Interpretable generalized Gaussian mixture modeling for risk-aware solar power forecasting

Reliable photovoltaic (PV) power forecasting requires not only accurate point predictions but also transparent and well-calibrated representations of uncertainty. This paper introduces an interpretable probabilistic forecasting framework that leverages a Generalized Gaussian Mixture Model (GGMM) parameterized by a Feedforward Neural Network (FNN) to capture the non-Gaussian and heteroscedastic nature of solar power generation. The contribution lies in using the GGMM mixture parameters (mean, scale, and shape) as intrinsic indicators of forecast confidence and distributional behavior. The model is trained end-to-end using a negative log-likelihood objective and tuned through Bayesian optimization. Experiments on the GEFCom2014 and UNISOLAR datasets show that this GGMM-based formulation provides better calibrated uncertainty estimates and improved robustness under variable weather conditions compared to several strong deterministic and probabilistic baselines. In addition, analyzing the temporal evolution of the mixture weights and shape parameters offers clear interpretability by revealing regime changes, tail behavior, and confidence levels across the forecast horizon. These properties highlight the proposed framework as a practical, explainable, and decision-oriented tool for solar power forecasting.