Optimized and Explainable ANN-Based Solar Radiation Forecasting for Arid Climates Using SHAP and LIME

Abstract

Accurate and transparent solar radiation forecasting is essential for optimizing photovoltaic (PV) performance in arid climates, where localized atmospheric fluctuations and limited ground-based measurement density pose challenges for operational planning. This study presents an optimized and explainable Artificial Neural Network (ANN) framework for short-term Global Horizontal Irradiance (GHI) forecasting across five climatically diverse Saudi Arabian cities–Riyadh, Tabuk, AlUla, Abha, and Dammam–using hourly reanalysis data from the NASA POWER database.Explainability is integrated through Shapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), which guide feature selection and provide global- and local-level interpretability. The optimized ANN outperforms its baseline configuration and three classical regressors–Random Forest, Support Vector Regression, and Gradient Boosting Regressor. Across all cities, the model achieves normalized root mean square error (RMSE) values between 0.0109 and 0.0114 (based on min–max scaled GHI) and coefficients of determination ( $$R^{2}$$ R 2 ) exceeding 0.99.Benchmarking against Transformer-based models, including the Temporal Fusion Transformer (TFT), Informer, and Patch Time Series Transformer (PatchTST), shows that the ANN delivers comparable accuracy while significantly reducing inference latency and training time. SHAP and LIME consistently identify hour of day, lagged irradiance, and humidity as dominant predictors, aligning with established physical drivers of solar variability.Edge deployment tests on a Raspberry Pi 4 demonstrate sub-2 ms inference latency and low computational overhead, confirming suitability for microgrids, embedded PV controllers, and IoT-enabled energy systems. The proposed approach offers an interpretable, efficient, and deployment-ready solution for real-time solar forecasting in arid and resource-constrained environments.