Assessment of machine and deep learning models integrated with variational mode decomposition for photovoltaic power forecasting using real-world data from the semi-arid region of Djelfa, Algeria

• Evaluated ML and DL models for PV power forecasting using real-world data. • Incorporated Variational Mode Decomposition for enhanced feature extraction. • CNN outperformed all standalone models for both 30-min and 3-hour forecasts. • ML models, especially ANN, SVM, and RF, benefited more from VMD preprocessing. • Monthly analysis revealed seasonal variability affecting forecasting accuracy. Accurate photovoltaic power forecasting is essential for grid stability and efficient energy management. While Deep Learning (DL) and Machine Learning (ML) models are widely used, the extent to which each can be effectively leveraged remains an open question. This study thoroughly compares several ML and DL models, applied to both short-term (30-minute) and medium-term (3-hour) horizon. The research is built upon real-world data from a 53 MW PV plant located in Algeria, offering practical insights under realistic conditions. Another critical element of the study is the incorporation of Variational Mode Decomposition (VMD) for feature processing, which mainly enhances information extraction. The study also includes a monthly performance analysis investigating the climatological variability on forecasting accuracy. Among standalone models, CNN performs best with an nMAE of 2.9 %, nRMSE of 5.45 %, and R 2 of 0.9678 at 30 min ahead, and nMAE of 3.15 %, RMSE of 4.15 %, and R 2 of 0.9839 at 3 h forecasting. When combined with VMD, ML models, particularly ANN, SVM, and Random Forest frequently outperform DL-VMD counterparts. For instance, ANN achieves an nMAE of 1.08 %, nRMSE of 1.89 %, and R 2 of 0.9961 at 30 min, and maintains excellent accuracy at 3 h with nMAE of 1.1 %, nRMSE of 1.4 %, and R 2 of 0.9982. Collectively, this research serves as a reference for a multidimensional evaluation of forecasting performance. The analysis highlights the importance of selecting appropriate models and preprocessing techniques in PV power forecasting, tailored to location and climatological conditions, contributing to effectively addressing abrupt fluctuations and facilitating large-scale integration.