Explaining Solar Radiation Concurrent Estimation and Conditional Multi-Horizon Prediction with SHAP and LIME: Interpretable CatBoost Benchmarking Using Long-Term NASA POWER Data
Аннотация
Abstract Accurate hourly solar-radiation estimation is important for photovoltaic and solar-cell applications because irradiance directly influences power generation, conversion behavior, and operational decision-making. This study develops a unified benchmarking and explainable artificial intelligence (XAI) framework for hourly solar-radiation modelling under two settings: (i) concurrent estimation and conditional (ii) multi-horizon prediction. Twenty-four years of hourly NASA POWER meteorological data for San Diego were used to benchmark eighteen statistical, machine-learning, and deep-learning models for (i) Task A: Concurrent Estimation $$\left( {h = 0,\;X\left( t \right) \to {\text{GHI}}\left( t \right)} \right)\;$$ and (ii) Task B: Conditional Multi-Horizon Prediction $$\left( {h = 1,3,6,24,{\;}X\left( t \right) \to {\text{GHI}}\left( {t + h} \right)} \right)$$. In Task B, only predictors available at time $$t$$ were used; therefore, the multi-horizon results should be interpreted as retrospective conditional prediction baselines under time-$$t$$ information rather than as a fully operational forecasting system with forward-looking forecast inputs. Model performance was evaluated using RMSE, MAE, MSE, $${R^2}$$, normalized RMSE, and a persistence-relative skill score. For concurrent estimation, CatBoost achieved the highest accuracy, with RMSE = 4.99 Wh/m 2 , MAE = 3.67 Wh/m 2 , MSE = 24.89 (Wh/m 2 ) 2 and $${R^2} = 0.9997$$. For conditional multi-horizon prediction, CatBoost provided the most robust overall performance across horizons, achieving RMSE = 73.71 Wh/m 2 at $$h = 1$$ and RMSE = 197.18 Wh/m 2 at $$h = 24$$, while consistently outperforming the persistence baseline. To address the interpretability limitations of high-performing ensemble models, SHAP and LIME were applied to the selected CatBoost models for both Task A and Task B. The XAI analysis showed that concurrent estimation is governed primarily by the Clearness Index, Solar Zenith Angle, and UV Index, whereas the conditional multi-horizon setting preserves the importance of these physically meaningful predictors but exhibits horizon-dependent changes in their relative influence, with broader meteorological variables becoming more relevant at longer lead times. Global SHAP results, dependence analysis, and local SHAP–LIME case studies consistently revealed physically coherent and horizon-aware explanatory patterns. Overall, the proposed framework provides not only strong predictive benchmarking but also transparent, physics-consistent interpretation for retrospective solar-radiation estimation and conditional multi-horizon prediction.