is crucial to balancing energy generation and demand, which is critical for the whole grid system. Deep neural networks have become the standard de facto in many fields, and also in forecasting. Despite that, these models are black-box, meaning that it is difficult to understand, given an input, why the corresponding output is produced. This aspect is crucial when a neural network is applied to a real-world scenario such as solar radiation forecasting. For this reason in this paper, we use an explainable model to forecast values on two closed datasets consisting of weather parameters in Basel–Switzerland. Both datasets contain measured values, one from January 2012 to March 2018 and the other from January to December 2020. Evaluating the results against existing models in the literature shows the superiority of the employed model in predicting solar radiation. It is anticipated that the applied model, which offers excellent performance and explainability, would resolve the black-box nature of neural network models in predicting and forecasting solar radiation.