Comparative Analysis of Multivariable Deep Learning Models for Forecasting in Smart Grids

Abstract

Clean-energy generation in smart grids is limited by the availability of the energy to be transformed and advanced energy management strategies requires solid and anticipated information about its dynamic behavior. This includes multivariable prediction of meteorological and user consumption data simultaneously in time series. The selection of a predicting model, from long short-term memory (LSTM), convolutional neural networks (CNN), gated recurrent units (GRU), or their hybrid models merging CNN with LSTM and GRU, is a very complex task. In this paper, a mean absolute error, absolute percentage error (MAPE), and root mean square error (RMSE) comparative analysis, for prediction of energy consumption, and solar and onshore wind generation, is presented. A three-day prediction-horizon is used, with four-year hourly training data from Madrid. The combination of the best GRU and CNN models found, subject to the given hyperparameters grid, has a better prediction performance, including if they predict separated. Relevant information about training and coding appreciations is also presented.