Solar Irradiance Estimation from SEVIRI: A Feature Importance and Channel Selection Analysis

This study proposes a deep learning framework for estimating global horizontal irradiance (GHI) from SEVIRI satellite data, incorporating feature importance analysis and using Heliosat-4 as a reference benchmark. Building on prior work, we refine temporal alignment between satellite inputs and ground truth by aggregating 1-minute BSRN GHI measurements to 15-minute and hourly intervals. A multilayer perceptron (MLP), trained on 12 SEVIRI channels and Mc-Clear clear-sky GHI, is evaluated using leave-one-location-out cross-validation (LOLO-CV) across 16 diverse stations. Results consistently show improved accuracy over Heliosat-4 at both temporal scales. To enhance model interpretability and reduce computational complexity, we apply SHAP (SHapley Additive exPlanations) analysis to assess channel relevance. A reducedinput model that uses only the most informative channels maintains comparable performance, demonstrating the potential for efficient and scalable deployment. These findings support the feasibility of data-driven, interpretable, and resource-efficient GHI estimation systems for operational and forecasting applications.