Abstract
Accurate photovoltaic (PV) power prediction is hindered by environmental stochasticity. Existing physical models often overlook ''atmospheric disturbers''—such as wind-induced cooling and humidity-based spectral absorption—while black-box deep learning models lack physical interpretability and stability. To address this gap, this study proposes an Atmospheric-Aware Residual Deep Learning (AARDL) framework. This ''Grey-box'' approach integrates a calibrated 8-parameter physical model with a 1D-ResNet to map and compensate for systematic residuals. By incorporating wind speed and humidity as input features, the model captures previously unmodeled dynamics. Validated on seasonal datasets from September 2018 and April 2019, the AARDL framework achieved significant error reductions, lowering RMSE from 36.11 V to 3.04 V (91.59%) and 36.78 V to 2.09 V (94.32%), respectively. Sensitivity analysis confirms that while temperature is the primary driver, atmospheric features are vital for high-fidelity precision. This research offers a scalable, computationally efficient solution for real-time IoT monitoring and high-fidelity PV digital twins.