Performance of alternative electricity price forecasting methods: Findings from the Greek and Hungarian power exchanges

• Published in: Applied energy 2020-11, Vol.277, p.115599, Article 115599
• Main Authors: Halužan, Marko, Verbič, Miroslav, Zorić, Jelena
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Data mining; Electricity price forecasting; ENTSOE-E transparency platform; Greek power exchange; Hungarian power exchange; Learning sample size; Machine learning; Rolling-window
• ISSN: 0306-2619; 1872-9118
• DOI: 10.1016/j.apenergy.2020.115599

•The paper forecasts the Greek and the Hungarian day-ahead electricity price.•Forecasting is performed based on the ENTSO-E Transparency Platform data.•Rolling-window analysis is based on more than 1000 out-of-sample forecasting days.•Linearity bias is reduced with selected data mining and machine learning algorithms.•Impact of learning sample size on forecasting performance is examined. This paper evaluates the performance of alternative algorithms for day-ahead electricity price forecasting. Forecasting performance is assessed based on evidence from the Greek and Hungarian Power Market simulation. The electricity price formation process is simulated on a long time series spanning from January 2015 to September 2018. The EPF models are structured upon the explanatory variables that are available to the market participants before the exchange gate closure, through the publicly available ENTSO-E transparency platform. Relationships between the electricity spot price and explanatory variables are estimated by the selected econometric, data mining, and machine learning algorithms. The econometric autoregressive model with exogeneous explanatory variables is a benchmark model, as the other alternative approaches are used to overcome the linearity bias in the ordinary least squares estimator. We analyse the impact of a different training sample size as well as the impact of training on an hourly clustered sample on the forecasting performance. The support vector machine algorithm turned out to be the best alternative approach, with the lowest mean absolute error and statistically confirmed better forecasts compared to the benchmark econometric autoregressive model. The majority of the tested algorithms perform better with smaller training samples, whereas neural network based approaches prefer large training samples. Models with hourly clustered training samples have higher accuracy based on the Hungarian evidence, while hourly non-clustered training is a superior training method based on the Greek findings.