Short-term photovoltaic power forecasting in Çanakkale, Türkiye: A comparative study of machine learning, deep learning, and hybrid models

This study investigates the short-term photovoltaic (PV) power forecasting problem using real field data and comparatively evaluates the performance of different forecasting approaches. The study utilizes active power and meteorological data from a 1 MW installed capacity PV plant located in Çanakkale province, northwest Turkey, with a 15-minute sampling interval. The dataset covers the period from August 2022 to August 2024, and only daytime data with solar irradiance above 20 W/m² were considered to minimize the negative impact of zero production on the model. Forecasting performance was analyzed for 15-minute (h = 1) and 60-minute (h = 4) forward forecasting horizons.In the comparative analysis, the persistence method was used as the basic reference model; Ridge regression, support vector regression (SVR), and LSBoost model were used as machine learning-based methods; Long-short-term memory (LSTM) and gated recurrent unit (GRU) networks were evaluated as deep learning-based methods. Additionally, a hybrid VMD+LSTM model combining Variational Mode Decomposition (VMD) with an LSTM network was investigated as a current signal decomposition-based approach. The models were evaluated using RMSE, MAE, normalized RMSE (nRMSE), and coefficient of determination (R²) metrics on a dataset separated by 70% training, 15% validation, and 15% testing ratios without time-order distortion.The results showed that the persistence model offered competitive performance at a very short prediction horizon (h = 1), but the accuracy of this approach decreased significantly as the prediction horizon increased. For 60-minute forward prediction, deep learning models produced more successful results; the optimized LSTM model achieved the best performance with 9.69% nRMSE and an R² value of 0.884. In contrast, while the VMD+LSTM hybrid model produced promising results during the validation phase, it exhibited poor generalization performance on the test set. This finding reveals that decomposition-based hybrid approaches are not superior in all conditions.