A Predictive Deep Learning Model to Estimate the Power Generated by a Photovoltaic Plant

This work presents a predictive model based on deep learning to estimate the electrical power generated by a photovoltaic rooftop plant in the very short term. LSTM and GRU models and variants were implemented in the predictive system as a basis for their excellent performance. The relationship between the variables in the database was analyzed to define an architecture that takes into account both weather and relevant physical input variables, including the electric power generated 24 hours before the moment to be predicted. The final model consists of a LSTM architecture with 3 layers and sequence length of size 50, which has improved the results of models such as NARX and SVR, with respect to previous published results based on a CNN-GRU architecture on the same dataset. The experimental investigations and case of study took place in the city of Mar del Plata in the southeast of Argentina. This study has also revealed geographical issues in this coastal city and how water vapor and wind interact and affect power efficiency.