Forecasting Electricity Prices for Intraday Markets with Machine Learning : An exploratory comparison of the state of the art

Abstract: Electricity needs to be consumed when it is produced, making sure that supply closely meets demand at all times. To account for the rapidly changing operational status and the need for increasing the flexibility of power systems, financial instruments have been put in place creating markets where electricity is traded as a commodity across different time frames; from months or days to minutes before, or even after, planned delivery. In this work, the focus is placed on the short-term electricity markets and particularly on forecasting the intraday volume-weighted average price of the last three hours of trading of hourly power products. To this end, two state-of-the-art recurrent neural network architectures, namely the Temporal Fusion Transformer and the DeepAR network, are compared against well-established statistical models, such as the Linear Regression, ARX and SARIMAX models, with respect to their forecast accuracy on each of the 24 hourly delivery products. Two different experimental setups are applied, with one utilizing two input features drawn specifically from the findings of relevant literature and the other blindly exploiting all available streams of information in either their raw or aggregated form. All models are trained individually per hourly product per experimental setup to support a fair and decisive comparison, leading to 240 unique model instances being trained in total. Furthermore, the input feature importance is inferred by exploiting the inbuilt attention mechanism of the Temporal Fusion Transformer architecture. Finally, by using various realworld historical market data originating from the Nord Pool power exchange as well as from the Svenska Kraftnät, available up until the day of delivery, it is shown that the statistical models outperform both contemporary neural network architectures, with the latter suffering from the inability to generalize to elevated price levels—which are absent from the training dataset.