Multistep PV power forecasting using deep learning models and the reptile search algorithm

Forecasting Photovoltaic (PV) power output is a key challenge in renewable energy systems, particularly for short- to mid-term operational planning. Accurate multi-step PV forecasting supports efficient energy scheduling, grid stability, and integration of solar resources. This study investigates the performance of three advanced deep learning models: Temporal Convolutional Network (TCN), Minimal Gated Unit (MGU), and Temporal Fusion Transformer (TFT), applied to one-day-ahead and three-day-ahead PV power forecasting. The Reptile Search Algorithm (RSA), a novel metaheuristic optimizer, is employed for hyperparameter tuning. Results show that TFT consistently outperforms the other models, achieving Root Mean Square Error (RMSE) values of 6.256 kWh and 8.353 kWh and coefficient of determination (R²) scores of 98.92 % and 98.07 % for the one-day and three-day forecasts, respectively. RSA optimization yields significant performance gains, reducing TFT's RMSE by 44.57 % and 42.43 % relative to its non-optimized baseline. While MGU had the weakest overall performance, particularly under summer conditions and zero-generation periods, it performed better in the one-day-ahead setting than TCN did in the three-day-ahead scenario. Further, RSA was benchmarked against the Whale Optimization Algorithm (WOA), Gray Wolf Optimizer (GWO), and Constrained Particle Swarm Optimization (CPSO), and consistently outperformed all three across all models in both forecasting horizons. However, this is the first study to evaluate MGU in the context of PV forecasting, and its performance may vary under different case studies. It is shown that TFT and RSA offer superior accuracy and generalization across forecast horizons. These findings support the broader adoption and further benchmarking of TFT future research.