A Deep Learning Technique for Electricity Price Forecasting in Consideration of Spikes

Abstract

This paper presents a Deep Neural Network (DNN) method for electricity price forecasting in power markets. They are inclined to generate spikes that are dozens to hundred times as large as the normal prices so that the prediction is hard to handle. This paper focuses attention on the prediction of spikes to suppress the forecasting errors. This paper deals with the pretraining technique of Autoencoder (AE) in Deep Learning. To enhance the performance of AE, this paper presents a Denoising-Autoencoder (DAE)-based method that consists of DAE and Multilayer Perceptron (MLP) of ANN with the clustering technique. DAE is an extension of AE in a sense that noisy learning data is used with random numbers. The use of clustering enhances the model accuracy due to data similarity. The effectiveness of the proposed method is tested for data of New England ISO, USA.