SVM Based Maximum Power Consumption Excess Forecast Alert for

Large-Scale Power Consumers

Seigo Haruta

1 a

, Ken-ichi Tokoro

2 b

and Takashi Onoda

1 c

Aoyama Gakuin University School of Science and Engineering, K anagawa, Japan

Central Research Institute of El ectric Power Industry, Kanagawa, Japan

Keywords:

Support Vector Machine (SVM), Machine Learning, Discrimination Problem, Excess Forecast Alert,

Large-Scale Power Consumers, Maximum Power Consumption, Imbalanced Data.

Abstract:

Large-scale power consumers, such as buildings and factories, make high-voltage power contracts with the

Japanese electric power companies. The basic fee for high-voltage power contracts is based on the maximum

power consumption in the past year. If the power consumption in the present month does not exceed the

maximum power consumption in the past year, large-scale power consumers can suppress the basic fee. So,

large-scale power consumers need the alert to prevent the maximum power consumption in the present month

from exceeding the maximum power consumption in the past year. In this study, excess forecasting was per-

formed considering the characteristics of power consumption in each industry. In addition, we proposed SVM

improvements for imbalanced data. We applied this method to power consumption data, which is imbalanced

data, to perform excess forecast. As a result, we have improved the accuracy of the excess forecast and con-

tributed to effective alerts to many large-scale power consumers.

1 INTRODUCTION

Consumers with large-scale electricity demand, such

as buildings, factories, and hospitals, are called large-

scale power consumers. In Japan, electric power

compan ie s sold 837.4 billion kWh of electricity in

2021. Of this, 552 billion kWh, or 62.3%, is used

by large-scale power consumers. Therefore, large-

scale power consumers are important custo mers for

electric power comp anies. In a ddition, electric ity lib-

eralization is progressing in Japan. Electric p ower

compan ie s must continue to provide services to large-

scale power consumers who use a lot of electricity.

Large-scale power consumers make the high-voltage

power contract with the Japanese electric power com-

panies. The power contrac t for general hou seholds

has a low basic fee a nd a high power usage fee that

is calculated in proportion to the amount of power

used. The high-voltage power contract has a low

power usage fee and a high basic fee. The basic

fee accounts for a large proportion of the elec tric-

ity charges of large-scale power consumers. The ba-

sic fee for high-voltage power contracts is based on

https://orcid.org/0000-0002-1311-7246

https://orcid.org/0000-0001-5074-8250

https://orcid.org/0000-0002-5432-0646

Figure 1: How the basic fee is determined for large-scale

power consumers.

the maximum power co nsumption in the past year.

Figure 1 shows how the basic fee for high voltage

power con tracts is determined. Large-scale power

consumers want to reduce th eir electricity charges. If

the power consumption in the presen t month does not

exceed the maximum power consumption in the past

year, large-scale p ower consumers can suppre ss the

basic fee(Tokoro et al., 2019). It is most important

for large-scale power consumers to prevent exceed-

ing the maximum power c onsumptio n for the past

year. Figure 2 shows the maximum power consump-

tion that determines the basic fee for August 2022.

As sh own in ﬁgure 3, if the maximum power con-

sumption is exceeded in August 2022, the maximum

power co nsumption that determines the b a sic fee in

806

Haruta, S., Tokoro, K. and Onoda, T.

SVM Based Maximum Power Consumption Excess Forecast Alert for Large-Scale Power Consumers.

DOI: 10.5220/0011666100003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 806-813

ISBN: 978-989-758-626-2; ISSN: 2184-4313

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

Figure 2: Maximum power consumption to determine basic

fee in August 2022.

Figure 3: Maximum power consumption to determine basic

fee in September 2022.

September 2022 changes. In order to solve such prob-

lems, power demand forecasting has been studied in

various ways from general households to large-scale

power consumers(Suganthi and Samuel, 2012)(Haghi

and Toole, 2013)(Motamedi et al., 2012).

2 RELATED RESEARCH

This chapter describes related r e search on fore-

cast alerts for exceeding maximu m power consump -

tion based on support vector machine (SVM)(Cortes

and Vapnik, 1995)(Burges et al., 1 999)(Vapnik,

1999b)(Vapnik, 1999a) to reduce basic fee for large-

scale power consumers(Tokoro et al., 2019).

2.1 Excess Forecast Alert

As one of the various forms of power demand fore-

casting being studied, an electric power company pro-

posed an excess forecast alert based on a discrimina-

tion problem as a service to be provided to large-scale

power consumers. Figur es 4 and 5 show the ﬂow of

excess forecast alerts. At for e cast execution time t,

Figure 4: How alerts are sent.

Figure 5: Large-scale power consumers curb power con-

sumption.

SVM is used to forecast whether the power consump-

tion at forecast target time t + 3 after 3 hours will

exceed the threshold. The threshold is 90% of the

maximum power consumption in the past year. If the

threshold is forecast to be exceeded, an excess fore-

cast alert is sent. Large-scale power consume rs who

receive alerts can reduce their power consumption so

that they do not exceed th eir m aximum power con-

sumption.

2.2 Results and Issues of Related

Research

Table 1 shows the prediction accuracy (recall, pre-

cision, F-measure)(Ahmad et al., 2018)(Powers,

2020)(Goutte and G a ussier, 2005)of an excess fore-

cast alert in the power consumption data of 5,727

large-scale power consumer s in Japa n. Th e issue of

related research is that the discriminan t function for

an excess forecast alert is created under the same con-

ditions for all indu stries, even though the amount of

electricity used differs depending on th e industry of

large-scale power consum ers. Large-scale power con-

sumers are diverse as shown in ﬁgure 6. Because th e

discriminant function is created without co nsidering

the ch aracteristics of the industry, the forecast accu-

racy of an excess forecast alert may be low.

SVM Based Maximum Power Consumption Excess Forecast Alert for Large-Scale Power Consumers

807

Table 1: Forecast accuracy for an excess forecast alert.

Recall Precision F-measure

62.17% 78.74% 68.28%

Figure 6: Industry of large-scale power consumers.

2.3 The Purpose of This Research

We perform excess forecasts based on the character-

istics of eac h industry to improve forecast accuracy.

We also improve the forecast accuracy by improving

the SVM.

3 PROPOSED METHOD

Power consump tion data for 5,727 large-scale power

consumers in Japan can be divided into nine indus-

tries. To improve the accuracy of excess forecast

alerts, we analyzed the characteristics of each indus-

try. We also improved the SVM for imbalanced data

to improve the accuracy of excess forecast alerts.

3.1 Characteristics of Power

Consumption in Each Industry

There is a difference in the power consumption be-

tween weekdays and holidays depending on indus-

try of the large-scale consumer. Th e ﬁgure 7 shows

a graph of c hanges in power consumptio n by whole-

salers and retailers in th e Ka nto region of Japan (Mon-

day, April 1, 2013 to Sunday, April 7, 2013). T he ﬁg-

ure 8 shows a graph of changes in power consumption

by th e ﬁnancial and insurance industries in the Kanto

region of Japan (Mon day, April 1, 2013 to Sunday,

April 7, 2013). The power consumption of whole-

salers and r etailers is constant regardless of the day of

the week.

Also, depending on the type of business of large-

scale p ower consumers, there are cases wher e there

are many excesses only in the summer, and there are

cases where th ere are many excesses in both summer

and winter. Figure 9 shows a case of high excesses

only in the summer. Figure 9 shows the number of

monthly excesses for all wholesaler and retailer cus-

tomers in the Kanto region of Japan. Figure 1 0 shows

a case of high excesses in both summer and winter.

Figure 7: Changes in power consumption by wholesalers

and retailers in the Kanto region of Japan (Monday, April 1,

2013 to Sunday, April 7, 2013).

Figure 8: Changes in power consumption by the ﬁnancial

and insurance industries in the Kanto region of Japan (Mon-

day, April 1, 2013 to Sunday, April 7, 2013).

Figure 10 shows the number of m onthly excesses for

all ﬁnancia l and insurance industry customers in the

Kanto region of Japan.

In the related research, the featur es used for the

discriminant function did not take into consideration

the industries of large-scale power consumers. In this

research, we select features that affect the excess of

the max imum power consumption for each industry

of large-scale power consumers. The method of se-

lecting features that affect exceeding the maximum

power consum ption is shown below.

1. Normalize all fe atures from minimum 0 to maxi-

mum 1

2. Find the median value of each feature when ex-

ceeding and when not exceeding

3. Select features with a media n difference of 0.2 or

Figure 11 is a boxplot of the feature quantity

”Power consumption of forecast execution time” at

Figure 9: The number of monthly excesses for all whole-

saler and retailer customers in the Kanto region of Japan.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

808

Figure 10: The number of monthly excesses for all ﬁnan-

cial and insurance industry customers in the Kanto region

of Japan.

Figure 11: The difference between the median values of the

feature for excess and non-exceeding.

the time of excess and non-exceeding in the Kanto

region wholesale and retail industry. As shown in ﬁg -

ure 11, when the difference between the median val-

ues of th e feature for excess and non-exceeding is 0.2

or more, 7 5% of the excess data and 75% of the non-

exceeding data have different values.

3.2 SVM Improvements for Imbalanced

Data

In related research, the ratio o f when consumers

power consumption exceeds the maximum power

consumption of the past year and whe n it does not is

imbalanced. In a discrimination problem using such

imbalanced data, th ere are c a ses where the model is

tilted by the majority class and the minority class is

neglected. Therefore, an excess forecast a ccuracy of

related research was low. Other related research uses

methods such as oversampling and undersampling as

solutions to the pro blem(Mohammed et al., 2020)(He

and Garcia, 2009)(Sun et al., 2009)(Krawczyk, 2016).

In this research, we propose a method to solve the

discrimination problem of imbalanced da ta by mov-

Figure 12: Discrimination problem of imbalanced data.

ing the op timal hyperplane of SVM by the ratio of the

number of training failures in each class. SVM cal-

culates discriminan t function f (x) from training data.

Let f (x) = 0 be the optimal hyperplane for discrim-

ination. However, in the discrimination problem of

imbalanced data as shown in ﬁgure 12, there are cases

where learning is not suc cessful. We will improve the

recall by moving the optimal hy perplane as shown

in ﬁgure 13. Let the optimal hyperplane after being

moved be the modiﬁed optimal hyperplane g(x) = 0.

Equation (1) shows the modiﬁed optimal hyperplane

g(x) = 0 using FP, FN and f (x) = 0 in the ﬁgure12.

By moving the optimal hyperplane, misidentiﬁcation

of class 2 incr eases, but misidentiﬁcation of class 1

decreases. Equation (2) shows the recall calculation.

Equation (3) shows the precision calculation.

g(x) = f (x) −

−1 × FP + 1 × FN

FP + FN

(1)

Recall =

T P

T P + FP

(2)

Precision =

T P

T P + FN

(3)

As shown in ﬁgure 13, movin g the optimal hyper-

plane decreases FP and increases TP. Instead, FN also

increases. Equation (4) shows the calculation of the

recall

′

after the improvement in ﬁgure 13, and equa-

tion ( 5) shows the calculation of the precision

′

after

the imp rovement in ﬁgure 13. Equation (4) shows that

the recall

′

increases. Equation (5) shows that TP in -

creases even if FN increases. Therefore, the precision

′

does not drop sign iﬁca ntly.

Recall

′

T P

′

T P

′

+ FP

′

T P + 4

T P + FP

(4)

Precision

′

T P

′

T P

′

+ FN

′

T P + 4

T P + 4 + FN + 1

(5)

SVM Based Maximum Power Consumption Excess Forecast Alert for Large-Scale Power Consumers

809

Figure 13: Move the optimal hyperplane.

Figure 14: How to move the optimal hyperplane.

Figure 14 shows how to move the optimal hyperplane.

Move the optimal hyperplane according to the propor-

tion o f the number of training failures within the mar-

gin. In the example in ﬁgure 14, class 1 has 4 learning

failure data and class 2 has 1 learning failur e data. In

that case, move the optim al hyperplane so that the ra-

tio within the margin is 4:1. However, this proposed

method assumes a ha rd-margin.

4 EXPERIMENTAL CONDITIONS

4.1 Data

The forecast accuracy was veriﬁed using the power

consumption data of 5,72 7 large-scale power con-

sumers nationwide published by the Japanese Sus-

tainable open I nnovation Initiative (SII) and the data

of th e Japan Meteorological Agency published on the

website of the Japan Meteorological Agency. As

shown in ﬁgure 10 , for an industry that has many

excesses in both summer and winter, it is assumed

that the features that affect excesses will change be-

tween summer (May to October) and winter (Novem-

ber to April). Discriminant functions for such indus-

tries were created by dividing the year into summer

(May to Octob e r) and winter (November to April).

4.2 Features

Feature candidates used for excess forecast alerts are

described below. In a ddition to the feature used in

the related research, we added “average temperature

for the week up to the forecast execution time t” and

“power consumption m hours before the forecast ex-

ecution time (m = 1, 2, 3)” was ad ded. In addition,

as shown in the ﬁgure 8, we added ”holiday informa-

tion” with 1 for Saturdays, Sundays, and holidays and

0 for wee kdays, taking into consideration industries

where there is a difference in power consumption be-

tween weekdays and holidays.

1. Power con sumption at the forecast execution time

2. Power con sumption at the forecast execution time

t on the previous day

3. Power consumption at the forecast target tim e t +

3 on the previous day

4. Temperature at the fo recast execution time t

5. Sun shine hours from time t − 1 to prediction exe-

cution time t

6. Cooling index at the fo recast execution time t

(How many degrees above 20 degrees Celsius)

7. Heating index at the forecast execution time t

(How many degrees below 18 degrees Celsius)

8. Average power consumption for the past m days

from the day before the forecast target time t + 3

(m = 7, 14, 21, 28)

9. Average power consumption for 4 weeks from the

previous week on the same day of the week at the

forecast target time t + 3

10. Average power consumption for x hours up to the

forecast execution time t (x = 2, 4, 6)

11. Difference between the amount of power con-

sumption at the forecast execution time t and

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

810

Table 2: Number of features used for each industry (Describe the number of features used fr om Section 2 of Chapter 4).

Industry Used featu res number

Wholesale industry 1,2,3,4,5,6, 7,8,9,14,15

Accommo dation and restaura nt industry 1,2,3,4, 5,6,7,8,9,14,15,16

Financial indu stry (sum mer) 1,2,3,4,5,6, 8,9,11, 14,15,16

Financial indu stry (winter) 3,4,7,8,12,14,16

Manufacturin g industry ( summer) 1,2,3,4,5,6, 8,9,11,14,15,16

Manufacturin g industry ( w inter) 1,2,3,4,5,7, 8,9,11,14,15,16

Medical welfare industry (summer) 1,2,3,4,5,6, 8,9,11,14,15,16

Medical welfare industry (winter) 1,2,3,4,5,7, 8,9,11,14,15,16

Life-related service industry 1,2,3,4,5,6, 7,8,9,14,15

Transportation and postal industry 1,2,3,4,5,6, 7,8,9,14,15,16

Education industry (summer) 1,2,3,4,5,6, 8,9,11,14,15,16

Education industry (winter) 1,2,3,4,5,7, 8,9,11,14,15,16

Other industry 1,2,3,4,5,6, 7,8,9,14,15,16

the average amount of power consumption for

the past m days from the previous day ( m =

7, 14, 21, 28)

12. Difference be tween power consum ption at fore-

cast execution time t and average power consump-

tion for 4 weeks from the previous week on the

same day of the wee k

13. Day of the week information (dummy variable for

each d a y of the wee k) and national holiday infor-

mation (dummy variable for whether it is a holi-

day)

14. Average tempera ture for the we ek up to the fore-

cast target time t

15. Power co nsumption m hours before the prediction

execution time (m = 1, 2, 3)

16. Holiday inform a tion

Based on Chapter 3, the characteristics of each in-

dustry were extracted from the above characteristics.

The fe a tures used in each industry ar e shown in table

2. For the industrie s divided into summer and winter,

the feature is shown for each.

4.3 An Excess Alert Based on SVM

Suppor t vector machine (SVM) is used to create a dis-

criminant function that forecasts whether th e power

consumption after 3 h ours will exceed the threshold

based on the input data. Soft-margin SVM, which

can handle overlapping class distributions, was used

to create discriminant fu nctions. Moreover, RBF (Ra-

dial Basis Function Kernel) was used as the SVM ker-

nel. Th e cost parameter C ranges from 100 to 1000,

and the RBF kernel parameter γ ranges from 0.01

to 100. G rid search (Syarif et al., 2016) is used to

ﬁnd the most appropriate hyperp arameters. Further-

more, a discriminant f unction was created f or each

Table 3: The ratio to move the optimal hyperplane, which

is determined for each industry.

Industry Ratio

Wholesale industry 1.00:2 .23

Accommo dation and restaurant 1.00:3 .00

Financial industry 1.00:4 .56

Manufacturin g industry 1.00:3 .47

Medical welfare industry 1.00:2 .77

Life-related service industry 1.00:3 .17

Transportation and postal industry 1.00:3.08

Education industry 1.00:3 .88

Other industry 1.00:3 .88

consumer for each hour. However, in our experiments

with the proposed improved SVM method for imbal-

anced data, we used a hard-margin SVM. This is be-

cause the improvement proposal m ethod assume s a

hard-margin. In addition, we determined how best to

move the hyperplane for each industry when using the

improved SVM for imbalanced data. Table 3 shows

the ratio for eac h industry.

5 EXPERIMENTAL RESULTS

Section 1 of Chap te r 5 presents the resu lts of the pro-

posed method in Section 1 of Chapter 3 . Section 2 of

Chapter 5 presents the resu lts of the proposed method

in Section 2 of Chapter 3.

5.1 Excess Forecast Alert Considering

Industry Characteristics

Table 4 shows the accuracy results of the excess fore-

cast alerts in this research, in which discriminant

functions were created based o n the characteristics

SVM Based Maximum Power Consumption Excess Forecast Alert for Large-Scale Power Consumers

811

Table 4: An excess alert results by industry using the proposed method (S ection 1 of Chapter 3).

Industry Recall Precision F-measure

Wholesale industry 73.44% 83.86% 77.87%

Accommo dation and restaura nt industry 66.04% 80.95% 72.29%

Financial indu stry 57.78% 76.61% 65.27%

Manufacturin g industry 59.61% 77.31% 66.55%

Medical welfare industry 67.98% 82.22% 74.01%

Life-related service industry 63.96% 80.19% 70.34%

Transportation and postal industry 64.65% 81.06% 71.17%

Education industry 61.77% 79.11% 68.90%

Other industry 63.49% 81.21% 70.43%

Average of all consumers 67.36% 81.33% 73.11%

Table 5: Comparison of the results of related research and this research using the proposed method (Section 1 of Chapter 3).

Recall Precision F-measure

Related research 62.17% 78.74% 68.28%

Proposed metho d(Section 1 of Chapter 3)(Soft-margin) 67.36% 81.33% 73.11%

of each industry of consumers. Even w hen discrim-

inant functions are used for summe r and winter re -

spectively, the for e cast acc uracy is summed for ea ch

industry. Table 5 shows a comparison of accuracy re-

sults for excess forecast alerts in related research and

excess fore cast alert accuracy results in this study,

in which a discrim inant function was created based

on the cha racteristics of each industry of consume rs.

We were able to improve the F-measure by about 4.8

points.

Since the amount of power consumption by each

consumer industry has its own cha racteristics, we

changed the conditions for creating the d iscr iminant

function for each consum er industry to improve the

accuracy of the excess foreca st alert. As a result of

creating a discriminant function according to the char-

acteristics of each consumer industry, we were able

to improve the foreca st accuracy (re call, precision, F-

measure) of the average excess forecast alert f or all

consumers. As a result, w e c an send effective alerts to

consumers in various industries. However, the recall

is 67.36%, which is low compared to the precision of

81.33%. Lower recall means more oversights. We

want to aim for an excess forecast a lert w ith a high

F-measure and recall.

5.2 Excess Forecast Alerts Based on

Improved SVM for Imbalance Data

Table 6 shows the accuracy results of the excess fore-

cast alert in this study, which created the d iscriminant

function based on the improved SVM for the imbal-

anced data . Table 7 compares the average results for

all con sumers in Table 6 and the accuracy r e sults for

the excess forecast alerts in this study, in which dis-

criminant functions were created based on the charac-

teristics of each industry of consumers using the pro-

posed method(Section 1 of Chapter 3). We were able

to improve the recall by about 6.8 points and the F-

measure by about 1.3 points.

The excess f orecast alert result based on improved

SVM for imbalanced data could obtain high recall

while improving F-measure. A high rec a ll means a

reduction in the number of missed oversights.

6 CONCLUSION

In this research, since the amount of electricity used

is characteristic for each indu stry of consumers, we

changed the conditions for creating the d iscr iminant

function for each industry of c onsumers to improve

the accuracy of excess forecast alerts. As a result of

creating a discr iminant f unction based on the char a c-

teristics of each consumer’s industry, it was possible

to improve the recall, precision, and F-measure of the

average excess forecast alert for all consumers. In ad-

dition, excess forecast alerts using the improved SVM

for imbalanced data were able to produce results with

high F- measures and r e call. This makes it possible

to contribute to the transmission o f effective alerts

to con sumers in a wide variety o f industries. How-

ever, in order to have a wide range of consumers use

it in the future, further improvement in a ccuracy is re-

quired. The accuracy of excess forecast alerts will be

improved in our future researc h.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

812

Table 6: An excess alert results by industry using the proposed method(Section 2 of Chapter 3).

Industry Recall Precision F-measure

Wholesale industry 75.64% 74.55% 74.94%

Accommo dation and restaura nt industry 72.11% 66.87% 69.28%

Financial indu stry 68.50% 61.66% 64.66%

Manufacturin g industry 67.02% 64.84% 65.71%

Medical welfare industry 72.54% 70.27% 71.27%

Life-related service industry 68.33% 68.65% 68.21%

Transportation and postal industry 68.06% 69.46% 68.49%

Education industry 72.34% 62.24% 66.79%

Other industry 71.44% 66.26% 68.61%

Average of all consumers 72.33% 69.55% 70.72%

Table 7: Comparison of the results of this research using the proposed method (Section 1 of Chapter 3) and this research using

the proposed method (Section 2 of Chapter 3).

Recall Precision F-measure

Proposed metho d (Section 1 of Chapter 3) (Hard-margin) 65.41% 74.75% 69.41%

Proposed metho d (Section 2 of Chapter 3) (Hard-margin) 72.33% 69.55% 70.72%

REFERENCES

Ahmad, M., Aftab, S., Bashir, M. S., Hameed, N., Ali, I.,

and Nawaz, Z. (2018). Svm optimization for sen-

timent analysis. International Journal of Advanced

Computer Science and Applications, 9(4).

Burges, C. J., Smola, A. J., and Scholkopf, B. (1999). Ad-

vances in kernel methods. Support Vector Learning,

53.

Cortes, C. and Vapnik, V. (1995). Support-vector networks.

Machine learning, 20(3):273–297.

Goutte, C. and Gaussier, E. (2005). A probabilistic interpre-

tation of precision, recall and f-score, with implication

for eval uation. In European conference on informa-

tion ret rieval, pages 345–359. Springer.

Haghi, A. and Toole, O. (2013). The use of smart meter data

to forecast electricity demand. CS229 course paper.

He, H. and Garcia, E. A. (2009). Learning from imbalanced

data. IEEE Transactions on knowledge and data engi-

neering, 21(9):1263–1284.

Krawczyk, B. (2016). Learning from imbalanced data: open

challenges and future directions. Progress in Arti ﬁcial

Intelligence, 5(4):221–232.

Mohammed, R., Rawashdeh, J., and Abdullah, M. (2020).

Machine learning with oversampling and undersam-

pling techniques: overview study and experimental

results. In 2020 11th international conference on in-

formation and communication systems (ICICS), pages

243–248. IEEE.

Motamedi, A., Zareipour, H., and Rosehart, W. D. (2012).

Electricity price and demand forecasting in smart

grids. IEEE Transactions on Smart Grid, 3(2):664–

674.

Powers, D. M. (2020). Evaluation: from precision, recall

and f-measure to roc, informedness, markedness and

correlation. arXiv preprint arXiv:2010.16061.

Suganthi, L. and S amuel, A. A. (2012). Energy models for

demand forecasting - a review. Renew able and sus-

tainable energy reviews, 16(2):1223–1240.

Sun, Y., Wong, A. K., and Kamel, M. S. (2009). Classiﬁca-

tion of imbalanced data: A review. I nternational jour-

nal of pattern recognition and artiﬁcial intelligence,

23(04):687–719.

Syarif, I., Prugel-Bennett, A., and Wills, G. (2016). Svm

parameter optimization using grid search and ge-

netic algorithm to improve cl assiﬁcation performance.

TELKOMNIKA (Telecommunication Computing Elec-

tronics and Control), 14(4):1502–1509.

Tokoro, K., Tsurumi, T., and Hi go, T. (2019). Excessive

power consumption alert using svm (in japanese). In

The Institute of Electrical Engineers of Japan Na-

tional conference.

Vapnik, V. (1999a). The nature of st ati stical learning the-

ory. Springer science & business media.

Vapnik, V. N. (1999b). An overview of statistical learn-

ing theory. IEEE transactions on neural networks,

10(5):988–999.

SVM Based Maximum Power Consumption Excess Forecast Alert for Large-Scale Power Consumers

813