Empirical study of day-ahead electricity spot-price forecasting: Insights into a novel loss function for training neural networks

• Published in: Applied energy 2022-08, Vol.319, p.119182, Article 119182
• Main Authors: Loutfi, Ahmad Amine, Sun, Mengtao, Loutfi, Ijlal, Solibakke, Per Bjarte
• Format: Article
• Language: English
• Subjects: Day-ahead forecasting; Loss functions; Machine learning; Neural networks; Optimization
• ISSN: 0306-2619; 1872-9118
• DOI: 10.1016/j.apenergy.2022.119182

•We proposed Theil UII-S as a novel loss function for training neural networks.•Theil UII-S loss function provides more accurate forecasts on the average, best, and, worst case scenarios, converges faster, is twice differentiable, and has a variable gradient.•Theil UII-S loss function is a good candidate for training neural networks for forecasting day-ahead electricity spot prices. Within deregulated economies, large electricity volumes are traded in daily spot markets, which are highly volatile. To develop profitable trading strategies, all stakeholders must be empowered with robust forecasting tools. Although neural network approaches have become increasingly popular for time-series forecasting, they do not optimally capture unique features of financial datasets. A major factor hindering their performance is the choice of the backpropagation loss function. We performed a systematic and empirical study of loss functions that can optimize the forecasting of day-ahead electricity spot prices. We first outlined a set of properties that such a loss function should meet. We proposed Theil UII-S as a novel loss function, which is derived from Theil’s forecast accuracy coefficient. We also implemented five neural network models and trained them on the two most used loss functions—mean squared error and mean absolute error—and our Theil UII-S. We finally tested our models on a real dataset of the electricity spot market of Norway. Our results show that Theil UII-S provides more accurate forecasts on the average, best, and, worst case scenarios, converges faster, is twice differentiable, and has a variable gradient.