Adaptive Segmentation and Machine Learning Based

Potential DR Capacity Analysis

Wen-Jun Tang Yi-Syuan Wu Hong-Tzer Yang
[email protected] [email protected] [email protected]
Research Center for Energy Technology and Strategy,
Department of Electrical Engineering
National Cheng Kung University
Tainan, Taiwan

Abstract—By shedding their electricity consumption or getting introducing models of DR aggregator, more studies are shifted
ready to be shed, the Demand Response (DR) program to analyze the consumption behavior among low voltage users.
compensates the shortage of generation or acts as spinning Based on the data collected by low voltage advanced metering
reserve. Depending on how the end-users join in the program. the infrastructure (LVAMI), consuming behavior clustering [5],
potential capacity of DR is, therefore, a key issue no matter to consumption prediction [6], direct load control, and price-based
system operator or DR aggregator. The proposed method load control [7]-[11] are proposed. A. Asadinejad et al. [11]
employs adaptive k-means approach to evaluate the potential consider the probability of changing the consumption behavior
consumer as candidate participants of DR. The consumption to schedule the appliance usage. All these mentioned studies
prediction models of controllable appliances are constructed by
assume the appliances controllable, perfectly obeying the
Gaussian Processes for Machine Learning (GPML). Through
combining the candidates’ data and prediction models, the
schedule, and their historical consumptions well known.
potential capacity is then achieved. Case studies evaluate the However, except some demo sides, to most of consumers with
accuracy and efficacy of the proposed method with practical low just AMI installed, the consumption data of each appliance
voltage advanced metering infrastructure (LVAMI) data cannot be obtained. The capacity of individual consumer’s DR
achieved from Taiwan. The results show good efficiency and potential is, therefore, difficult to be investigated.
practicability of the proposed method. There are some studies deliberating the classification of
AMI data for even further analysis. K-means, hierarchical
Index Terms-- Demand Response, LVAMI, Adaptive K-means,
Machine Learning
clustering, and finite mixture modeling are traditional
clustering algorithms. References [12] and [13] compare these
I.   INTRODUCTION algorithms and divide them as direct clustering and indirect
clustering categories. A suitable clustering method groups
Owing to growth of energy demand and environmental similar data within same bunch, and also has the ability to
protection issues, distributed renewable energy sources (RESs) distinguish dissimilar data. Some indices can be used to
have been introduced to worldwide distribution systems. The evaluate the clustering quality [12]. K-means is often used for
penetration of DGs, such as cogeneration, photovoltaic, gas AMI data clustering [6], [14]. Reference [6] uses the clustering
turbines, wind power, and fuel cells, is meanwhile fast scheme to find similarities between customers’ behaviors to
increasing. Adequate controllable and fast-reacting resources improve the accuracy of load forecast. Y. Zhang et al. propose
are required to maintain system’s security and reliability [1]. a method to calculate the baseline for DR based on the
energy storage system (ESS) and demand response (DR) are clustering result.
thus attracting public’s attention as the compensation for
traditional generation units. However, the number of groups k must be determined in
advance, and how to find a suitable k is an issue. To solve this
With the popularity of advanced metering infrastructure problem, adaptive k-means and Bayesian information criterion
(AMI), the real-time power consumption is fully aware by users are employed by References [15] and [16]. These two methods
and enable them to participate in active bidding programs and determine the number k according to error-tolerant rate and
managements [2]. Many countries have been pushing DR threshold [15] or Bayes Information Criterion [16]. The
program, while some of them can provide 5%-10% of total strategy employed by Reference [15] evaluates the potential a
consumption as DR capability. The advantages of DR include consumer has to participate in DR or energy efficiency (EE)
peak shaving, RES efficiency enhancement, short action time, program based on adaptive k-means algorithm.
and low cost.
This paper proposes a method to deal with large quantities
From previous researches, the analysis of DR is mainly of AMI data for the total consumption. The potential capacity
focused on medium or high-voltage level consumers with high to join DR is obtained with potential segmentation, prediction
quantity and stable usage of consumption [3],[4]. By

This work was supported by the Ministry of Science and Technology,

Taiwan, under Grant106-3113-E-006-010 & 104-2221-E-006 -116 -MY3.

Authorized licensed use limited to: IEEE Xplore. Downloaded on July 28,2020 at 03:43:50 UTC from IEEE Xplore. Restrictions apply.
 Consumer with DR program
The potential duration (peak
LVAMI Data Adaptive K-means stable consumption
users’ data time) data
behavior
Final Prediction
Segmentation&Screening Daytype Result

Controllable
Potential Weather Data
Appliances’ Data
Capacity
Sensitivity Analysis

Open Data Weather Data Analysis &

Prediction Model
Construction
Daytype
Machine
Learning

Figure 1.   The overall flowchart of proposed method

models construction and capacity estimation. The details are dividing the maximum consumption L of the day. Let s (t ) be
described as follows. the normalized load profile and i(s) a tentative representative
II.   PROPOSED STRATEGY load pattern. Euclidean distance E(s, i(s)) is computed in (2) to
assign the element to each cluster. E(s, i(s)) is minimized to
The research proposes a method to evaluate the users with determine k representative patterns i*(s), which is the
DR participating potential and the related capacity for DR. The representative load profile determined from (3) for minimal
capacity is calculated based on some kinds of appliances. The square error (MSE) and called centroid to signify the s sample
categories of appliances are determined by 1) the fixed and curves of data [5].
unadjustable load, 2) controllable load, and 3) adjustable load.
The second and third kinds of loads may contribute to DR By defining a threshold θ in (4), the number of i*(s) is
program. Limited by diverse facts, such as environment, updated. θ is the threshold which is normally limited within
location, weather, season, family structure and so on, prediction 0-2 [2],[15]. A near-to-2 value of θ denotes a loose
of some appliance consumption is complex and difficult classification environment with relatively less dictionaries. On
especially due to lack of historical consumption data. the contrary, a smaller θ leads to a stricter segmentation.
In the view of data mining, to evaluate the potential of DR Compared with the traditional k-means algorithm, the adaptive
capacity, the consumers with DR-participating potential are one generates optimal ko within a given boundary (kmin to kmax)
firstly screened. In other words, the consumer with stable and determine it referring to the MSE within the limitation by a
consumption behavior is preferred as a candidate to be further certain proportion θ of centroid value’s summation.
evaluated. The load data evaluation of candidate’s second and
The algorithm proceeds with the steps described in Figure
third categories of appliances (as defined previously) is then
2. It follows a simple principle that if a clustering process results
made. The quantities of energy that can be shed is computed
in violation of condition in (4), an additional 2-means algorithm
through the prediction models of individual appliances. The
is used to partition the violating segmentation and two new
overall flowchart of the proposed strategy is described as in
centroids are obtained. The updated result is then checked again
Figure 1 with the details given in the sections that follow.
with the re-defined boundary. By introducing equation (4), the
A.   Potential Segmentation classification error tolerance becomes stricter, while the
K-means algorithm is traditionally used for classification. segmentation number is adaptable dynamically.
The value k is also be treated as the dictionary of s data and L = max .(l (1),..., l (t ),..., l (24)),
must be determined prior to the start of the algorithm. While
l (t ) (1)
dealing with large quantity of data, how to determine a suitable s(t ) =
k is a debatable issue. The target of the research is to find the L
consumer with stable consumption behavior through 24
classifying different patterns of power-consumption behavior. E ( s, i) = ∑t =1 ( s(t ) − i( s))2 (2)
That is, the value k should be different as for users with different
consumption data and behavior.
i* (s) = arg min E (s, i) (3)
Adaptive k-means is employed to solve the problems i
mentioned above. It starts with an initial kmin and forms the
seeds of centroid of clusters. Normalization of the data is very 24
important to avoid being dominated by one or few properties. E (s, i* (s)) ≤ θ ∑t =1 i * (s) 2 (4)
Equation (1) is used to normalize the daily profile l(t) by

Authorized licensed use limited to: IEEE Xplore. Downloaded on July 28,2020 at 03:43:50 UTC from IEEE Xplore. Restrictions apply.
 Start

HVAC Consumption
Set k= kmin

Traditional k-means
algorithm
Temperature Dew Point Total Consumption

Calculate the number Nv of

clusters that violate the threshold Figure 3.   The relationship between HVAC prediction and predictors

Training
Yes Data
Nv=0?
Temperat
No ure
Temperat

Data of the time to be

ure
k=k+Nv End

Historical Data
Well-Known
Dew

predicted
Point Prediction
GPML Model of Dew Point
Yes Total HVAC
k>kmax? Consump
Total
tion
No Consump
tion
For the clusters violating the Consump HVAC
tion of consumpt
threshold limitation do 2-means HVAC ion
clustering
Figure 4.   GPML prediction model for HVAC construction
Update the cluster and
check the violation

Figure 2.   The flowchart of adaptive k-means algorithm

The features involved in the prediction modeling consist of
time, date, date type, weather summation, ozone density,
When evaluating the users’ consumption, average load temperature, dew point, humidity, visibility, total consumption
and peak consumption are normally used as indices. However, of a consumer, etc. Take as an example the heating, ventilation
these two indices are not enough to denote the features of and air conditioning (HVAC) consumption prediction model.
consumption behavior, especially the variability. Entropy of The screening result shows three features, temperature, dew
each consumer introduces the probability of load profiles point, and total consumption, are most related to the air
owned to each centers p(i*(s)). Therefore, the entropy is consumption and meanwhile have less inter-correlation with
computed according to each consumer’s clustering results to each other.
estimate the stability or variability of their consumption Figure 3 shows the relationship between these three features
behavior as shown in Equation (5), where k is the total number and conditioner consumption. All the data displayed are
of clustering. By employing entropy, some extreme but not normalized. The correlation between HVAC consumption and
frequent consumption behavior can be ignored. total consumption shows almost positively linear. It
k demonstrations the relationship between the power
Sn = −∑ p(i* ( s)) log p(i* ( s)) consumption and temperature/dew point above some constant
(5)
i =1 extent. In a word, the pre-processing of input variables
increases the efficiency and reduce the redundancy of data to
B.   Prediction Model of Appliance the following model construction.
1)   Sensitivity Factors Selection 2)   Consumption Prediction Models of Appliances
To build the prediction model by machine learning (ML), Supervised ML can be divided as the one of regression with
input data should just include the features related to the continuous value output and the other of classification with
predicting values and expected output. Selection of the features discrete value output. Gaussian Processes for Machine
is thus important. An algorithm is used to evaluate the worthy Learning (GPML), as one of regression solution, is good at
subset of feature data. The algorithm determines the solving high dimension, varying samples, and non-linear
redundancy degree of each individual feature for the prediction complicated problems. It utilizes the view of function-space by
[17]. Preferred is the subset which is highly correlated with the changing self-variable space to high-dimension space based on
classified training data within the subset but is less inter- kernel function. Kernel function (6) is a binary function and
correlated with the other subsets. defined as the inner product of two vectors. The function can
be treated as the similarity of two vectors. The Gaussian
Processes (GP) can be understood as a process where feature
vector is changed to kernel vector. During the processing, every
input vector is combined with a stochastic vector which follows

Authorized licensed use limited to: IEEE Xplore. Downloaded on July 28,2020 at 03:43:50 UTC from IEEE Xplore. Restrictions apply.
Gaussian Distribution (GD). Meanwhile, any combination B.   Potencial Capacity Reuslt
ofthe input vectors also follows GD. GP turns to Gaussian In the case study, the load data are segmented by adaptive
random field, when the vector is multi-dimensional. Similar to k-means method. The parameter of threshold θ is determined
GD, GP is also described as a distribution function.
as 2 according to [15], while kmin and kmax are set to be 5 and 105,
respectively. Sample consumers with stable consumption
k ( xm , xn ) = φ ( xm )T φ ( xn ) (6) behavior and with nonstable behavior are shown by load profile
number and entropy in Table I. Obviously, the stable ones have
Where xm is the prediction input, while xn is the exiting less segmentations of daily profiles. However, the number of
training data , such like total consumption,  temperature and dew segmentation is not the index that can represent the variability
point are the input of HVAC model. It is noted that all the input completely. Entropy (function (5)) is calculated according to
attributes’ categories and format should be the same. With the the probability of consumption behavior being divided to each
result obtained from Kernel function (6), the similarity between load pattern. As presented in Table I, entropy varies even with
prediction data and training data is calculated. The predicted the same load pattern number. The table lists part of the
output ym is thus computed with GP distribution function consumers from Taiwanese LVAMI dataset. Some of them
according to the relationship between ym and training data have larger load pattern number. However, the user M00249
output yn, HVAC’s historical consumption for HVAC model with 49 load patterns (meaning 49 consumption behavior types)
for instance. With GP and the features selected by previous has higher entropy than the one M00052 with 64 load patterns,
sensitivity screening, every appliance has its individual a fact which means M00249 actually has more nonstable
forecasting model. The same, HVAC model is taken as an consumption behavior than M00052. Overall, the candidates
example shown in Figure 4. The historical training data as with DR potential are selected with the entropy value lower
shown are associated with the HVAC power consumption at the than 25% of the highest entropy. 77.4% of the consumers are
same time interval. The more training data we have, the more thus concluded as the candidates.
accurate prediction model can be achieved. It should be pointed The users from two locations, M and T, are employed as
out that the input of prediction model for future time points examples. With the previous assortment, the numbers of the
must have the same categories of appliance, same units, and candidates are 213 and 496, separately, accounting for 89.5%
same time interval as the training data used to construct the and 84.9% of M and T consumer groups. Figure 6 is the
model. summation of predicted HVAC consumption of candidates
III.   CASE STUDY AND DISCUSSION TABLE I.   SOME PRESENTATIVE USERS’ LOAD PATTERN NUMBER AND
ENTROPY
The proposed method is applied to the LVAMI data
LOAD LOAD
provided by Industrial Technology Research Institute of CUSTOMER CUSTOMER
PATTERN ENTROPY PATTERN ENTROPY
Taiwan. The data contain 9,588 residential AMIs’ data for ID ID
NUMBER NUMBER
every 15 minutes in year 2016. For demonstration of training
M00012 50 5.6739 M00001 5 0.4712
the models, the biggest database for family energy consumption,
M00050 55 5.2924 M00002 5 0.3708
Pecan Street, is utilized [18]. The data of Pecan Street are
M00052 64 3.5292 M00003 5 0.2913
collected from communities of the cities in US. The database
M00082 53 4.3213 M00004 5 0.3189
provides the consumption data of different appliances per 15
minutes. All the data and codes are implemented in Matlab M00126 62 3.3015 M00005 5 0.2908

2017 for the model construction and potential DR capacity M00221 53 3.4485 M00006 5 0.2727

prediction. M00243 49 5.5204 M00007 5 0.2442

M00249 50 4.8069 M00010 5 0.3119
A.   Accuracy of Prediction Model M00266 70 3.4235 M00011 5 0.3119
As mentioned above, Pecan Street data are employed to M00306 67 3.9120 M00013 5 0.4420
build and test the accuracy of the models. Eighty percent (80%)
kW
consumers are randomly chosen as training data, and the rest
2.5 3 kW
Prediction Real Prediction Real
2.5

20% as the test data. Figure 5 shows four accuracy evaluation

2

results. Figure 5(a) and (b) represent two randomly-chosen

Air Consumption

1.5
Air Consumption

1.5

consumers’ actual and predicted HVAC consumptions at the

1
1

same day. Compared to the actual data, most trends and 0.5
0.5

consumptions are accurately predicted, except some 0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
h
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
h
consumption changing points, such as the one during hour 7-8 -0.5
Time
(a)
-0.5
Time
(b)
kW
of the user in (a). Similar results are also displayed in Figure 3 kW
Prediction Real
3

Prediction Real

5(c) and (d). Tested on the data for yearly appliance 2.5 2.5
Air Consumption

consumption of 283 users, the average prediction Mean 2

Air Consumption

1.5

absolute percentage error (MAPE, calculated by the equation

1.5

1
1

(7)) reaches 10.4%. 0.5

0.5

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

100 n ActualValu e − Pr edictionValue 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Time h -­‐0.5
Time
h

MAPE =
(d)
∑ (7) (c)

n t =1 ActualValu e Figure 5.   Prediction accuracy evaluation of HVAC model

Authorized licensed use limited to: IEEE Xplore. Downloaded on July 28,2020 at 03:43:50 UTC from IEEE Xplore. Restrictions apply.
 45 kWh

40
total consumption predicted air condition consumption
200
kWh
180
Total Consumption Predicted Air Condition Consumption
candidates and prediction models, the potential capacity can
35

30
160

140
thus be calculated. Realistic Taiwanese LVAMI data has been
120
tested to evaluate the applicability of proposed strategies. The
Air Consumption

Air Consumption
25
100
20

15
80

60
proposed scheme is believed to be a promising tool for
10

5
40 investigating the potential DR participants and their related
capacity for system operator or DR aggregators.
20
0 0
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 1 4 7 10 13 1619 22 25 28 31 34 3740 43 46 49 52 55 5861 64 67 70 73 76 7982 85 88 91 94
-5 Time per 15 mins per 15 mins
Time
(a) (b)

Figure 6.   DR potential capacity provided by HVAC from two areas (a) M

REFERENCES
area; (b) T area. [1]   M. H. Albadi and E. F. El-Saadany, “Demand response in electricity
markets: an overview,” in Proc. IEEE Power Eng. Soc. Gen. Meeting, pp.
1–5.
[2]   A. J. Conejo, J. M. Morales, and J. M. Morales, “Real-time demand
response model,” IEEE Trans. Smart Grid, vol. 1, no. 3, pp. 236–242, Dec.
2010
[3]   G. Chicco, R. Napoli, and F. Piglione, “Comparisons among clustering
techniques for electricity customer classification,” IEEE Trans. Power
Systems, vol. 21, no. 2, pp. 933–940, May 2006.
[4]   G. Chicco, “Overview and performance assessment of the clustering
methods for electrical load pattern grouping,” Energy, vol. 42, no. 1, pp.
68–80, June 2012.
[5]   C. Flath, D. Nicolay, T. Conte, C. van Dinther, and L. FilipovaNeumann,
“Cluster analysis of smart metering data—an implementation in practice,”
Business & Information Systems Engineering, vol. 4, no. 1, pp. 31–39, Jan.
2012.
[6]   F. L. Quilumba, L. Wei-Jen, H. Huang, D. Y. Wang, and R. L. Szabados,
Figure 7.   The shedding capacity provided by the aggreagator demonstration “Using smart meter data to improve the accuracy of intraday load
forecasting considering customer behavior similarities,” IEEE Trans.
Smart Grid, vol. 6, no. 2, Mar. 2015.
compared with the actual total energy consumption of all the [7]   M. Muratori and G. Rizzoni, “Residential demand response: dynamic
consumers in the location. Location M has relatively stable total energy management and time-varying electricity pricing,” IEEE Trans.
consumption, as compared to Locations T. On the contrast, Power Systems, vol. 31, no. 2, pp. 1108-1117, Mar. 2016.
HVAC consumption demonstrates an obvious dependency of [8]   D. Zhang, S. Li, M. Sun, and Z. O’Neill, “An optimal and learning-based
demand response and home energy management system,” IEEE Trans.
time, which shows higher consumption during noon and night, Smart Grid, vol. 7, no. 4, pp. 1790-1801, Jul. 2016.
while lower value during 3 a.m. to 11 a.m.. For the user group [9]   N. Forouzandehmehr, M. Esmalifalak, H. Mohsenian-Rad, and Z. Han,
T, the total consumption also represents peaks at noon and night, “Autonomous demand response using stochastic differential games,”
IEEE Trans. Smart Grid, vol. 6, no. 1, pp. 291-300, Jan. 2015.
but the overall variation is gentler. [10]  H. Yang, J. Zhang, J. Qiu, S. Zhang, M. Lai, and Z. Dong, “A practical
pricing approach to smart grid demand response based on load
An aggregator is assumed to allocate the potential capacity classification,” IEEE Trans. Smart Grid, no.99, pp.1-1, Mar. 2016.
of users in Locations M and T (709 consumers in total). All the [11]  A. Asadinejad, K. Tomsovic, and C. f. Chen, “Sensitivity of incentive
candidates in the two areas are assumed to join the DR project. based demand response program to residential customer elasticity,” in
Referring to the papers discussing the schedule of HVAC 2016 North American Power Symposium (NAPS), pp. 1-6.
[12]  Y. Wang, Q. Chen, C. Kang, M. Zhang, K. Wang, and Y. Zhao, “Load
[8][19][20], the average saving energy 14% of the total power profiling and its application to demand response: a review,” Tsinghua
consumption is used here for the shedding capacity evaluation. Science and Technology, vol. 20, pp. 117-129, Apr. 2015.
The results show that 1.4%-9% energy saving is obtained [13]  G. Chicco, R. Napoli, and F. Piglione, “Comparisons among clustering
techniques for electricity customer classification,” IEEE Trans. on Power
through the daytime (7:30-22:30) from HVAC. The Systems, vol. 21, no. 2, pp. 933–940, May 2006.
quantitative analysis during daytime can be found in Figure 7, [14]  Y. Zhang, W. Chen, R. Xu, and J. Black, “A cluster-based method for
which contains the total consumption of these areas, HVAC calculating baselines for residential loads,” IEEE Transaction on Smart
consumption prediction of candidates, and the DR potential Grid, vol. 7, no. 5, pp. 2368 – 2377, Sep. 2016.
[15]  J. Kwac, J. Flora, and R. Rajagopal, “Household energy consumption
capacity from HVAC of the aggregator. segmentation using hourly data,” IEEE Trans. Smart Grid, vol. 5, no. 1,
pp. 420 – 430, Jan. 2014.
IV.   CONCLUSION [16]  S. Haben, C. Singleton, and P. Grindrod, “Analysis and clustering of
This paper has presented the approach to estimate the DR residential customers energy behavioral demand using smart meter data,”
IEEE Trans. Smart Grid, vol. 7, no. 1, pp. 136 – 144, Jan. 2016.
potential capacity to contribute to power system reliability [17]  “The Code of Class weka.attributeSelection.CfsSubsetEval,”
especially during peak time. In the proposed approach, http://www.cs.tufts.edu/~ablumer/weka/doc/weka.attributeSelection.Cfs
Adaptive k-means method classifies consumption patterns SubsetEval.html.
[18]  “Dataport from Pecan Street”, https://dataport.cloud/, Austin.
referring to users’ different daily profiles. The pattern number [19]  J. Pepanen, M. J. Reno, and S. Grifalva, “Thermal Energy Storage for Air
and pattern size are processed via the entropy function to Conditioning as an Enabler of Residential Demand Response,” in 2014
evaluate the stability and complexity of consumption. The North American Power Symposium (NAPS), pp.1-6.
GPML reads the open data for each appliance from different [20]  Y. I. Khalid, N. U. Hassan, C. Yuen, and S. Huang, “Demand Response
Management For Power Throttling Air Conditioning Loads In Residential
users to build individual models corresponding to data for Smart Grids,” in 2014 IEEE International Conference on Smart Grid
weather, time, and date. Combining the stable consumer Communications, pp. 1-6.

 
 

Authorized licensed use limited to: IEEE Xplore. Downloaded on July 28,2020 at 03:43:50 UTC from IEEE Xplore. Restrictions apply.