An interpretable machine learning model for daily global solar radiation prediction

Abstract

Machine learning (ML) models are commonly used in solar modeling due to their high predictive accuracy. However, the predictions of these models are difficult to explain and trust. This paper aims to demonstrate the utility of two interpretation techniques to explain and improve the predictions of ML models. We compared first the predictive performance of Light Gradient Boosting (LightGBM) with three benchmark models, including multilayer perceptron (MLP), multiple linear regression (MLR), and support-vector regression (SVR), for estimating the global solar radiation (H) in the city of Fez, Morocco. Then, the predictions of the most accurate model were explained by two model-agnostic explanation techniques: permutation feature importance (PFI) and Shapley additive explanations (SHAP). The results indicated that LightGBM (R2 = 0.9377, RMSE = 0.4827 kWh/m2, MAE = 0.3614 kWh/m2) provides similar predictive accuracy as SVR, and outperformed MLP and MLR in the testing stage. Both PFI and SHAP methods showed that extraterrestrial solar radiation (H0) and sunshine duration fraction (SF) are the two most important parameters that affect H estimation. Moreover, the SHAP method established how each feature influences the LightGBM estimations. The predictive accuracy of the LightGBM model was further improved slightly after re-examination of features, where the model combining H0, SF, and RH was better than the model with all features.