An electricity price interval forecasting by using residual neural network

This article proposed a new electricity price interval forecasting method based on a novel Residual Neural Network (ResNet) for the electricity price interval forecasting. The significant outcome of the ResNet model was that the model performs excellently on normal and spike price interval forecasting in accuracy and reliability point of view. The proposed ResNet was consisting of two network layers. The first neural network layers were probabilistic normal, high, and spike prices prediction part. The Lower and Upper Bound Estimation (LUBE) formulates the price interval forecasting from the output of the second neural network layers. The LUBE methods included Quantile Regression and Mean and Variance estimation. The proposed forecasting models were demonstrated with the GEFCom2014 dataset. The dataset is consisting of 15 tasks for electricity prices forecasting. The results of proposed ResNet models compared with GEFCom2014's benchmarks, Quantile Regression Average, and Multilayer Perceptron Network approaches. The performances of forecasting models are evaluated in terms of accuracy and reliability metrics by Pinball Loss Function and Coverage Width-based Criterion (CWC), respectively. Also, this article shows the effect of an increase in confidence level, which could generate lower CWC values and represent high reliability's satisfaction.