Artificial Intelligence and Nord Pool’s intraday electricity market Elbas : a demonstration and pragmatic evaluation of employing deep learning for price prediction : using extensive market data and spatio-temporal weather forecasts

Abstract

This thesis demonstrates the use of deep learning for automating hourly price forecasts in continuous intraday electricity markets, using various types of neural networks on comprehensive sequential market data and cutting-edge image processing networks on spatio-temporal weather forecast maps. Deep learning is a subfield of artificial intelligence that excels on problems such as these with multifarious input data and manifold interacting factors. It has seen tremendous success on a range of problems across industries, and while it is important to have realistic expectations, there is little reason to believe that intraday electricity markets are different. Focusing on Nord Pool’s intraday market Elbas, we predict Nordic buyers’ volume-weighted average price over the last six hours of trading prior to each delivery hour. Aggregating this window gives buyers flexibility from many trades and sufficient time in which to act on the predictions, and solves issues with data sparsity while keeping sufficient resolution for predictions to be informative. We develop various neural networks via extensive experimentation, with inspiration from other research and problem domains. To make the findings relevant in practice, we impose constraints on the input data based on what would be available to Elbas market participants six trading hours ahead of delivery. The neural networks are benchmarked against a set of simple domainbased heuristics and traditional methods from econometrics and machine learning. We conclude with a holistic evaluation of the efficacy of deep learning on our problem, whether it is economically justifiable in light of its value-add, what the salient hurdles are to implementing it in practice, and what the implications are for broader applications of AI in intraday markets. The deep learning models1 are relatively accurate and reliable under normal market conditions. The average price across all delivery hours in the held-out data is 30.95 EUR/MWh, where our best network is on average off by 2.72 EUR/MWh. It beats the best simple heuristic by 21–25%, and the best benchmark model by 12–16%. The network also anticipates major fluctuations in prices relatively consistently, and generally outperforms all alternative methods when prices are especially volatile or trading activity particularly high. In contrast to the benchmarks, there are also ample avenues for improving the network further. Beyond being promising in its own right, we also argue that the network demonstrates the wider potential of deep learning in a range of applications in intraday markets, and that these are worthy of serious consideration — though one should be aware of the practical hurdles to implementing them operationally.