Photovoltaic energy forecast; influence of two numerical weather forecast datasets on the performance of an analytical and three machine learning models

The European Union aims to reduce reliance on fossil fuels by achieving a 42.5% share of renewable energy sources by 2030. Self-consumption and the associated energy management systems (EMS) are essential to the integration of renewable energy sources, of which photovoltaic energy is a key component. As part of the EMS tasks, photovoltaic power forecasting is critical due to the weather-related variability in generation. This research study evaluates the performance of 24-hour photovoltaic production forecasting considering meteorological data obtained from two different numerical weather prediction models, the European ECMWF and the Galician MeteoGalicia, three machine learning (ML) models, Feedforward Neural Networks (FFNN), Support Vector Regression and Nonlinear Autoregressive Neural Network with Exogenous Inputs, and an analytical model. Results show that the FFNN performs best, especially in summer, with a R 2 of 0.9 when using predicted weather data. Furthermore, ML models trained with MeteoGalicia data outperform ECMWF-based models. For instance, the FFNN obtains an improvement over benchmark of 8.6% with MG data and 5.7% with ECMWF data, during November. Winter forecasting challenges highlight the need for good ML models to address variability. Moreover, analytical models underperformed compared to ML methods when using forecasted weather data, emphasizing the partial compensation from ML models for weather prediction errors.