Hybrid VMD-TCN-GRU Model for Short-Term Photovoltaic Power Forecasting: A Case Study of 20 MW and 3 MW Grid-Connected Solar Power Plants in Adrar, Algeria

Abstract

Precise Short-Term Photovoltaic Power Forecasting (STPVPF) is crucial for the reliable integration of solar energy into modern power systems. This study investigates data-driven forecasting strategies for two grid-connected PV power plants located in Adrar, Algeria, with installed capacities of 20 MW and 3 MW. Two deep learning models, Temporal Convolutional Networks (TCNs) and Gated Recurrent Units (GRUs), are evaluated and further extended through a hybrid framework based on Variational Mode Decomposition (VMD). The developed VMD–TCN–GRU scheme decomposes the PV power signal into intrinsic modes prior to learning, enabling improved representation of multi-scale temporal patterns. The models are trained and validated using long-term operational data collected from the two utility-scale PV power plants, with performance assessed under different historical input windows. Results show that VMD substantially improves forecasting accuracy. At the respective optimal input windows, the VMD–TCN–GRU model achieves rRMSE values of 7.11% for the 20 MW Adrar power plant and 10.82% for the 3 MW Kaberten power plant, corresponding to relative rRMSE reductions of 42.43% and 28.96%, respectively, compared with the corresponding non-decomposed TCN–GRU models. The optimal configurations achieve coefficients of determination of 98.89% for Adrar and 97.16% for Kaberten. The results also show that the optimal historical input window is power plant-dependent. Overall, the findings demonstrate that combining VMD with hybrid deep learning provides an effective forecasting framework for short-term PV power prediction under real-world grid-connected operating conditions.