Professional Documents
Culture Documents
Machine Learning in Power Markets: Bilal Asghar Farooqi Dr. Ali Abbas Kazmi Abdul Kashif Janjua
Machine Learning in Power Markets: Bilal Asghar Farooqi Dr. Ali Abbas Kazmi Abdul Kashif Janjua
Abstract—Intelligent operational planning ensures supply and office on no-profit, no-loss and electric cooperatives which are
demand matching in a power system, which is achieved governed by members responsible for investing and policy
traditionally by optimization and scheduling of government and making [3].Furthermore, concept of power market is widely
private owned non-renewable power plants. Recently, adopted throughout the world-which includes a system
encouragement has been extended to local and distributed power
generators, due to uncertainty and long-term variability in
operator, buying electricity from a facility on a negotiated
renewable rich system. In the study, machine learning approaches price and sells electricity to the customer with profit, serving
are proposed in the context of power markets to learn and predict as an escrow responsible for management of generation,
usage patterns to avoid power deficit. A case study is presented, transmission , distribution and billing [4].
which includes a solar plant and a wind farm for base load along
with bio-gas plant for peak load. To predict peak load, logistic Traditionally electric market is large, monopolistic, vertical
regression, a supervised machine learning approach, has been and state owned responsible for generation, transmission and
employed to classify the time of engagement, in order to ensure distribution as shown in Figure-1 [5].
supply and demand balance. Applying logistic regression will result
in reduced operational and economic cost for utility and price for
consumer.
I. INTRODUCTION
Electricity being the efficient form of secondary energy
storage is considered a commodity and regarded as a major
source of modern survival. The demand of electricity is
increasing due to increase proliferation of smart devices,
electronic appliances, population and economic growth;
however, achievement of supply-demand balance is still a
critical task. Additionally, sufficient growth of electricity
production could not be witnessed to meet the demand that is
why consumers face load shedding of 8-12 hours a day in
developing countries [1]. To mitigate supply demand
imbalance, modern world is shifting from traditional energy
resources to renewable energy resources. The Danish
Government’s plan to create green growth by shifting power Fig. 1 Traditional Power Market
generation from 100 percent renewable energy resources by To increase compliance with the proliferation of renewable
2050 [2]. This transformation inherently brings benefits and and distributed generation, increase in competition within the
challenges of complex nature. However, evolution of markets is established by deregulation as shown in Figure-2.
restructured power markets around the world making local This is achieved by splitting up vertically integrated power
generation feasible and viable. Developed countries are producers and privatization of state-owned utilities. To
heavily investing in renewable infrastructure development as enhance performance of aforementioned entities, deregulation
well as electric utilities. is introduced to break natural monopoly in order to establish
competition within the market. However, consumer-end
Electric utilities include investor own faculties governed by deregulation is achieved by introducing competition to reduce
Power Purchase Agreement between investor and controller, power cost for customers.
public power system or municipalities governed by public
Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:57:11 UTC from IEEE Xplore. Restrictions apply.
Supply-demand balance is achieved by establishing wholesale Artificial intelligence (AI) and machine learning are poised to
and retail markets where customer get the least price and revolutionize the way utilities produce, transmit and consume
supplier get maximum profit as shown in Figure-3 [6]. energy by powering the modern smart grid.AI can help cutting
Moreover, markets use either day ahead or intraday trading consumer bills and managing power generation. Global
and price clearing mechanisms, which can be achieved via demand of low-carbon, green electricity with low cost is
increasing and AI applications are increasingly being used to
meet this demand with potentially huge long-term impact [11].
Machine learning is a field of artificial intelligence that uses
statistical techniques to give computer systems the ability to
learn from data without being explicitly programmed [12].
Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:57:11 UTC from IEEE Xplore. Restrictions apply.
Different electricity sources and different control strategies are theory. [41] gives a comparison of various time-series
the difference between microgrid and conventional grid in modeling approaches for forecasting of spot electricity prices.
power market. However, the main distinction among them is [42] compares the linear and non-linear time-series models for
different participants and different purpose of introducing forecasting of electricity prices for California electricity
electricity markets [20]. Energy trading market is flexible as market. [43] describes approaches to model the electricity
the change of energy price in real time pricing environment is prices.
due to changing operational parameters. Commonly used In literature, focus is on forecasting and pricing but
energy pricing schemes are Real Time Pricing (RTP), Time- classification of scenarios for predictive analysis is not widely
Of-Use (TOU) pricing, Peak Rebate Pricing (PRP), Critical discussed. Thus, logistic regression, a classification-based
Peak Pricing (CPP), and Day ahead (DA) pricing. The machine learning algorithm will forecast the need for a new
ultimate goal of power markets is to achieve supply-demand generation by predicting increase in demand in a day ahead
balance as well as maximization of benefits for all. Therefore, market [44]. It will help utilities predicting the demand from
machine learning approaches are studied to forecast the consumers in days, weeks, and months to ensure availability
predicted load and consumer behavior to pre-plan the of sufficient generation from power resources. Hence, no need
generation (renewable or non-renewables) based on the price, of demand side management or transactive energy where
location, efficiency, need and stability of the power system consumers were engaged to vary their consumption pattern
[21]. [45].
Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:57:11 UTC from IEEE Xplore. Restrictions apply.
In logistic regression, sigmoid function is taken as hypothesis
and Gradient Decent minimizes the cost function to attain
acceptable decision boundary as shown in Figure-6.
Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:57:11 UTC from IEEE Xplore. Restrictions apply.
Step-5: Set threshold for measuring algorithm REFERENCES
Step-6: If 1-P is greater than threshold, '‘Engage New [1] G. D. Valasai, M. A. Uqaili, H. R. Memon, S. R. Samoo, N. H. Mirjat,
Facility’’ otherwise move to next iteration and K. Harijan,” Overcoming electricity crisis in Pakistan: A review of
sustainable electricity options,” Renewable and Sustainable Energy
Reviews, vol. 72, pp. 734-745, 2017.
Proposed algorithm will predict possible need of engaging a
[2] http://www.buildup.eu/en/news/securing-denmarks-energy-future
new facility depending on the threshold set by the operator.
[3] https://www.coursera.org/lecture/electric-utilities/2-3-utility-types-and-
This will open up new dimensions to operator’s management prices-overview-public-power-and-electric-cooperatives-WV3GN
strategy which might result in incentives for customers to [4] http://www2.econ.iastate.edu/tesfatsi/ElectricPowerMarketDesign.TESH
reduce demand to save cost. However, if the predicted demand andbookChapter.LTesfatsion.pdf
is still high, smart contracts might be drafted to save time, [5] Nguyen, Hieu Trung; Battula, Swathi; Takkala, Rohit Reddy; Wang,
improve efficiency and increase output with maximum Zhaoyu; and Tesfatsion, Leigh, "Transactive Energy Design for
Integrated Transmission and Distribution Systems" (2018). Economics
utilization of all resources. Working Papers: Department of Economics, Iowa State University.
18004.
V. CONCLUSIONS & FUTURE WORK
[6] Mayer, K., & Trück, S. (2018). Electricity markets around the
Precision is acknowledged as the most efficient performance world. Journal of Commodity Markets, 9, 77-100.
measurement matrix for regression analysis-based [7] Dynamic Pricing for Decentralized Energy Trading in Micro-grids. /
classification. All these metrics may be obtained from Liu, Youbo; Zuo, Kunyu; Liu, Xueqin; Liu, Junyong ; Kennedy, Jason
.In: Applied Energy, Vol. 228, 06.07.2018, p. 689.
confusion matrix. Precision is represented as the number of
[8] http://www.europarl.europa.eu/RegData/etudes/BRIE/2016/593519/EPR
correct examples divided by the number of all the examples S_BRI(2016)593519_EN.pdf
labeled by the classifier. Recall is the number of examples [9] L. Xue, Y. Teng, Z. Zhang, J. Li, K. Wang and Q. Huang, "Blockchain
correctly classified divided by the number of all the examples technology for electricity market in microgrid," 2017 2nd International
in the data [47]. We checked the correctness of the proposed Conference on Power and Renewable Energy (ICPRE), Chengdu, 2017,
pp. 704-708.
algorithm by taking hourly data from CES, NUST from 8am
[10] Chen S, Liu C-C (2017) From demand response to transactive energy:
to 6pm of 5 working days for 2 consecutive weeks of state of the art. J Mod Power Syst Clean Energy 5(1):10–19
July,2018 (100 Hours). Precision turned out to be 60% and [11] https://www.raconteur.net/technology/giving-power-to-the-people-with-
recall 40 % subsequently. Often, there is an inverse ai-tech
relationship between precision and recall, where it is possible [12] https://www.expertsystem.com/machine-learning-definition/
to increase one at the cost of reducing the other. This case [13] http://www.greatachievements.org/
study provides an illustrative example of the tradeoff. [14] M. A. Badri, A. Al-Mutawa, D. Davis, and D. Davis,” EDSSF: A
Consider an operator tasked with performing load shedding. decision support system (DSS) for electricity peak-load forecasting,”
The operator may be more liberal to perform load shedding. Energy, vol. 22, no. 6, pp. 579-589, 1997.
This decision increases recall but reduces precision. [15] H. Farahbakhsh, V. Ugursal, and A. Fung,” A residential end use energy
consumption model for Canada,” International Journal of Energy
Conversely, the operator may be more conservative to perform Research, vol. 22, no. 13, pp. 1133-1143, 1998.
load shedding. This decision increases precision but reduces [16] M. Alizadeh, L.Xiao ,W.Zhifang ,A.Scaglione ,and R.Melton, “Demand-
recall. Hence, greater recall increases the chances of load Side anagement in the Smart Grid :Information Processing for the
shedding (negative outcome) and increases the chances of PowerSwitch,” IEEE Signal Processing Magazine, vol.29,no.5,pp.55–
uninterrupted power (positive outcome). Conclusively, Greater 67,Sep.2012.
precision decreases the chances of load shedding which leads [17] T.H. MouftahandM.E.Kantarci,“Wireless Sensor Networks for Cost-
Efficient Residential Energy Management in the Smart Grid,” IEEE
to customer satisfaction and improved quality of life. Transactions on Smart Grid, vol.2,no.2,pp.314–325,Jun.2011.
Classification performed via logistic regression will [18] Farrokh A Rahimi and Ali. 2012. Transactive energy techniques:
help in pre-planning and scheduling of new facility closing the gap between wholesale and retail markets. The Electricity
engagement beforehand. In this way, operational cost would Journal 25, 8 (2012),29–35.
be reduced and incentive schemes might be offered to [19] Xuesong Zhou, Tie Guo, Youjie Ma, “An overview on microgrid
technology”, 2015 IEEE International Conference on Mechatronics and
customers to meet supply-demand balance by using demand Automation (ICMA), IEEE Conference Publications, 2015, pp.
side management or via implementation of a novel demand 76 – 81, doi: 10.1109/ICM A.2015.7237460
framework. To predict expected increase in demand resulting [20] L. Xue, Y. Teng, Z. Zhang, J. Li, K. Wang and Q. Huang, "Blockchain
in increased customer satisfaction proposed algorithm can be technology for electricity market in microgrid," 2017 2nd International
tested on expanded data set. Our case study is limited to a Conference on Power and Renewable Energy (ICPRE), Chengdu, 2017,
specific site; however, it provides a proof of concept. To pp. 704-708.
increase efficiency and performance, machine learning [21] Muhammad Asghar Khan,CIIT/FA12-REE-030/ISB,MS Thesis In
Electrical Engineering COMSATS Institute of Information Technology
approaches can be utilized at every level of electric power Islamabad – Pakistan Spring, 2014.
market to establish standardization. From the introduction of [22] Suganthi, L., Samuel, A.A. Energy models for demand forecasting—A
competition within the decentralized market, players will be review. Renewable and Sustainable Energy Reviews 16: 1223 -1240
encouraged to test all machine learning algorithms on their (2002).
data set to find out which learning algorithm gives optimal [23] Amjady, N. Short-term hourly load forecasting using time-series
modeling with peak load estimation capability. Power Systems, IEEE
results for performance maximization. Transactions on 16:498–505 (2001).
Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:57:11 UTC from IEEE Xplore. Restrictions apply.
[24] Charytoniuk, W., Chen, M.S., Van Olinda, P. Nonparametric regression [36] Qiuge Liu, Qing He, and Zhongzhi Shi. Extreme support vector
based short-term load forecasting. IEEE Transactions on Power Systems machine classifier. PAKDD, pages 222{233, 2008.
13:725–730 (1998). [37] J. A. K. Suykens and J. Vandewalle. Training multilayer perceptron
[25] Xia, C., Wang, J., McMenemy, K. Short, medium- and long-term load classifiers based on a modified support vector method. IEEE
forecasting model and virtual load forecaster based on radial basis transactions on Neural Networks, 10:907{911, 1999.
function neural networks. International Journal of Electrical Power & [38] Guang-Bin Huang, Xiaojian Ding, and Hongming Zhou. Optimization
Energy Systems 32:743 –750 (2010). method based extreme learning machine for classification. Neurocom-
[26] Elattar, E.E., Goulermas, J., Wu, Q.H. Electric Load Forecasting Based puting, 74, December 2010.
on Locally Weighted Support Vector Regression. Systems, Man, and [39] Guang-Bin Huang, Xiaojian Ding, Hongming Zhou, and Rui Zhang.
Cybernetics, Part C: Applications and Reviews, IEEE Transactions on Extreme learning machine for regression and multiclass classification.
40:438–447 (2010). IEEE transactions on systems, man and cybernetics, 42(2), April 2012.
[27] Liao, S.-H. Expert system methodologies and applications—a decade [40] Marcelo Espinoza, J. A. K. Suykens, Ronnie Belmans, and Bart
review from 1995 to 2004. Expert Systems with Applications 28:93 – Demoor. Electric load forecasting. IEEE Control Systems
103 (2005). Magazine,27:43{57, 2007.
[28] Amjady, N., Keynia, F. A New Neural Network Approach to Short [41] Rafal Weron and Adam Misiorek. Forecasting spot electricity prices: A
Term Load Forecasting of Electrical Power Systems. Energies 4:488– comparison of parametric and semiparametric time series models.
503 (2011). MPRA, (10428), 2008.
[29] Vapnik, V.N. The nature of statistical learning theory. Springer-Verlag [42] Adam Misiorek, Stefan Trueck, and Rafal Weron. Point and interval
New York, Inc., New York,NY, USA (1995). forecasting of spot electricity prices: Linear vs. non-linear time series
[30] Jakkula, V. Tutorial on Support Vector Machine (SVM) (2006). models. Studies in Nonlinear Dynamics and Econometrics, 10, 2006.
[31] Haykin, S. Neural Networks: A Comprehensive Foundation, Prentice [43] Dr. Aurelio Fetz. Fundamental aspects of power markets - price
Hall International Editions Series. Prentice Hall (1999). forecasting. Part of Strommarkt-II lectures at ETH Zurich, Spring 2012,
[32] Sra, S., Nowozin, S., Wright, S.J. Optimization for Machine Learning. 29 February 2012.
MIT Press (2011). [44] https://www.coursera.org/learn/machine-learning/supplement/EYX8
[33] J. A. K. Suykens, T. Van Gestel, J. de Brabanter, B. De Moor, and J. /lecture-slides
Vandewalle. Least Square Support Vector Machines. World Scientific [45] Chen S, Liu C-C (2017) From demand response to transactive energy:
Publishing Co. Pte. Ltd., 2002. state of the art. J Mod Power Syst Clean Energy 5(1):10–19.
[34] Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme [46] https://onlinecourses.science.psu.edu/stat504/node/150/
learning machine: Theory and applications. Neurocomputing, [47] Tomin N.V., Kurbatsky V.G., Sidorov D.N., Zhukov A.V.IFAC-
70(13):489 501, 2006. PapersOnLine. Vol.49. No.27. P.445-450.
[35] Benoit Frenay and Michel Verleysen. Using SVMS with randomized
feature spaces: an extreme learning approach. In ESANN 2010
proceedings, European Symposium on Artificial Neural Networks-
Computational Intelligence and Machine Learning., April 2010.
Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:57:11 UTC from IEEE Xplore. Restrictions apply.