Transformer-based neural network for short-term photovoltaic and wind power prediction

Abstract

Short-term photovoltaic (PV) and wind forecasting are essential for managing connected systems to the electrical grid and purchasing and selling energy in the daily and intraday market. Thus, the development of accurate prediction models becomes necessary in the dynamics of the electricity sector. In this context, machine learning models are widely used for their excellent performance in the complex extraction of atmospheric features that directly induce renewable production. This work proposed using a vector representation learning model as embedding, namely Time2Vec, to improve the learning models used for prediction and a transformer-based neural network architecture. The experiment is carried out on two different PV power plants in India and two wind farms using the proposed architecture, Multilayer Perceptron (MLP) and Long-Short Term Memory (LSTM), which are then compared to several models used as baseline prediction in this type of forecasting. In most cases, the results showed a relevant improvement over the reference models, reaching more than 20% improvement over the mean absolute error and correlation coefficient in some horizons.