本文为挪威经济学院（作者：Johannes Krokeide Kolberg & Kristin Waage）的硕士论文，共138页。

本文介绍了利用深度学习技术实现连续日内电力市场小时电价预测的自动化，在综合序列市场数据上使用各种类型的神经网络，在时空天气预报图上使用尖端的图像处理网络深度学习是人工智能的一个分支，它擅长于处理各种输入数据和多种交互因素的问题，已经在各个行业的一系列问题上都取得了巨大成功，尽管现实的预期很重要，但几乎没有理由相信日内电力市场是不同的。以Nord Pool的日内市场Elbas为重点，我们预测了北欧买家在每个交割时间之前的最后六个交易小时的成交量加权平均价格。

我们在其他研究和问题领域的启发下，通过广泛的实验开发了各种神经网络。为了使调查结果在实践中具有相关性，我们对输入数据施加了限制，这是基于Elbas市场参与者在交割前6个交易小时可以获得的数据。神经网络的基准是一套简单的基于领域的启发式的传统方法，来自于计量经济学和机器学习。最后，我们对深度学习的有效性进行了全面评估，从其附加值来看，深度学习是否在经济预测上合理，在实践中实施深度学习的主要障碍是什么，以及人工智能在日内市场中更广泛应用的影响是什么。

在正常市场条件下，深度学习模型1相对准确可靠。延迟数据中所有交货时间的平均价格为30.95欧元/兆瓦时，其中我们的最佳网络平均价格为2.72欧元/兆瓦时。它比最好的简单启发式方法高出21 - 25%，比最好的基准模型高出12 - 16%。该网络还相对一致地预测价格的重大波动，当价格波动较大或交易活动特别高时，通常优于所有其他方法。与基准相比，还存在着进一步改善网络的充分途径。除了本身具有希望之外，我们还认为，该网络在日内市场的一系列应用中显示出更广泛的深度学习潜力，这些都值得认真考虑——但人们应该意识到在实际操作中实施这些方法时可能遇到的障碍。

This thesis demonstrates the use of deeplearning for automating hourly price forecasts in continuous intraday electricitymarkets, using various types of neural networks on comprehensive sequentialmarket data and cutting-edge image processing networks on spatio-temporalweather forecast maps. Deep learning is a subfield of artificial intelligencethat excels on problems such as these with multifarious input data and manifoldinteracting factors. It has seen tremendous success on a range of problemsacross industries, and while it is important to have realistic expectations,there is little reason to believe that intraday electricity markets aredifferent. Focusing on Nord Pool’s intraday market Elbas, we predict Nordicbuyers’ volume-weighted average price over the last six hours of trading priorto each delivery hour. Aggregating this window gives buyers flexibility frommany trades and sufficient time in which to act on the predictions, and solves issueswith data sparsity while keeping sufficient resolution for predictions to beinformative.

We develop various neural networks viaextensive experimentation, with inspiration from other research and problemdomains. To make the findings relevant in practice, we impose constraints onthe input data based on what would be available to Elbas market participantssix trading hours ahead of delivery. The neural networks are benchmarkedagainst a set of simple domain based heuristics and traditional methods fromeconometrics and machine learning. We conclude with a holistic evaluation ofthe efficacy of deep learning on our problem, whether it is economicallyjustifiable in light of its value-add, what the salient hurdles are toimplementing it in practice, and what the implications are for broaderapplications of AI in intraday markets.

The deep learning models1 are relativelyaccurate and reliable under normal market conditions. The average price acrossall delivery hours in the held-out data is 30.95 EUR/MWh, where our bestnetwork is on average off by 2.72 EUR/MWh. It beats the best simple heuristicby 21–25%, and the best benchmark model by 12–16%. The network also anticipatesmajor fluctuations in prices relatively consistently, and generally outperformsall alternative methods when prices are especially volatile or trading activityparticularly high. In contrast to the benchmarks, there are also ample avenuesfor improving the network further. Beyond being promising in its own right, wealso argue that the network demonstrates the wider potential of deep learningin a range of applications in intraday markets, and that these are worthy ofserious consideration — though one should be aware of the practical hurdles toimplementing them operationally.

1. 引言

2. 北欧电力市场

3. Elbas价格决定因素的识别

4. 数据

5. 研究方法

6. 分析与结果

7. 讨论

8. 结论

附录A 理论：深度学习

附录B 关于Elbas的其它信息

附录C 标准化还是归一化？

附录D 深度学习模型

附录E 基于标准模型的理论

附录F 市场数据变量

更多精彩文章请关注公众号：