Deep Learning for Photovoltaic Power Plant Forecasting

Abstract

Deep Learning is getting a relative importance in the field of machine learning due to better performance in fields of classification and pattern recognition. However, deep models have seen little use in time series forecasting. Thus, the purpose of this work is to investigate the performance of such models in power plant output forecasting. A classical Artificial Neural Network with one hidden layer and two Deep Learning models were developed to forecast the output from a photovoltaic power plant. A Recurrent Deep Neural Network with Long Short Term Memory and a Deep Neural Network were proposed to predict future values; trained by the Adam algorithm and validated using R, RMSE and MAPE statistical criteria. Using deep models improves the accuracy of forecasting better than models without a large hidden layer size. This improvement is demonstrated by training several structures of Deep Models and feed forward Neural Networks models. Correlation coefficient of 1.0 is achieved using a deep architecture for this case study.