International Core Journal of Engineering 2020-26 | Page 28

reach 95.38%. Calculate the BPA output of PCA and ICA
statistics, and partial result is shown in Table II.
effectively, and multiple statistics can be used as evidence to
identify abnormal data from various aspects. The paper
considers the multi-evidence fusion method of evidence
conflict to eliminate the pseudo data in the process of
abnormal data identification, and weights the remaining
evidence, which makes the evidence fusion more accurate
and further improves the abnormal data. The accuracy of
recognition also reduces the rate of misrecognition.
The existing evidence is processed according to the steps
shown in Fig. 1 to determine whether there is pseudo
evidence. If the evidence fusion of a variable is used, it is
judged that there is evidence to support the primary focus. If
the degree is less than the average support of the primary
coke, it is considered as pseudo-evidence. After elimination,
the remaining evidence is weighted and averaged, and the
Pignistic probability method is used to make the decision.
C. Results and Analysis of Risk Processing in Power market
Settlement Data
The results of the multi-evidence fusion decision of
some abnormal data in the first random setting abnormal
data experiment are shown in Table V.
TABLE II. BPA O UTPUT FOR S OME V ARIABLES
Number PCA  T 2 PCA  SPE ICA  I 2 ICA  SPE
1 0.020028 0.030032 0.022099 0.032177
2 0.026934 0.023239 0.031077 0.021452
3 0.032459 0.026814 0.031768 0.024312
4 0.01174 0.034322 0.025552 0.031105
5 0.022099 0.023597 0.029696 0.022167
6 0.037293 0.027172 0.032459 0.023954
7 0.009669 0.035038 0.012431 0.034322
8 0.024171 0.023239 0.031077 0.026457
9 0.03453 0.028602 0.037293 0.033607
10 0.009669 0.033607 0.020718 0.031105
11 0.036602 0.023239 0.036602 0.025027
12 0.03453澳 0.030032 0.040746 0.03182
TABLE V. M ULTI - EVIDENCE F USION D ECISION R ESULT OF S OME
S ETTLEMENT A BNORMAL D ATA
Number
X  1,28 
X  1,29 
X  1,30 
X  16,28 
X  16,29 
PCA
ICA
SPE I 2 SPE Multiple
evidence fusion
1 27.8% 47.2% 58.3% 30.7% 94.4%
2 33.3% 36.1% 30.6% 47.2% 86.1%
3 41.7% 27.8% 38.9% 63.9% 91.7%
4 38.9% 30.6% 19.4% 27.8% 88.9%
5 27.8% 44.4% 47.2% 58.3% 83.3%
PCA
ICA
SPE I SPE Multiple
evidence fusion
1 2.78% 2.19% 4.68% 5.26% 1.17%
2 1.9% 2.05% 3.22% 1.75% 0.58%
3 3.8% 2.34% 2.19% 4.39% 0.88%
4 3.95% 4.53% 3.51% 3.65% 1.75%
5 1.9% 4.24% 1.75% 4.82% 0.73%
T
2
2
178000 kWh 78000 kWh 0.2641
164980 kWh 64980 kWh 0.2225
0.2¥/kWh 0.229¥/kWh 0.7792
80880 kWh 64129 kWh 0.5497
0 64129 kWh 0.0119
TABLE VI. A BNORMAL S ETTLEMENT D ATA R ECONSTRUCTION R ESULT
TABLE IV. A BNORMAL D ATA A LARM R ATE OF D IFFERENT R ECOGNITION
M ETHODS
Number
Multiple evidence
fusion results
0.1017
The paper sets the reconstruction threshold T 0.3, then
the data that needs to be reconstructed is X(1,28), X(1,29),
X(1,30) and X(16,30). The relevant data is selected for
polynomial fitting, and the maximum curvature of the fitted
curve is calculated. Because curvature is related to the
magnitude of the numerical value, the data needs to be
normalized. The paper sets the curvature threshold Q 0.2.
The degree of deviation between the reconstructed data of
the two methods and the original normal data is calculated,
the results are shown in Table VI.
TABLE III. A BNORMAL D ATA R ECOGNITION R ATE OF D IFFERENT
R ECOGNITION M ETHODS
T 2 Correct data
value
0.246¥/kWh
X  16,30 
B. Risk Identification Results and Analysis of Power market
Settlement Data
The paper re-randomly sets the anomaly data to the
original data samples to judge the accuracy of the evidence
theory's recognition of data risks. The pairs of identification
methods are shown in Table III and Table IV.
Number
Abnormal
data value
0.12¥/kWh Number Fit
curve
curvature Fit
result Fit
offset
rate Historical
Data
calculation
results Historical
Data
calculation
result
offset rate
X  1,28  0.1607 0.226 8.13% 0.212 16.04%
X  1,29  0.2663 98416 26.17% 87360 10.71%
X  1,30  0.3318 117134 80.26% 85570 24.06%
X  16,30  0.4191 50632 21.05% 57830 10.89%
It can be found that the accuracy of the polynomial
fitting results tends to be higher when the curvature of the
fitted curve is low. However, due to the large fluctuations of
data X(1,29), X(1,30) and X(16,30), the polynomial fitting
results can not accurately reflect the normal data values,
especially the polynomial fitting offset rate of X(1,30)
reaches 80.26%, which is unacceptable. Through the
combination of polynomial fitting and historical correlation,
the final reconstructed data offset rates are 8.13%, 10.71%,
24.06% and 10.89%, respectively, within a reasonable range.
The calculation results show that according to the method
proposed in the paper, the advantages of the two
reconstruction methods can be combined. By discriminating
the curvature of the fitted curve, a more appropriate data
reconstruction method is selected as the alternative.
Through the experimental results, it can be found that
when a single piece of evidence is used to identify abnormal
data, the effective recognition rate is lower. After apply
multiple evidence fusion, the data recognition rate is greatly
improved, and can reach more than 80%. Through multiple
evidence fusion, PCA and ICA advantages can be combined
6