11 - AI application for load forecasting: a comparison of classical and deep learning methodologies

Abstract

Electricity demand forecasting has become one of the main research topics in the energy management system. For this reason, many different methodologies for load forecasting have been proposed. Artificial Intelligence forecasting methodologies are divided into classical machine learning and deep machine learning. The most used classic machine learning methodology is multilayer perceptron, whereas long short-term memory (LSTM) and gated recurrent unit (GRU) are the most used deep machine learning methodologies. All the methodologies mentioned require a proper choice of hyper-parameters to improve forecasting. This chapter reviews the three methodologies and the hyper-parameters that influence them. For this purpose, a real dataset, taken from an Irish project with 260 residential consumers sampled every hour, is used. Finally, the methodologies are compared to each other with the best choice of hyperparameters. The three methodologies presented similar results, because the loss function (mean squared error) is optimized with Adam. Adam has adaptive learning rate and is the hyperparameter most difficult to define by classical methodologies. This hyperparameter always had an influence on the improvement of the forecasting.