Neural networks and arima for short-term electricity price forecasting

Abstract

ABSTRACT: Electricity price forecasting provides significant information for the different elec tricity market agents so that their profits can be maximized. This work is meant to make a univariate and multivariate comparison between state-of-the-art statistical models such as ARIMA and Transfer Function Models, and the promising Deep Learn ing models, such as Recurrent Neural Networks and Convolutional Neural Networks, in order to make 24 hours ahead predictions of the electricity price in the Spanish elec tricity market for the 2020 timespan. In addition, an ensembling model composed of models from both backgrounds will be suggested to improve the predictions of either individual model. In the experiments, Convolutional Neural Networks outperformed all other Neural Networks at univariate and multivariate level and had similar results to the state-of-the-art statistical models at univariate level, outperforming them at mul tivariate level. Additionally, it has been shown that the ensembling model obtains considerably better results than each of the individual models