Residual-enhanced hybrid deep learning framework for day-ahead PV power forecasting with cross-site generalization

Accurate day-ahead photovoltaic (PV) power forecasting is still a challenging task due to the nonlinearities in solar irradiance, structured temporal dependencies, and climatic variability across regions. In this study, a structured residual-enhanced hybrid forecasting methodology is introduced, in which the forecasting problem is decomposed into a primary nonlinear regression task and a secondary residual sequence learning task. The proposed method combines a physics-informed PV power modeling approach, SHapley Additive explanations-based feature selection, leakage-free chronological data preprocessing, Bayesian hyperparameter optimization of a multilayer perceptron (MLP) architecture, and structured residual sequence modeling using a long short-term memory autoencoder. The residual sequence modeling task is performed independently and integrated with the primary forecasting task to create a hierarchical hybrid forecasting architecture. The effectiveness of the proposed method is validated using high-resolution (10-min time-step) PV power data from climatically distinct regions in India the humid coastal region of Mumbai and the semi-arid region of Udaipur. Significant performance gains are observed over a traditional single-regional forecasting paradigm in comparison with the optimized results of the MLP on the Mumbai dataset, the hybrid approach resulted in an improvement of 48.8% in terms of mean absolute error (MAE), with an improvement in R² values from 0.9467 to 0.9882. The cross-validation results verified the generalization capability of the approach, while adaptive residual fine-tuning resulted in up to 63.6% improvement in terms of MAE under domain shift scenarios. Statistical hypothesis test results verified the robustness of improvements, thus validating that explicit modeling of structured residuals leads to improved forecasting accuracy, robustness, and adaptability, which can be scaled up for reliable PV power forecasting in heterogeneous climatic environments.