A Variational Mode Decomposition-Enhanced Deep Convolutional Network for Day-Ahead Solar Irradiance Prediction

The effective integration and operation of photovoltaic (PV) generation in power systems are critically dependent on the availability of reliable and accurate day-ahead global horizontal irradiance (GHI) forecasts. Traditional methods, however, face serious issues due to strong nonstationary and nonlinear GHI characteristics that are influenced by varying atmospheric conditions. To address these issues, this paper proposes a hybrid forecasting methodology that combines variational mode decomposition (VMD) and the convolutional neural network (CNN) for day-ahead GHI prediction. In the proposed method, the VMD approach is specifically applied to the GHI to reduce its non-stationary characteristics, while the temperature, RH, and clear sky GHI are used in their raw form to maintain their physical interpretability. The CNN model then uses the meteorological variables and the decomposed GHI components as inputs to identify significant temporal patterns. The VMD-CNN approach is validated using real-world scenarios with three different Indian Locations-Jodhpur, Kochi, and Manali-and tested using benchmark models such as RF, ANN, and standalone CNN. The results demonstrate consistent improvements across all locations, with the proposed technique achieving an average reduction of 33.52% in RMSE and 33.7% in MAE compared to the CNN model, along with high$\mathbf{R}^{\mathbf{2}}$across all locations. The outcome of these results highlights the effectiveness and robustness of the presented framework in effectively capturing complex irradiance characteristics, which makes it suitable for real-world day-ahead solar forecasting applications.