In the future, it could be beneficial to study more thoroughly the effect of different factors and factor combinations that were used in this study and also some that were not. For example, the use of other weather factors and weather forecasts could be useful. Also implementing some hyperparameter tuning for ReduNet 6, which was the best DeepAR model, could improve the learning of the model and produce more accurate predictions.
Since DeepAR models appear to have problems forecasting the fast changes of the imbal-ance power price, it could be beneficial to try out some Dynamic Time Warping (DTW) method with the DeepAR. One example is presented in [24]. The paper proposes quite promising results against Euclidean loss when producing time-series predictions, espe-cially when there are sharp changes in the data.
7 CONCLUSION
Using available literature, multiple machine learning methods, that were implemented into problems in the electricity market or some similar environment, were compared.
DeepAR was chosen as the studied method since based on the comparison it had the most promising results. SARIMAX was chosen as the benchmark method due to its universality and the assumed seasonality of the data.
A wide variety of factors related to the electricity market and weather were used as ex-ogenous factors. Data was gathered from a three-year period and divided into training, validation, and test sets. Both models were experimented with different combinations of hyperparameters. Moreover, resampled data, recursive prediction, and reduced number of factors were tried with DeepAR model.
The SARIMAX models suffer from an inability to adjust to the different magnitudes of the prices, whereas the DeepAR models’ biggest problem is the lack of learning while training the model. Neither of the models gives accurate predictions, but in the end DeepAR model’s predictions are considered better, since its MAE and MSE values are smaller.
The problems of the models may relate to the high complexity of the imbalance power price or the use of factors, that do not contain information about the price. In the fu-ture, these issues could be tackled with more in-depth experiments with different factors, studying the phenomenon more precisely, or experimenting with some other methods.
REFERENCES
[1] Nord pool - trading. https://www.nordpoolgroup.com/trading/, 2020.
[2] Nika Nizharadze, Arash Farokhi Soofi, and Saeed D. Manshadi. Learning the Gap in the Day-Ahead and Real-Time Locational Marginal Prices in the Electricity Market.
arXiv:2012.12792 [cs, eess], December 2020. arXiv: 2012.12792.
[3] Fingrid. https://www.fingrid.fi/en/.
[4] Umut Ugurlu, Ilkay Oksuz, and Oktay Tas. Electricity Price Forecasting Using Recurrent Neural Networks.Energies, 11(5):1255, May 2018. Number: 5 Publisher:
Multidisciplinary Digital Publishing Institute.
[5] Enislay Ramentol, Florian Schirra, and Andreas Wagner. Short- and long-term fore-casting of electricity prices using embedding of calendar information in neural net-works. arXiv:2007.13530 [stat], July 2020. arXiv: 2007.13530.
[6] Epex spot. https://www.epexspot.com/en/tradingproducts.
[7] Youru Li, Zhenfeng Zhu, Deqiang Kong, Hua Han, and Yao Zhao. EA-LSTM:
Evolutionary Attention-based LSTM for Time Series Prediction. arXiv:1811.03760 [cs, stat], November 2018. arXiv: 1811.03760.
[8] Hao Li, Y. Shen, and Yanmin Zhu. Stock Price Prediction Using Attention-based Multi-Input LSTM. InACML, PMLR 95:454-469, 2018.
[9] Joel Janek Dabrowski, YiFan Zhang, and Ashfaqur Rahman. ForecastNet: A Time-Variant Deep Feed-Forward Neural Network Architecture for Multi-Step-Ahead Time-Series Forecasting.arXiv:2002.04155 [cs, stat], 12534:579–591, 2020. arXiv:
2002.04155.
[10] Yitian Chen, Yanfei Kang, Yixiong Chen, and Zizhuo Wang. Probabilistic forecast-ing with temporal convolutional neural network. Neurocomputing, 399, July 2020.
[11] Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. Modeling Long-and Short-Term Temporal Patterns with Deep Neural Networks. arXiv:1703.07015 [cs], April 2018. arXiv: 1703.07015.
[12] Siteng Huang, Donglin Wang, Xuehan Wu, and Ao Tang. DSANet: Dual Self-Attention Network for Multivariate Time Series Forecasting. InProceedings of the 28th ACM International Conference on Information and Knowledge Management, pages 2129–2132, Beijing China, November 2019. ACM.
[13] David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. DeepAR:
Probabilistic forecasting with autoregressive recurrent networks.International Jour-nal of Forecasting, 36(3):1181–1191, July 2020.
[14] Ilkay Oksuz and Umut Ugurlu. Neural network based model comparison for in-traday electricity price forecasting. Energies, 12(23):4557, jan 2019. Number: 23 Publisher: Multidisciplinary Digital Publishing Institute.
[15] Aleksei Mashlakov, Toni Kuronen, Lasse Lensu, Arto Kaarna, and Samuli Honka-puro. Assessing the performance of deep learning models for multivariate prob-abilistic energy forecasting. Applied Energy, 285:116405, March 2021. ZSCC:
NoCitationData[s1].
[16] Rob J Hyndman and George Athanasopoulos. Forecasting: Principles and Practice (2nd ed). Melbourne, Australia, 2018.
[17] Christopher Bennett, Rodney A. Stewart, and Junwei Lu. Autoregressive with Ex-ogenous Variables and Neural Network Short-Term Load Forecast Models for Resi-dential Low Voltage Distribution Networks. Energies, 7(5):2938–2960, May 2014.
Number: 5 Publisher: Multidisciplinary Digital Publishing Institute.
[18] Fingrid open data. https://data.fingrid.fi/en/.
[19] Entso-e. https://transparency.entsoe.eu/dashboard/show.
[20] Kendall tau metric. encyclopedia of mathematics. http://
encyclopediaofmath.org/index.php?title=Kendall_tau_
metric&oldid=51572.
[21] Emad A. Mohammed, Christopher Naugler, and Behrouz H. Far. Chapter 32 -Emerging Business Intelligence Framework for a Clinical Laboratory Through Big Data Analytics. In Quoc Nam Tran and Hamid Arabnia, editors,Emerging Trends in Computational Biology, Bioinformatics, and Systems Biology, Emerging Trends in Computer Science and Applied Computing, pages 577–602. Morgan Kaufmann, Boston, January 2015.
[22] Statsmodels. https://www.statsmodels.org/stable/index.html.
[23] Gluonts. https://www.statsmodels.org/stable/index.html.
[24] Marco Cuturi and Mathieu Blondel. Soft-DTW: a differentiable loss function for time-series. InProceedings of the 34th International Conference on Machine Learn-ing - Volume 70, ICML’17, pages 894–903, Sydney, NSW, Australia, August 2017.
JMLR.org. ZSCC: 0000153.
Figure A1.1.A part of the test period of SARIMAX 1.
Figure A1.2.A part of the test period of ReduNet 6.
(continues)
Figure A1.3.A part of the test period of SARIMAX 1.
Figure A1.4.A part of the test period of ReduNet 6.
(continues)
Figure A1.5.A part of the test period of SARIMAX 1.
Figure A1.6.A part of the test period of ReduNet 6.
Table A2.1.aFRR is automatic frequency restoration reserve, mFRR is manual frequency restora-tion reserve, FCR-N is frequency controlled normal reserve and FCR-D is frequency controlled disturbance reserve.
ID Description Unit Source
1 Volume of upregulation capacity, aFRR MW Fingrid Oy
2 Volume of downregulation capacity, aFRR MW Fingrid Oy
33 Ordered downregulation, mFRR MWh/h Fingrid Oy
34 Ordered upregulation, mFRR MWh/h Fingrid Oy
51 Price of upregulation capacity, aFRR e/MW Fingrid Oy 52 Price of upregulation capacity, aFRR e/MW Fingrid Oy 53 Activated downregulated power, aFRR MWh/h Fingrid Oy
54 Activated upregulated power, aFRR MWh/h Fingrid Oy
79 Price of activated FCR-N e/MW Fingrid Oy
80 Volume of activated FCR-N MW Fingrid Oy
81 Price of activated FCR-D e/MW Fingrid Oy
82 Volume of activated FCR-D MW Fingrid Oy
92 Imbalance power price e/MW Fingrid Oy
105 Sum of downregulation bids, mFRR MW Fingrid Oy
106 Downregulation price, mFRR e/MW Fingrid Oy
123 Activated FCR-N MWh/h Fingrid Oy
165 Next day’s consumption forecast in Finland MWh/h Fingrid Oy 166 Consumption forecast in Finland, updates hourly MWh/h Fingrid Oy
177 Frequency of the electricity grid Hz Fingrid Oy
178 Temperature in Helsinki ◦ C Fingrid Oy
181 Wind power production MW Fingrid Oy
182 Temperature in Jyväskylä ◦ C Fingrid Oy
183 Peak load power production MW Fingrid Oy
185 Temperature in Rovaniemi ◦ C Fingrid Oy
188 Nuclear power production MW Fingrid Oy
191 Hydropower production MW Fingrid Oy
192 Electricity production in Finland MW Fingrid Oy
193 Electricity consumption in Finland MW Fingrid Oy
194 Electricity’s net import/export MW Fingrid Oy
196 Temperature in Oulu ◦ C Fingrid Oy
198 Finland’s production surplus/deficit MW Fingrid Oy
201 District heat production MW Fingrid Oy
202 Industry’s power production MW Fingrid Oy
205 Other production MW Fingrid Oy
209 Power system state 1-5 Fingrid Oy
213 Downregulation, other power transactions MWh/h Fingrid Oy 214 Upregulation, other power transactions MWh/h Fingrid Oy 241 Production forecast in Finland, updates hourly MWh/h Fingrid Oy 242 Next day’s production forecast in Finland MWh/h Fingrid Oy
243 Sum of upregulation bids, mFRR MW Fingrid Oy
244 Upregulation price, mFRR e/MWh Fingrid Oy
245 Wind power production forecast, updates hourly MWh/h Fingrid Oy 248 Solar power production forecast, updates hourly MWh/h Fingrid Oy
DAY.AHEAD.PRICE Day-ahead price in Finland e/MW ENTSO-E