Abstract

In the current context of energy transition, accurately forecasting solar power production is a major challenge for the efficient operation of electricity grids that include renewable energy sources. This work aims to improve prediction performance by combining a nonlinear physical model based on the principles of photovoltaic conversion with a recurrent neural network designed for time series analysis. The underlying hypothesis is that integrating physical knowledge with data-driven learning can better capture the complexity of solar energy patterns. The proposed method involves careful selection of input variables through correlation analysis and embedding the physical model within a deep learning structure. Results, evaluated using standard error metrics, demonstrate a clear improvement in forecasting accuracy compared to conventional approaches. While the physical model considered on its own produces a high error level (root mean square error = 338.55 and mean absolute error = 182.08), methods based on artificial intelligence significantly reduce these values (long short-term memory network: root mean square error = 3.29; recurrent neural network: 2.87). The hybrid method developed in this study achieves the best overall performance (root mean square error = 2.83 and mean absolute error = 1.26). This study contributes to the development of more reliable prediction systems capable of anticipating fluctuations in solar power generation due to changing environmental conditions.