Tree‐Boost–Guided CNN–BiLSTM–Transformer for Solar Irradiance Forecasting: Cross‐Regional Evidence for Sustainable Energy Planning

ABSTRACT Accurate solar irradiance (SI) forecasting is essential for photovoltaic operation, grid stability, and sustainable energy planning, yet its nonlinear dependence on temporal, meteorological, and radiation‐related variables remains difficult to model across diverse climatic regions. Existing forecasting studies often emphasize single‐site accuracy, with limited attention to dataset‐specific feature relevance, interpretability, ablation‐based architectural validation, and cross‐regional robustness. This study proposes a reproducible tree‐boost–guided convolutional neural network (CNN)–bidirectional long short‐term memory (BiLSTM)–transformer framework for SI forecasting. A scikit‐learn Gradient Boosting Regressor is first used for dataset‐specific feature selection, followed by 24‐h sequence modeling through CNN‐based local temporal extraction, BiLSTM‐based bidirectional dependency learning, and Transformer‐based attention refinement. The framework is evaluated on 7 hourly NASA POWER sites and one utility‐scale QASP ground dataset. The model achieved strong cross‐site performance, with best‐site MAE = 0.129, RMSE = 0.224, and R 2 = 0.947, while on QASP‐GDATA it obtained MAPE = 0.551, MAE = 0.139, RMSE = 0.246, and R 2 = 0.942. Compared with Prophet, Seasonal ARIMA (SARIMA), artificial neural network (ANN), gated recurrent unit (GRU), CNN–LSTM, temporal convolutional network (TCN), and Informer, the proposed model showed robust and competitive performance across different forecasting families. SHAP analysis confirmed physically meaningful predictors, and ablation results verified the contribution of the complete hybrid architecture. Overall, the framework offers an interpretable and scalable forecasting tool for solar planning and grid‐support applications.