A Multivariate Prediction Model for Short-Term Photovoltaic Plant Generation Using Bi-LSTM and CNN

Abstract

The short-term prediction of the energy produced by a photovoltaic plant is a widely studied topic, and it is an important issue for the stability of the grid and its correct operation, as well as for reducing the operating costs and increasing the lifetime of the elements that make up it. The creation of a tool to more accurately predict the solar generation of the PV plant, specifically the prediction of ramps 5-10 minutes in advance. In this work, a multivariate prediction model is presented that combines images, historical production data, and solar position at each moment. The model consists of two parts: image processing, with a convolutional neural network (CNN) and time series processing using a Bidirectional Long Short-Term Memory (Bi-LSTM) capable of detecting long-term nonlinear features. CNNs will be trained to automatically detect the relationship between the images taken of the sky and cloud movement and the current power of the solar array. Then, the recurrent neural networks (RNNs) created will be used to give rise to a 5-minute prediction and a 10-minute prediction. The prediction results are compared using different error metrics, like skill score and the mean squared error (RMSE).