Day-ahead forecasting of regional photovoltaic production using deep learning

Power production based on solar energy is directly related to the state of the atmosphere. As the atmospheric state is undergone, this connection makes the solar energy a non-dispatchable source as opposed to controllable renewable sources such as hydroelectricity. In a context of growing photovoltaic generation, accurate forecast tools at regional scale are then increasingly important to grid operators. Indeed, forecasts allow getting information about future production over the next minutes, hours and days. Forecasting tools offer the possibility of a better grid management strategy specifically for transmission system operator (TSO) that are responsible for balancing renewable power production. High forecast accuracy could also lead to reduced costs for energy trading. In light of this situation, this study focuses on the development and analysis of a regional forecasting tool based on a deep learning approach. The selected model consists in a combination of a convolutional neural network (CNN) with a long short-term memory architecture (LSTM). The CNN layers allow extracting spatial features from Numerical Weather Prediction outputs while the LSTM part supports the temporal relationship. The day ahead regional forecast for Germany is chosen as a case study. The CNN-LSTM is compared to the classical Random Forest model known to be one of the reference techniques for this kind of problematic. Simpler deep learning models are also tested to validate the improvement brought by the CNN-LSTM architecture. All the comparisons are based on the classical root mean squared error (RMSE) metrics. The main result of this study shows that CNN-LSTM model can improve forecast accuracy when compared to state-of-the-art Random Forest. As expected, this improvement is strongly correlated to the amount of historical data which must cover several years according to the sensitivity study realized in this work.