Integration of Updated Meteorological Data in Machine Learning Models for Accurate Photovoltaic Forecasts

This paper presents the implementation and performance comparison of different machine learning techniques to forecast Photovoltaic (PV) production for the next 23 hours. Specifically, the performance of Recurrent Neural Networks (RNNs) and ensemble learning methods are compared. The models receive input values of meteorological variables as well as the PV production over the previous 23 hours. A key feature is that the models are designed to hourly receive updated inputs, including the most recent measurements of both meteorological variables and PV production, which avoids relying on outdated information. Inputs are accessed from open repositories that provide measurements with sufficient timeliness to allow the algorithm to be rerun hourly. To assess the usefulness of receiving hourly updates of the model inputs, the forecasting ability of the algorithms are compared with the ones obtained when inputs are updated daily. The results obtained from applying the models to a real PV plant show that RNNs produce better estimates than ensemble learning models. Besides, outcomes show that incorporating the most recent available data significantly improves forecasting performance. This confirms that updating the model with the hourly measurements leads to highly accurate predictions of PV production, provided that data are reliable and timely available.