Enhancing Solar Energy Forecasting: 1D CNN With Explainable AI Techniques

Accurate solar energy forecasting is essential for optimizing renewable energy systems and supporting SDG 7: Affordable and Clean Energy and SDG 13: Climate Action. This paper presents a 1D CNN framework for day-ahead solar irradiance prediction, using data from the Global Ensemble Forecast System and the “KENT” Mesonet station, as part of the American Meteorological Society’s Solar Energy Prediction Contest. Enhanced with residual inception squeeze-and-excitation blocks and attention mechanisms, our model improves forecast accuracy, reducing mean absolute error (MAE) by 6.56% over the contest’s best model. Explainable AI methods, including Integrated Gradients and Permutation Feature Importance, identify key influences such as downward shortwave radiation and surface temperature. Future work will explore transformer models for smart grid applications, further supporting SDG 9: Industry, Innovation, and Infrastructure.