Multivariate prediction in photovoltaic power plants by a stacked deep neural network

In this paper, a new approach on energy time series prediction is carried out. We propose a deep learning technique with the employment of specific neural network architectures: Convolutional Neural Network and Long Short-Term Memory network. The goal is to exploit the correlation between several time series, joining and filtering them together as to bring out the long-term dependencies among all the observations. We superpose many different functional layers, thus providing a stacked scheme that can result in a greater approximation capability. The novel architecture is assessed in a real-world prediction problem, in order to evaluate the performance regarding prediction accuracy.