Electricity Price Forecasting: Regression vs. Neural Networks

Btissame EL Mahtout, University of Duisburg-Essen and TU Dortmund

Co-authors: Florian Ziel, University of Duisburg-Essen

Abstract: Despite the high performance of linear models in predicting the day-ahead electricity price, they fail to capture nonlinear relationships that might boost the forecasting accuracy. In this paper, we propose a  novel approach that combines linear and nonlinear models to enhance the performance employing two primary architectures. The first architecture is a simple multilayer perceptron (MLP) that connects inputs directly to outputs without any activation function, effectively learning only linear relationships. The second architecture extends this by including both a direct input-to-output layer and an additional hidden layer with a nonlinear activation, thereby capturing more complex patterns. In addition to improving performance, we focus on minimizing computational cost by employing an online learning approach. Our results reveal that the suggested models achieve competitive performance in day-ahead electricity price forecasting in Germany.