An Interpretable Probabilistic Model for Short-Term Solar Power Forecasting Using Natural Gradient Boosting

The stochastic nature of photovoltaic (PV) power has led both academia and industry to a large amount of research work aiming at the development of accurate PV power forecasting models. However, most of those models are based on machine learning algorithms and are considered as black boxes which do not provide any insight or explanation about their predictions. Therefore, their direct implementation in environments, where transparency is required, and the trust associated with their predictions may be questioned. To this end, we propose a two stage probabilistic forecasting framework able to generate highly accurate, reliable, and sharp forecasts yet offering full transparency on both the point forecasts and the prediction intervals (PIs). In the first stage, we exploit natural gradient boosting (NGBoost) for yielding probabilistic forecasts while in the second stage, we calculate the Shapley additive explanation (SHAP) values in order to fully understand why a prediction was made. To highlight the performance and the applicability of the proposed framework, real data from two PV parks located in Southern Germany are employed. Initially, the natural gradient boosting is thoroughly compared with two state-of-the-art algorithms, namely Gaussian process and lower upper bound estimation, in a wide range of forecasting metrics. Secondly, a detailed analysis of the model's complex nonlinear relationships and interaction effects between the various features is presented. The latter allows us to interpret the model, identify some learned physical properties, explain individual predictions, reduce the computational requirements for the training without jeopardizing the model accuracy, detect possible bugs, and gain trust in the model. Finally, we conclude that the model was able to develop nonlinear relationships following human logic and intuition based on learned physical properties.