A power data evaluation method based on IGEP-CNN algorithm

Abstract

The present invention relates to the technical field of power risk assessment, and in particular to a power data assessment method based on the IGEP-CNN algorithm, including S1, performing harmonic, voltage deviation and frequency deviation, three-phase imbalance, voltage fluctuation power quality analysis on the collected power data; S2, using the LOFR-NDCV algorithm to filter out abnormal collected data and repair the group data, and identify abnormal data in the power data; S3, improving the classic GEP algorithm; S4, performing power data risk assessment and prediction based on the IGEP-CNN algorithm. The present invention optimizes the algorithm from two aspects: dynamic generation of initialization parameters of the GEP algorithm and replacement of inferior individuals in the evolution process; constructs a neural network prediction model based on the CNN algorithm, and uses the improved GEP algorithm to optimize the initialization parameters of the CNN neural network model to improve the accuracy and generalization ability of the CNN prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of power risk assessment, and in particular to a power data assessment method based on an IGEP-CNN algorithm. Background Art

[0002] Generally, the monitoring data collected from various links of the power system are complex and not suitable for direct analysis; the data monitored normally consists of effective components and error components; the error part can be eliminated or not. For the error data that can be eliminated, it follows the probability distribution principle, and this part is also called white noise. For the error part that cannot be eliminated, it deviates from the distribution trajectory of the original data and does not follow any distribution law. This part is also called outlier data; these error data are generated during the measurement process and data transmission, and the causes are also diverse, but basically caused by factors such as measurement asynchrony, measurement equipment failure, and unstable measurement environment. Therefore, the data collected from the power system must be effectively classified, analyzed and processed, the white noise part must be effectively eliminated, the outlier data must be fully eliminated, and the abnormal data must be reasonably identified.

[0003] In addition, data prediction technology has been continuously exerting its social application value since its inception. Applying data prediction technology to the identification and detection of abnormal data in power grids will inevitably enhance the identification of power grid noise data. At the same time, with the development of artificial intelligence technology, technical limitations have been continuously broken through in the prediction direction, because the quality of the method directly affects the accuracy of the prediction. At the same time, artificial intelligence prediction methods are superior to traditional methods.

[0004] Choi et al. proposed a GAN-based anomaly detection and location method for multivariate time series data of power plants, as well as a transformation method for time series imaging, called distance image; anomalies are detected and located by executing residual images and anomaly scoring functions; finally, based on the actual power plant data set, the effectiveness of anomaly detection of the proposed method on the data set was verified through experiments. Chahla et al. proposed a power data anomaly detection method based on deep learning methods, combining clustering-based methods with prediction-based methods to learn typical behavior scenarios and predict future power data values. This new method can specify the time when anomalies occur; the effectiveness of the proposed method was verified experimentally under the simulation of the data set.

[0005] Ren et al. proposed a method for quantitative assessment of power grid risk; through experimental simulation of real data sets, the effectiveness and superiority of the proposed method were verified. Huang et al. proposed a method for information security risk assessment of power systems based on genetic algorithms. Combined with the application cases of genetic algorithms, they analyzed their application in power system information security risk assessment scenarios, established corresponding model methods, established a security risk quantification model, and modeled the algorithm according to the empirical parameter settings of the genetic algorithm. The algorithm was compared with several existing algorithms through real data sets. The experimental results verified the feasibility of the proposed method in information security risk application scenarios.

[0006] The above-mentioned methods provide a lot of references for power data anomaly processing, but most of them have poor generalization ability and are only suitable for specific data application scenarios. At the same time, most data anomaly mining methods only consider fixed-point mining without multi-dimensional correlation analysis. Therefore, there are still many areas worthy of in-depth analysis in the direction of power data anomaly mining.

[0007] The above-mentioned artificial intelligence risk assessment method cannot characterize the functional model between input and output, cannot characterize the internal correlation of data, and cannot further explore the problem of breaking through the problem that neural network algorithms are prone to fall into local optimality. Therefore, how to establish a large number of combined prediction methods that show superiority in prediction accuracy has emerged. At the same time, how to combine the optimal models is also a direction that needs to be continuously explored. Summary of the invention

[0008] In order to solve the above technical problems, the present invention proposes a power data evaluation method based on the IGEP-CNN algorithm, comprising:

[0009] S1. Analyze the power quality of the collected power data for harmonics, voltage deviation, frequency deviation, three-phase imbalance, and voltage fluctuation;

[0010] Harmonics refer to voltages and currents that are no longer in the form of sine waves due to nonlinear components, where h is the harmonic order. The calculation of harmonics is shown in the following formula:

[0011]

[0012] Among them, w1 represents the angular frequency, c0 represents the DC component; c h , are the amplitude and initial phase of the hth harmonic respectively;

[0013] The voltage deviation ΔU represents the voltage value U at the measuring point r With nominal voltage U N The ratio of the deviation to the nominal voltage is calculated as follows:

[0014]

[0015] Among them, ΔU is the voltage deviation, U r is the voltage value at the measuring point, U N is the nominal voltage;

[0016] The frequency deviation Δf also represents the frequency value f of the measurement point r With nominal frequency f N Deviation:

[0017] Δf=f r -f N (3)

[0018] Where Δf is the frequency deviation, f r is the frequency value of the measurement point, f N is the nominal frequency;

[0019] Voltage three-phase unbalanceε u 、Current three-phase unbalance ε i It refers to a system that does not meet the requirements of equal power, frequency, and phase difference. The calculation method of the three-phase imbalance is shown in the following formula:

[0020]

[0021] Among them, ε u , ε i are the three-phase unbalance of voltage and current, u1, i1 and u2, i2 represent the RMS values ​​of the positive-sequence voltage and negative-sequence current components, respectively;

[0022] Voltage fluctuation refers to the rapid change of effective voltage value. Voltage flicker is the result caused by voltage fluctuation. The measurement formula of voltage fluctuation is the equation U of two adjacent extreme values ​​in a series of voltage effective value changes. MAX with U MIN The difference between the nominal voltage U N The calculation formula is as follows:

[0023]

[0024] Among them, d is the voltage fluctuation, and the two adjacent extreme values ​​U in the voltage effective value change MAX with U MIN , U N is the nominal voltage;

[0025] S2, using LOFR-NDCV algorithm, filter out abnormal collection data and repair Liqun data, and identify abnormal data in power data;

[0026] Since the power data collection process is easily affected by external factors such as unstable detection environment, the collected power data will randomly distribute a certain number of spike data, that is, outlier data. The local anomaly factor algorithm is used to identify the outlier data. Since outlier data has no analytical significance, it is usually directly eliminated. However, this method will destroy the time series integrity of the collected data and is not conducive to overall analysis.

[0027] The LOFR algorithm deals with the process of outlier data identification and outlier data repair. First, the definition of outlier data is given;

[0028] S21. Suppose there is an m×n dimensional electric energy data set.

[0029] For each data point o in the m×n dimensional electric energy data set ij The local outlier factor is: lof(o ij ), define the power data o ij The outlier factor lof(o ij ):

[0030]

[0031] Among them, ij is a data point in the data set, d k It means calculating the distance between the point and the rest of the points and sorting them from small to large. The kth value is d k , P i represents all points within the k-distance domain, that is, points that are density-reachable;

[0032] Therefore, for any power data o ij ,i∈{1,2,...,m},j∈{1,2,...,n}, if lof(o ij )<1, it is normal data. If lof(o ij )>1, then the electric energy data is called ij For outlier data, the median filtering algorithm is used to replace it. The replacement formula is as follows:

[0033]

[0034] where x max 、x min The maximum and minimum values ​​of the data in the two points before and after the data point, including this point;

[0035] The time complexity is O(mn). Under normal circumstances, m≤n can satisfy the characteristics of the m×n-dimensional power data set that conforms to the normal distribution, so O(n) is ignored. Therefore, the time complexity of the outlier data identification and repair algorithm based on LOFR is O(m).

[0036] S22, abnormal data identification based on NDCV algorithm to solve the misjudgment of normal data;

[0037] Assume there is an m×n dimensional electric energy data set Among them, each power data o ij The mean and variance of the row and column are μ i , μ j , And, for any power data o ij , where i∈{1,2,...,m},j∈{1,2,...,n}, if and If the electric energy data is established, then the electric energy data o ij is the abnormal electric energy value;

[0038] The LOFR-NDCV algorithm first uses the LOFR algorithm to find the spike data that affects the identification of abnormal data in the power data set, and then uses the median filter algorithm to repair the identified spike data to ensure the integrity of the time series data. Finally, the rationality of the normal distribution of each data is verified based on the repaired data until all abnormal data are identified;

[0039] S3, improve the classic GEP algorithm, first initialize the population size 、population head and population number Parameters, respectively, represent the number of populations, the number of monomer genes, and the number of front ends of monomer genes;

[0040] population size The calculation method of is shown in formula (8) and (9):

[0041]

[0042] Among them, ROUND() is the rounding function; n is the iterative evolution parameter of the algorithm; f i is the fitness size; C (i，j) is the model prediction value; T j is the true value;

[0043] population number Parameters are dynamically generated as shown in formula (10):

[0044]

[0045] For population head Parameters are dynamically generated as shown in formula (11):

[0046]

[0047] From mathematical expressions 8-11, it can be seen that the initialization parameter formula generated by dynamic parameters fully considers the fitness value of the contemporary group after operation, and according to this value, the size of the initialization parameter is appropriately adjusted, making the algorithm more flexible;

[0048] Secondly, since the probability value of the genetic operation of the algorithm is fixed and changes during the iteration process, the design is extremely unreasonable. According to the properties of the random search in the early stage and the stable search in the later stage during the population evolution process, the present invention performs the recombination rate P of the three genetic operators on the basis of the random search in the early stage and the stable search in the later stage. (i,r) , mutation rate P (i,m) , transposition rate P (i,ro) Set the adaptive evolution probability value. The specific setting method is as follows:

[0049] For the parameter P (i,r) , the dynamic generation method is shown in formula (12):

[0050]

[0051] For the parameter P (i,m) , the dynamic generation method is shown in formula (13):

[0052]

[0053] For the parameter P (i,ro) , the dynamic generation method is shown in formula (14):

[0054]

[0055] The dynamic generation formula of genetic operators integrates the factors of the algorithm evolution environment in which the individual is located. The probability value changes from large to small. The value is large in the early stage, which can improve the search speed of the algorithm. The value is small in the later stage to maintain the stability of the algorithm.

[0056] The entire evolution process of the classic GEP algorithm is carried out on the basis of the population formed at the beginning, which makes the diversity of the population worse and easily falls into the local optimum. In order to solve the local optimum problem, the improvement algorithm of inferior individuals in population (IA-IP) is designed. That is, in the iterative evolution process, the frequency of the maximum fitness value of the previous generation in the evolution process is detected. If it exceeds the threshold, the evolution algorithm of inferior individuals is used to replace several individuals with lower fitness values ​​in the previous generation population.

[0057] The basic steps of the IA-IP algorithm are as follows:

[0058] (1) Starting from the first generation of evolution, record the individual with the highest fitness in the population and detect the number of times this maximum fitness value appears in the subsequent evolution process;

[0059] (2) If the number of occurrences is within the specified threshold (δ (i) ), then record the updated fitness value and detect the number of times the updated value appears in the evolution process. The threshold is shown in formula (15):

[0060]

[0061] (3) If the number of occurrences exceeds the specified threshold, new individuals are introduced to replace the individuals with lower fitness in the generation that exceeds the specified threshold; the number of new individuals introduced is shown in formula (16):

[0062]

[0063] Where: Δ is the number of eliminated individuals, and also the number of new individuals added; P1, P2, P3, P4, P5, P6, P7 are the 3, 5, or 7 non-repeating frequencies with the smallest fitness values ​​in the contemporary evolving population, P size The number of groups.

[0064] (4) The number of individuals to be replaced is equal to the number of new individuals introduced; the individuals to be replaced are replaced using the roulette algorithm, and the individuals to be replaced are all the individuals ranked in the last 3, 5, or 7 positions from the largest to the smallest fitness value;

[0065] (5) After replacement, recalculate the fitness value of each individual in the population and enter process (1) again;

[0066] S4. Power data risk assessment and prediction based on IGEP-CNN algorithm;

[0067] In the direction of sequence prediction, the prediction work of the classic CNN algorithm is relatively excellent, but there is a common problem encountered by neural network algorithms, that is, the problem of gradient vanishing occurs during the training process, and the direction will be lost in the process of finding the optimal solution. Therefore, the result of sequence prediction using the classic CNN algorithm is unstable, sometimes high and sometimes low; therefore, the improved GEP algorithm is integrated with the CNN algorithm to make up for the defects of the CNN algorithm; because one of the main reasons for the gradient vanishing of the CNN algorithm is the unreasonable setting of network parameters, how to select a set of suitable network parameters for different data sets is a difficult problem, so the present invention uses the IGEP algorithm to automatically find a set of more suitable network parameters.

[0068] The normal iterative evolution judgment function of the GEP algorithm is the fitness function. Therefore, the fusion algorithm of the GEP algorithm and the CNN algorithm is to replace the fitness function with the neural network loss function to judge the quality of the current neural network parameters. The adaptive equation of the CNN algorithm applied to the IGEP algorithm is formula 17. At the same time, the expression of the population individuals of the GEP algorithm is combined with the neural network parameters. After the chromosome individual expression, it is a set of neural network parameters.

[0069]

[0070] Among them, f i is the total negative loss value, which is also the fitness value; loss (i，j) is the CNN prediction value; T j is the true value;

[0071] The beneficial effects of the present invention are:

[0072] In view of the shortcomings of the GEP algorithm, the algorithm is optimized from two aspects: dynamic generation of initialization parameters of the GEP algorithm and replacement of disadvantaged individuals during the evolution process, and the effectiveness of the improved method is verified experimentally; a neural network prediction model based on the CNN algorithm is constructed, and the improved GEP algorithm is used to optimize the initialization parameters of the CNN neural network model to improve the accuracy and generalization ability of the CNN prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flow chart of the power data evaluation method based on IGEP-CNN algorithm of the present invention;

[0074] Figure 2 is the experimental result of correlation analysis between real data and outlier data of the present invention;

[0075] Figure 3 This is a graph of the pre-training results of the CNN algorithm of the present invention. DETAILED DESCRIPTION

[0076] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0077] like Figure 1 As shown, a power data evaluation method based on the IGEP-CNN algorithm includes the following steps:

[0078] S1. Analyze the power quality of the collected power data for harmonics, voltage deviation, frequency deviation, three-phase imbalance, and voltage fluctuation;

[0079] Harmonics refer to voltages and currents that are no longer in the form of sine waves due to nonlinear components, where h is the harmonic order. The calculation of harmonics is shown in the following formula:

[0080]

[0081] Among them, w1 represents the angular frequency, c0 represents the DC component; c h , are the amplitude and initial phase of the hth harmonic respectively;

[0082] The voltage deviation ΔU represents the voltage value U at the measuring point r With nominal voltage U N The ratio of the deviation to the nominal voltage is calculated as follows:

[0083]

[0084] Among them, ΔU is the voltage deviation, U r is the voltage value at the measuring point, U N is the nominal voltage;

[0085] The frequency deviation Δf also represents the frequency value f of the measurement point r With nominal frequency f N Deviation:

[0086] Δf=f r -f N (3)

[0087] Voltage three-phase unbalanceε u 、Current three-phase unbalance ε i It refers to a system that does not meet the requirements of equal power, frequency, and phase difference. The calculation method of the three-phase imbalance is shown in the following formula:

[0088]

[0089] Among them, ε u , ε iare the three-phase unbalance of voltage and current, u1, i1 and u2, i2 represent the RMS values ​​of the positive-sequence voltage and negative-sequence current components, respectively;

[0090] Voltage fluctuation refers to the rapid change of effective voltage value. Voltage flicker is the result caused by voltage fluctuation. The measurement formula of voltage fluctuation is the equation U of two adjacent extreme values ​​in a series of voltage effective value changes. MAX with U MIN The difference between the nominal voltage U N The calculation formula is as follows:

[0091]

[0092] Among them, d is the voltage fluctuation, and the two adjacent extreme values ​​U in the voltage effective value change MAX with U MIN , U N is the nominal voltage;

[0093] S2, using LOFR-NDCV algorithm, filter out abnormal collection data and repair Liqun data, and identify abnormal data in power data;

[0094] The training set part of the standard power data set provided by the European Intelligent Technology Network (EUINITE 1997-1999) is used. The data set information is shown in the following table.

[0095] Table 1 Experimental dataset information

[0096] Dataset Total number of items Number of features Data date EUINITE (1997-1999) training set 730 48 01 / 01 / 1997-12 / 31 / 1998

[0097] In order to test the feasibility of the outlier data identification and repair algorithm proposed in this chapter, the electric energy values ​​of the data points in the EUNITE training data set are manually intervened according to a certain proportion, and the information of any column of data points in the data set is randomly modified by 3 to 5 times the size of each other in every other row; the LOFR algorithm data set information is shown in Table 2:

[0098] Table 2 LOFR algorithm dataset

[0099]

[0100] In order to test the feasibility of the abnormal data recognition algorithm, the electric energy values ​​of the data points in the EUNITE training data set are manually intervened according to a certain proportion. The information of any column of data points is randomly modified by 1.5 to 3 times every 5 to 8 rows in the data set. The NDCV data set information is shown in Table 3:

[0101] Table 3 NDCV algorithm dataset

[0102] Abnormal dataset name Injection ratio Total number of data Number of abnormal data Number of features Abnormal data identification algorithm dataset 6％ 730 43 48

[0103] The experimental results of correlation analysis of real data and outlier data are as follows: Figure 2 shown; from Figure 2 From the experimental results, it can be seen that the correlation coefficient between the repaired outlier data and the real data in the regression analysis index reached 0.90927, which theoretically reflects the feasibility of LOFR algorithm data repair.

[0104] In order to verify the effectiveness and feasibility of the abnormal data identification method based on the NDCV algorithm, the NDCV algorithm is experimentally verified using an abnormal data set; at the same time, the IQR algorithm, DBSCAN algorithm, K-means algorithm, normal distribution algorithm and the NDCV algorithm proposed in this invention are used for comparative experiments. The experimental results are shown in Table 4:

[0105] Table 4. Abnormal data detection results under different algorithms

[0106] algorithm Missed detection rate False positive rate Accuracy IQR 0 0.075 0.925 DBSCAN 0 0.055 0.945 K-means 0 0.115 0.885 normal distribution 0 0.055 0.945 NDCV 0 0 1

[0107] From the experimental data in Table 4, it can be seen that the abnormal data recognition algorithm based on cross-normal distribution proposed in the present invention is higher than the other four algorithms in accuracy, which is 0.055 higher than the accuracy of normal distribution, 0.115 higher than the accuracy of K-means algorithm, 0.055 higher than the accuracy of DBSCAD algorithm, and 0.075 higher than the accuracy of IQR algorithm. The experiment proves the feasibility and effectiveness of the algorithm proposed in this paper.

[0108] In order to verify the effectiveness and feasibility of the improved IGEP-CNN algorithm proposed in the present invention, the present invention conducts experimental analysis on the running time and iteration speed performance of the GEP algorithm to verify the effectiveness and feasibility of the three improved methods;

[0109] In order to verify the effectiveness of the improved IGEP-CNN algorithm proposed in this paper, the present invention conducts a comparative experiment on the algorithm model. First, in order to prevent the CNN algorithm from overfitting, the CNN algorithm is pre-trained and the iteration time required for this data set is selected. The results of pre-training are as follows Figure 3 As shown;

[0110] from Figure 3 It can be seen that in the pre-trained model verification, the loss value of the verification set reaches the lowest at the 8th time, so the iteration parameter for CNN training of the data set in this chapter is selected as 8; the CNN model and the IGEP-CNN and GEP-CNN models are compared and the experimental results are shown in Tables 5 and 6 below:

[0111] Table 5 Comparison of evaluation indicators of CNN, GEP-CNN, and IGEP-CNN models

[0112]

[0113]

[0114] Table 6 Comparison of evaluation indicators of CNN, GEP-CNN, and IGEP-CNN model fitting and algorithm prediction results

[0115] index CNN GEP-CNN IGEP-CNN MAPE 0.376266552 0.028255882 0.025456566 RMSE 15.16656525 12.88415925 10.51665655 MAE 14.26415816 12.55628585 12.15615633 <![CDATA[R 2 ]]> 0.321598959 0.755265226 0.802023569

[0116] From Table 5, it can be found that the fitting parameters of the model after the CNN algorithm is combined with the GEP algorithm have been greatly improved, which further illustrates that the optimized combination of the GEP algorithm and the CNN model can improve the nonlinear fitting ability of the model. The correlation coefficient after the combination changes from 0.825612059 to 0.972325137, an increase of 0.147. At the same time, the GEP algorithm is combined with CNN before and after improvement. It can be found in Table 5 that the fitting parameters of the combined model of the improved GEP algorithm have been further improved, and the correlation coefficient has increased from 0.972325137 to 0.9910559596, an increase of 0.0187, which further illustrates that the improved GEP algorithm is more advantageous for the combined model of CNN.

[0117] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A power data evaluation method based on IGEP-CNN algorithm, characterized in that: The steps include: S1. Analyze the power quality of the collected power data for harmonics, voltage deviation, frequency deviation, three-phase imbalance, and voltage fluctuation; S2, using LOFR-NDCV algorithm, filter out abnormal collection data and repair Liqun data, and identify abnormal data in power data; S3, improve the classic GEP algorithm, including initializing population size 、population head and population number Parameters; Recombination rate P for the three major genetic operators (i,r) , mutation rate P (i,m) , transposition rate P (i,ro) Set the adaptive evolution probability value; solve the local optimal problem through the IA-IP algorithm; The population size The calculation is as follows: Among them, ROUND() is the rounding function; n is the iterative evolution parameter of the algorithm; f i is the fitness size; C (i，j) is the model prediction value; T j is the true value; The population number The calculation is as follows: The population head The calculation is as follows: The recombination rate P (i,r) The calculation is as follows: The mutation rate P (i,m) The calculation is as follows: The transposition rate P (i,ro) The calculation is as follows: S4. Power data risk assessment and prediction based on IGEP-CNN algorithm; The power data risk assessment and prediction based on the IGEP-CNN algorithm is to use the IGEP algorithm to automatically find the appropriate network parameters. The adaptive equation of the CNN algorithm applied to the IGEP algorithm is formula 8. The expression of the population individuals of the GEP algorithm is combined with the neural network parameters. The chromosome individual expression is the neural network parameter. Among them, f i is the total negative loss value, which is also the fitness value; loss (i，j) is the CNN prediction value; T j is the true value.

2. The power data evaluation method based on the IGEP-CNN algorithm according to claim 1 is characterized in that: The harmonics are calculated as follows: Among them, w1 represents the angular frequency, c0 represents the DC component; c h , are the amplitude and initial phase of the hth harmonic respectively; The voltage deviation is calculated as follows: Among them, ΔU is the voltage deviation, U r is the voltage value at the measuring point, U N is the nominal voltage; The frequency deviation is calculated as follows: Δf=f r -f N (11) Where Δf is the frequency deviation, f r is the frequency value of the measurement point, f N is the nominal frequency; The calculation formula for the three-phase imbalance is: Among them, ε u , ε i are the three-phase unbalance of voltage and current, u1, i1 and u2, i2 represent the RMS values ​​of the positive-sequence voltage and negative-sequence current components, respectively; The voltage fluctuation formula is: Among them, d is the voltage fluctuation, and the two adjacent extreme values ​​U in the voltage effective value change MAX with U MIN , U N is the nominal voltage.

3. The power data evaluation method based on the IGEP-CNN algorithm according to claim 1, characterized in that: The method of using the LOFR-NDCV algorithm to filter out abnormal data and repair the Liqun data to identify abnormal data in the power data includes the following steps: S21. Suppose there is an m×n dimensional electric energy data set O m×n , define the power data o ij The outlier factor lof(o ij ), for any power data o ij ,i∈{1,2,...,m},j∈{1,2,...,n}, if lof(o ij )<1, it is normal data. If lof(o ij )>1, then the electric energy data is called ij For outlier data, the median filtering algorithm is used to replace them; S22, abnormal data identification based on NDCV algorithm to solve the misjudgment of normal data; each power data o ij The mean and variance of the row and column are μ i , μ j , And, for any power data o ij , where i∈{1,2,...,m},j∈{1,2,...,n}, if and If the electric energy data is established, then ij Abnormal electric energy value.

4. The power data evaluation method based on the IGEP-CNN algorithm according to claim 1 is characterized in that: The IA-IP algorithm solves the local optimal problem by: Record the individual with the highest fitness in the population, and detect the number of times this maximum fitness value appears in the subsequent evolution process; If the number of occurrences is within the threshold δ (i) If it is updated within , the updated fitness value is recorded and the number of times the updated value appears in the evolution process is detected; If the number of occurrences exceeds the threshold δ (i) , then new individuals are introduced to replace several individuals with lower fitness in the generation that exceeds the specified threshold; The number of replacement individuals is equal to the number of new individuals introduced; After replacement, the fitness value of each individual in the population is recalculated.

Images