ASHRAE Journal - February 2021 - 43

TECHNICAL FEATURE

can be used to test the algorithm proTABLE 1 Detection results of three datasets.
posed in this study.
DATA SET 3
DATA SET 1
DATA SET 2
For the above three sets of data, the
k-Nearest
k-Nearest
k-Nearest
Z-Score
Z-Score
Z-Score
Neighbor
Neighbor
Neighbor
parameter selection method proTP
163
115
155
90
265
158
posed in the " Calculation Method
of N and k " section above is used to
TN
15,532
15,120
15,434
15,511
15,562
15,048
select N and k, and the abnormal data
P
171
115
168
90
282
158
detection algorithm is used to detect
Q
15,609
15,665
15,821
15,899
15,616
15,740
the abnormal data. For the ﬁrst set,
TP + TN
15,695
15,235
15,589
15,601
15,827
15,206
N = 50, k = 30. For the second set, N =
P+Q
15,780
15,780
15,989
15,989
15,898
15,898
60, k = 40. And for the third set, N =
ACC
99.5%
96.5%
97.5%
97.6%
99.6%
95.6%
50, k = 30. The data points of instantaneous ﬂow rate, supply and return water temperacollected from the heat meter installed in a thermal
ture on the primary side and the abnormal data detec- station located in Hejin City, Shanxi Province, China,
tion results were shown in Figures 6 through 8.
between Nov. 15, 2018, and March 15, 2019. The difThe accuracy, ACC, is used to evaluate the performance ference between two adjacent data is used to calculate
of the k-nearest neighbor method. ACC is calculated
the heat consumption data per minute. After that, the
using the following equation:
Z-score method is used to detect the outlier data. Then,
the method proposed in this article is used after impleTP + TN
(8)
menting the Z-score method to study the improvement
ACC =
P+Q
of detection performance.
where Q is the total number of data points detected
Figure 9a shows that before abnormal detection, there
as normal; TN, true negative, is the number of normal
are a small number of outliers in the data set because
data points correctly detected as normal. The detection
of the malfunction of wireless transmission of data.
results are shown in Table 1.
Using the Z-score method, you can see from Figure 9b that
Table 1 shows that for Data Set 1 to Data Set 3, the accumany outliers with very large deviations from normal
racy of the k-nearest neighbor method is 99.5%, 97.5%
values can be successfully removed, but outliers with
and 99.6%, respectively. For Data Set 1 and Data Set 3,
relatively small deviations from normal values cannot be
the accuracy of the k-nearest neighbor method is 3% and removed. This is because the abnormal data with a large
4% higher than that of the Z-score method, respectively.
deviation from the normal value causes a large deviation
Z-Score is a commonly used standard-normal-distribuin the calculation of the mean and variance when impletion-based abnormal data detection method. Although
menting the Z-score method. Figure 9c shows that outliers
all the abnormalities detected by the Z-score method are with small deviations from the normal values can also be
correct, the Z-score method cannot completely judge the detected successfully.
abnormal data (especially when the abnormal data is
Because the k-nearest neighbor method does not
close to the normal data), while for the k-nearest neighrequire calculating the mean and variance based on
bor method, after the calculation of k and N values, the
the original data that may include erroneous data
k-nearest neighbor method can detect the abnormal
and judging which one is abnormal using the erronedata accurately. At the same time, the correct data at the ous data, the performance resulting from detecting
edge of the interval can avoid being detected as abnorabnormal values with a small deviation from normal
mal data, which proves the feasibility of the abnormal
values will not be affected by abnormal data with a
detection algorithm proposed in this article.
large deviation. Therefore, using the algorithm inteThis method has already been applied to the real-life
grated in the smart heating system, the abnormal data
application of the abnormal detection of sensor data
can be automatically detected and rejected, enabling
collected from several smart heating networks in three
the analysis of energy consumption and building
China provinces, including Shanxi, Shandong and
thermal characteristic identiﬁcation to be carried out
Henan. Figure 9 presents the accumulated heat data
successfully.
FEBRUARY 2021

ashrae.org

ASHRAE JOURNAL

http://ashrae.org

ASHRAE Journal - February 2021

Table of Contents for the Digital Edition of ASHRAE Journal - February 2021