User abnormal load identification method based on stereo portrait library and dual-channel weighting

Abstract

The invention discloses a user abnormal load identification method based on a stereo portrait library and dual-channel weighting, and the method comprises the steps: collecting the load data and industry labels of a user, cleaning the data to obtain a load data set, and dividing a training data set and a test data set; a user load characteristic analysis module is constructed, the unbalance degree between different users and different industry types in the training data set is counted, and a user multi-time-scale typical power consumption curve and a multi-dimensional load characteristic index are extracted in a clustering mode; combining a user load characteristic analysis result with an industry label, performing weighting on each industry user, performing secondary clustering to extract an industry multi-time scale typical power consumption curve and a multi-dimensional load characteristic index, and constructing a three-dimensional portrait library; and after load curve-feature dual-channel weighting is carried out on each industry in the stereo portrait library, load type and abnormal power consumption behavior identification is carried out on unknown users in the test data set. According to the invention, high-accuracy identification of the user load type and the abnormal power consumption behavior is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of smart electric meters and electricity consumption big data analysis, and in particular to a method for identifying abnormal user load based on a stereoscopic portrait library and dual-channel weighting. Background Art

[0002] The characteristics and differences of electricity consumption behaviors of users with different load types at different time scales are very obvious and prominent. For example, during holidays, the electricity consumption of places such as hotels, restaurants and bars surges; the electricity consumption behavior of residential charging piles during commuting hours and non-commuting hours is obviously different, and the pressure to ensure power supply and power safety is huge. It is necessary to identify the user load type in order to formulate a reasonable power supply strategy; secondly, some users have abnormal electricity consumption behaviors, such as illegal commercial use of residential electricity, which poses a huge potential safety hazard. Therefore, accurate user abnormal load identification methods are of great significance to ensuring and maintaining a safe and standardized power supply and use environment.

[0003] At present, there are the following difficulties in identifying abnormal user loads:

[0004] 1. User load levels vary significantly with fluctuations in meteorological factors and economic environment, making it difficult to characterize user electricity usage habits and behavioral characteristics; 2. Data is unbalanced between different users and between different industries, and identification strategies are difficult to focus on some users and industries with small amounts of sample data, resulting in reduced load identification accuracy; 3. User load type differences are small between classes, while diversity within classes is high, resulting in high clustering difficulty, making it difficult to effectively extract load features.

[0005] The above problems have increased the difficulty of identifying abnormal loads of users to a certain extent, reduced the accuracy of abnormal load identification, and made the traditional abnormal load identification method have low accuracy in identifying user load types and abnormal power consumption behaviors.

[0006] Therefore, a user abnormal load identification method based on a three-dimensional portrait library and dual-channel weighting is provided to achieve high-accuracy identification of user load types and illegal electricity usage behaviors. As a part of distribution network operation and maintenance and planning, it can better characterize user electricity usage habits and behavioral characteristics, effectively assist operation and maintenance personnel in formulating reasonable power supply strategies, promote the application of advanced data processing technology in distribution network operation and maintenance and planning scenarios, support the digital and intelligent reform of distribution networks, and help build new power systems. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for identifying user abnormal loads based on a stereoscopic portrait library and dual-channel weighting, which can achieve high-accuracy identification of user load types and abnormal power consumption behaviors, and solve the problem of low accuracy of traditional load identification methods in identifying user load types and abnormal power consumption behaviors. As a part of distribution network operation and maintenance and planning, it will help to ensure the safety of power supply and use.

[0008] To achieve the above object, the technical solution provided by the present invention is: a method for identifying abnormal user load based on a stereo image library and dual-channel weighting, comprising the following steps:

[0009] 1) Collect users' load data and industry labels, and clean the load data, including outlier removal, missing value filling and standardization. The cleaned load data constitutes a load data set, which is divided into a training data set and a test data set according to the proportion;

[0010] 2) Construct a user load characteristic analysis module that includes a statistical module and a clustering module, calculate the imbalance degree between different users and different industries in the training data set, cluster and extract the user's multi-time scale typical electricity consumption curve, and then use the user's multi-time scale typical electricity consumption curve to calculate the user's multi-dimensional load characteristic index to obtain the user load characteristic analysis result;

[0011] 3) Combine the user load characteristic analysis results with industry labels, calculate the user and industry weight coefficients according to the imbalance degree between different users and different industries, and improve the density peak clustering algorithm by weight. Then, use the weighted improved density peak clustering algorithm to perform secondary clustering to obtain the typical power consumption curve of the industry at multiple time scales, and calculate the multi-dimensional load characteristic indicators of the industry. Use the typical power consumption curves of multiple time scales and multi-dimensional load characteristic indicators of each industry to build a three-dimensional portrait library;

[0012] 4) Design a load curve-feature dual-channel weighted identification strategy. After performing load curve-feature dual-channel weighting on each industry in the stereoscopic portrait library, calculate the total correlation coefficient between the unknown users in the test data set and each industry in the stereoscopic portrait library, and obtain the identification results of the unknown user's load type and abnormal power consumption behavior.

[0013] Further, the specific operation steps of step 1) are as follows:

[0014] 1.1) Obtain the load data of each user in the form of daily load series;

[0015] 1.2) Traverse the daily load sequence of each user one by one, and judge the missing situation of the user's daily load sequence. If the daily load sequence is missing more than 30% of the sampling data continuously, the daily load sequence will be abandoned; for the retained user daily load sequence, the average value and standard deviation of the daily load sequence are calculated using the following formula:

[0016]

[0017] In the formula, is the average value of the daily load sequence, x(i) is the load data at the i-th sampling time point in the daily load sequence, i = 1, 2, …, n, n is the total number of sampling points in the daily load sequence, and s is the standard deviation of the daily load sequence;

[0018] 1.3) Traverse the load data of each sampling time point in the daily load sequence. If the difference between the load data corresponding to the sampling time point and the average value of the daily load sequence is greater than 3 times the standard deviation of the load sequence of the day, it will be regarded as abnormal data and removed. After removal, the load data corresponding to the sampling time point will be treated as missing values;

[0019] 1.4) Traverse the user's daily load sequence and determine whether the load data corresponding to the adjacent sampling time points of the missing value exists. If the load data corresponding to the adjacent sampling time points before and after are all missing values, the missing value is assigned to 0; if the sampling time point corresponding to the missing value is the first sampling time point or the last sampling time point in the daily load sequence, the load data corresponding to the only adjacent sampling time point is used to replace the missing value; if the load data corresponding to the adjacent sampling time points are all normal values, the average value of the load data corresponding to the adjacent sampling time points before and after is used to replace the missing value;

[0020] 1.5) Traverse each sampling time point in the user's daily load sequence and map the corresponding load data to the interval [0,1]. The calculation formula is as follows:

[0021]

[0022] Where, L max , L min They represent the maximum and minimum values ​​of the user's daily load sequence, respectively, and x i ′ is the load data at the i-th sampling time point after standardization.

[0023] Further, the specific operation steps of step 2) are as follows:

[0024] 2.1) Construct a user load characteristic analysis module, which consists of a statistical module and a clustering module, which are used to count and divide load data and cluster to obtain typical power consumption curves of users at multiple time scales;

[0025] 2.2) The statistical module first counts the imbalance between different users and different industries in the training data set. The imbalance between different users includes the number of users included in the user load data and the number of daily load sequences included in each user; the imbalance between different industries includes the number of industry types included in the user load data and the total number of users in each industry;

[0026] Then, the user industry label data is counted to generate the user industry label sequence H Y , as shown below:

[0027] H Y ={H1,H2,…,H H}

[0028] In the formula, H H Represents the industry label corresponding to the Hth industry;

[0029] Finally, based on the time index in the user load data and combined with the statutory holiday information, the user load data is divided into four time scales: year, month, day and holiday;

[0030] 2.3) The clustering module first calculates the expected daily load sequence E of the user within the time scale T x 、Entropy E n and super entropy H e , T′∈{year, month, day, holiday}, generating cloud droplets Cloud(E X ,E n ,H e ), which is used to calculate the shape similarity between daily load sequences, as shown in the following formula:

[0031]

[0032]

[0033]

[0034] Where, L k and L j denote the kth and jth daily load sequences respectively, k = 1, 2, ..., N, j = 1, 2, ..., N, N is the total number of daily load sequences contained in the user time scale T′, D(L k ,L j ) represents the comprehensive similarity between the cloud droplets generated by the kth and jth daily load sequences, En k and En j denote the entropy of cloud droplets generated by the kth and jth daily load sequences, He k and He jdenote the super entropy of cloud droplets generated by the kth and jth daily load sequences, S(E xk -E xj |) represents the overlap area between the expected cloud droplet curves generated by the kth and jth daily load sequences, and S(0) represents the overlap area when the distance between the expected cloud droplet curves generated by the kth and jth daily load sequences is 0;

[0035] The density peak clustering algorithm is used for clustering. For the user daily load sequences within the same time range, the local density and relative distance of each daily load sequence within the corresponding time range are calculated. The calculation formula is as follows:

[0036]

[0037]

[0038] In the formula, ρ k and δ k denote the local density and relative distance of the kth daily load sequence, ρ j , i denote the local density of the jth and ith daily load series, respectively, d c represents the cutoff distance, X(·) represents a piecewise function with a value of 0 or 1. If the value in the brackets is negative, the function value is 1; if the value in the brackets is positive, the function value is 0;

[0039] After obtaining the local density and relative distance of the daily load sequence, the central value and outlier degree of the daily load sequence are calculated. The calculation formula is as follows:

[0040] λ k =ρ k 2 +δ k 3

[0041]

[0042] In the formula, λ k and k They represent the central value and outlier degree of the k-th daily load series respectively;

[0043] Arrange all daily load sequences within the time scale T′ from large to small according to the degree of outliers, and exclude the first q daily load sequences as outlier daily load sequences; then arrange all the remaining daily load sequences from large to small according to the center value, take the first w daily load sequences as the initial clustering centers, create d initial clusters, and assign the remaining daily load sequences to the clusters to which the clustering centers with the smallest relative distances belong in turn to obtain the initial clustering results;

[0044] Traverse the clusters in the initial clustering results and use the following formula to calculate the distance between clusters:

[0045]

[0046] Where D c (C a ,C b ) represents cluster C a and cluster C b The distance between a C and N b C Represent cluster C a and cluster C b The total number of daily load sequences within ;

[0047] Merge the two clusters with the smallest inter-cluster distance, traverse the remaining d-1 clusters, and recalculate the inter-cluster distance; if all the calculated inter-cluster distances are greater than the set threshold D c,max When , the merging stops and the clustering results are saved; among the k clusters finally saved, the daily load sequence with the smallest distance from other daily load sequences in the cluster is selected and output as the typical power consumption curve of multiple time scales of the user in the time scale T′;

[0048] 2.4) After obtaining the typical power consumption curve of the user at multiple time scales, calculate the user's multi-dimensional load characteristic index to form a user's multi-dimensional load characteristic index sequence, including daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient, and normal period power consumption percentage. The definition and specific calculation formula of each index are as follows:

[0049] The daily minimum load rate C1 is the ratio of the minimum value of the daily load sequence to the maximum value of the daily load sequence:

[0050]

[0051] The daily average load factor C2 is the ratio of the average value of the daily load sequence to the maximum value of the daily load sequence:

[0052]

[0053] Where, L a represents the average value of daily load series;

[0054] The daily peak-to-valley difference rate C3 is the ratio of the peak-to-valley difference of the daily load sequence to the maximum value of the daily load sequence:

[0055]

[0056] The valley power coefficient C4 is the ratio of the power consumption during the valley period to the total power consumption of the daily load sequence:

[0057]

[0058] Where, L G Indicates the power consumption during off-peak hours, L T Indicates the total power consumption of the daily load sequence;

[0059] The normal period electricity consumption percentage C5 is the ratio of normal period electricity consumption to the total electricity consumption of the daily load sequence:

[0060]

[0061] Where, L p Indicates the power consumption during normal periods.

[0062] Further, the specific operation steps of step 3) are as follows:

[0063] 3.1) Combine the user load characteristic analysis results with the industry label, calculate the user and industry weighting coefficients according to the imbalance degree between different users and different industry types, and save them to the compensation module in the stereo image library;

[0064] The calculation formula of user weight coefficient is as follows:

[0065]

[0066] In the formula, v Yy represents the user weight coefficient of the yth user, Y y represents the number of daily load sequences of the yth user, y = 1, 2, ..., p z , p z represents the total number of users, Y max and Y min They represent the maximum and minimum values ​​of the number of user daily load sequences respectively;

[0067] The industry weighting coefficient includes the industry curve weighting coefficient and the industry characteristic weighting coefficient. The calculation formula is as follows:

[0068]

[0069]

[0070] In the formula, v Ch and v fh They represent the industry curve weighting coefficient and industry characteristic weighting coefficient of the hth industry, respectively. h represents the total number of users in the hth industry, h = 1, 2, ..., H, H is the number of industry types included in the user load data, p max and p min Respectively represent the maximum and minimum values ​​of the total number of industry users;

[0071] 3.2) Obtain the typical power consumption curve of multiple time scales of each user in the time scale T′ of the training data set, divide the users with the same industry label into the same range, and perform secondary clustering using the improved density peak clustering algorithm. When calculating the local density and relative distance of the typical power consumption curve of multiple time scales of the user, the industry curve weighting coefficient is used to perform weighted improvement on the density peak clustering algorithm, as shown in the following formula:

[0072]

[0073]

[0074] In the formula, ρ' y and δ' y Respectively represent the local density and relative distance of the typical power consumption curve of the yth user, L' y and L' z They represent the typical electricity consumption curves of the y-th user and the z-th user respectively;

[0075] After secondary clustering using the improved density peak clustering algorithm, the typical power consumption curves of multiple time scales of all industries within the time scale T′ are obtained and saved in the curve module in the stereoscopic image library. For each industry, the typical power consumption curves of multiple time scales of the industry within the daily time scale are used as the reference sequence. The mean value of the correlation coefficient between the typical power consumption curves of multiple time scales of users and the reference sequence is calculated as the threshold value of the correlation coefficient of the industry curve, as shown in the following formula:

[0076]

[0077] In the formula, ξ ch represents the curve correlation coefficient threshold of the h-th industry, x h0 (i) is the load data at the i-th sampling time point in the typical electricity consumption curve of the h-th industry, x hy (i) is the load data of the ith sampling time point in the daily load sequence of the yth user in the hth industry, ρ is the resolution coefficient, and Respectively represent the minimum difference between two levels and the maximum difference between two levels;

[0078] 3.3) Obtain the multi-dimensional load characteristic index of each user in the training data set, weight it using the user weighting coefficient, calculate the multi-dimensional load characteristic index of each industry, and save it as the industry multi-dimensional load characteristic index sequence to the feature module of the stereoscopic image library, as shown in the following formula:

[0079]

[0080] In the formula, C hgrepresents the g-th multi-dimensional load characteristic index of the h-th industry, g = 1, 2, 3, 4, 5, corresponding to the five load characteristic indexes of daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and normal period power consumption percentage respectively; C yg represents the g-th multi-dimensional load characteristic index of the y-th user;

[0081] For each industry, the industry multi-dimensional load characteristic index sequence is used as the reference sequence, and the mean value of the correlation coefficient between the user multi-dimensional load characteristic index sequence and the reference sequence is calculated as the industry characteristic correlation coefficient threshold, as shown in the following formula:

[0082]

[0083] In the formula, ξ fh Indicates the characteristic correlation coefficient threshold of the h-th industry;

[0084] 3.4) Add the industry curve correlation coefficient threshold to the industry characteristic correlation coefficient threshold to obtain the industry correlation coefficient threshold, as shown in the following formula:

[0085] ξ h =ξ ch +ξ fh

[0086] In the formula, ξ h represents the correlation coefficient threshold of the h-th industry;

[0087] 3.5) According to the changes in user load data in different time scales and the electricity consumption characteristics of different industries, set the curve weight matrix θ C and the feature weight matrix θ f , where the curve weight matrix θ C =(θ C1 ,θ C2 ,θ C3 ,θ C4 ) represent the attention weights of the typical electricity consumption curves of each industry in the year, month, day and holiday time scales; the feature weight matrix θ f =(θ f1 ,θ f2 ,θ f3 ,θ f4 ,θ f5 ) respectively represent the attention weights of the five multi-dimensional load characteristic indicators, namely, daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and percentage of electricity consumption during normal hours. The two weight matrices are saved in the weight module in the stereo image library.

[0088] Further, the specific operation steps of step 4) are as follows:

[0089] 4.1) Obtain the unknown user daily load sequence L from the test dataset N , calculate the multi-dimensional load characteristic index of unknown users and form a multi-dimensional load characteristic index sequence C N , as shown below:

[0090] C N ={C N1 ,C N2 ,C N3 ,C N4 ,C N5}

[0091] In the formula, C N1 , C N2 , C N3 , C N4 and C N5 They represent the daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient, and normal-time power consumption percentage of unknown users respectively;

[0092] 4.2) Extract the typical electricity consumption curves of all industries in years, months, days and holidays from the stereoscopic image library, as shown in the following formula:

[0093] L Yh ={x Yh (1),x Yh (2),…,x Yh (n)}

[0094] L Mh ={x Mh (1),x Mh (2),…,x Mh (n)}

[0095] L Dh ={x Dh (1),x Dh (2),…,x Dh (n)}

[0096] L Jh ={x Jh (1),x Jh (2),…,x Jh (n)}

[0097] Where, L Yh , L Mh , L Dh and L Jh represents the typical electricity consumption curves of the hth industry in the year, month, day and holiday, x Yh (n), x Mh (n), x Dh (n) and x Jh(n) represents the load data of the nth sampling time point in the typical electricity consumption curve of the year, month, day and holiday of the hth industry;

[0098] Take the unknown user daily load sequence L N The curve correlation coefficient ζ between the reference sequence and the typical electricity consumption curve of the hth industry in the year, month, day and holiday at the i-th sampling time point is calculated. Yh (i), ξ Mh (i), ξ Dh (i) and ξ Jh (i), as shown below:

[0099]

[0100]

[0101]

[0102]

[0103] In the formula, and Respectively represent the minimum difference between the two levels and the maximum difference between the two levels, x Yh (i), x Mh (i), x Dh (i) and x Jh (i) represents the load data of the typical electricity consumption curve of the hth industry at the i-th sampling time point in the year, month, day and holiday respectively;

[0104] Combining the curve correlation coefficients of n sampling time points, we can get the curve correlation coefficient R between the reference sequence and the typical electricity consumption curves of the hth industry in the year, month, day and holiday. Yh , R Mh , R Dh and R Jh :

[0105]

[0106]

[0107]

[0108]

[0109] Traverse the typical electricity consumption curves of all industries in the stereoscopic image library by year, month, day and holiday, and obtain the 4×H curve correlation coefficient matrix R1:

[0110]

[0111] In the formula, RYH , R MH , R DH and R JH They represent the curve correlation coefficients between the reference sequence and the typical electricity consumption curves of the Hth industry in the year, month, day and holiday;

[0112] With the curve weight matrix θ C =(θ C1 ,θ C2 ,θ C3 ,θ C4 ) to obtain the 1×H weighted curve correlation coefficient matrix R of unknown users and industries. C :

[0113] R c =θ c R1

[0114] 4.3) Extract multi-dimensional load characteristic index sequences of various industries from the stereoscopic image library:

[0115] C h ={C h1 ,C h2 ,C h3 ,C h4 ,C h5}

[0116] In the formula, C h represents the multi-dimensional load characteristic index sequence of the h-th industry, C h1 , C h2 , C h3 , C h4 and C h5 They are the daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and percentage of electricity consumption during normal hours for the hth industry;

[0117] Taking the unknown user multidimensional load characteristic index sequence as the reference sequence, calculate the characteristic correlation coefficient ζ of the g-th multidimensional load characteristic index between the reference sequence and the multidimensional load characteristic index sequence of the h-th industry hg , as shown below:

[0118]

[0119] In the formula, C Ng represents the g-th multi-dimensional load characteristic index of the unknown user;

[0120] Traversing the characteristic sequences of each industry, we obtain the 5×H characteristic correlation coefficient matrix R2:

[0121]

[0122] In the formula, ξH1 , H2 , H3 , H4 and H5 They are the characteristic correlation coefficients between the reference sequence and the multi-dimensional load characteristic index sequence of the Hth industry in terms of daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and percentage of electricity consumption during normal periods;

[0123] Using the feature weight matrix θ f =(θ f1 ,θ f2 ,θ f3 ,θ f4 ,θ f5 ) weights the feature correlation coefficients to obtain a 1×H weighted feature correlation coefficient matrix R f :

[0124] R f =θ f R2

[0125] 4.4) Read the h×1 industry curve weighting coefficient matrix ν from the stereo image library C and the industry characteristic weighted coefficient matrix ν f , for R C and R f Perform dual-channel weighting as shown below:

[0126] ν C =(ν C1 ,ν C2 ,…,ν CH ) T

[0127] ν f =(ν f1 ,v f2 ,…,v fH ) T

[0128] R C =v C R c

[0129] R F =v f R f

[0130] In the formula, R C and R F They represent the curve correlation coefficient matrix and feature correlation coefficient matrix after dual-channel weighting respectively;

[0131] 4.5) Using the curve correlation coefficient matrix and feature correlation coefficient matrix after dual-channel weighting, calculate the total correlation coefficient matrix between unknown users and each industry in the stereoscopic portrait library, and take the index value corresponding to the maximum value as the index value corresponding to the user load type, as shown in the following formula:

[0132] R = argmax(diag(R C +R F ))

[0133] Where R is the index value corresponding to the unknown user load type, and the argmax(·) function represents the index value corresponding to the maximum value in the matrix;

[0134] According to the index value corresponding to the user load type, the corresponding industry label is read from the industry label sequence to obtain the user load type identification result;

[0135] 4.6) After obtaining the unknown user load type identification result, the corresponding industry correlation coefficient threshold is obtained from the constructed three-dimensional portrait library. If the correlation coefficient between the unknown user and the industry is less than the corresponding industry correlation coefficient threshold, it is considered that the unknown user's electricity consumption behavior deviates from the normal electricity consumption behavior habits of the corresponding industry, and there may be abnormal electricity consumption behavior. An abnormal alarm signal is output, which is further used to guide the work of troubleshooting electricity hazards.

[0136] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0137] 1. When cleaning user load data, the present invention uses the triple standard deviation principle to remove outliers and fill missing values ​​in user data, significantly reducing the number of days with missing user data, reducing the proportion of bad data, and significantly improving the quality of load data.

[0138] 2. When extracting typical electricity consumption curves of users at multiple time scales, the present invention uses the comprehensive similarity between cloud droplets generated by the user's daily load sequence as a standard, and has excellent characterization performance for the correlation between the user's daily load sequences. The density peak clustering method is improved by using weighted coefficients, and the clustering center can be determined more efficiently, thereby accurately extracting typical electricity consumption curves of users and industries and effectively extracting user load characteristics.

[0139] 3. The present invention fully considers the imbalance between different users and different industry types in the load data, introduces user and industry weighting coefficients when constructing a three-dimensional portrait library, compensates for industries with a small number of samples, effectively increases the attention paid to industries that are easy to identify errors, and thus improves the accuracy of user load type identification.

[0140] 4. The present invention adopts a load curve-feature dual-channel weighted identification strategy. After weighting the load curve-feature dual-channel of each industry in the three-dimensional portrait library, it can more comprehensively and accurately measure the degree of association between unknown users and each industry in the three-dimensional portrait library, and realize efficient and accurate identification of user load types and abnormal electricity consumption behaviors, which helps safe electricity management and is more suitable for actual projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0141] Figure 1 The figure is a logical flow diagram of the method of the present invention.

[0142] Figure 2 It is a typical electricity consumption curve diagram of each industry in the embodiment of the method of the present invention within a daily time scale. DETAILED DESCRIPTION

[0143] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0144] Table 1 below is a user load data structure of an embodiment of a method for identifying abnormal user load based on a stereo image library and dual-channel weighting of the present invention:

[0145] Table 1

[0146]

[0147]

[0148] Table 2 below is a user industry label data structure of an embodiment of a method for identifying abnormal user load based on a stereoscopic image library and dual-channel weighting of the present invention:

[0149] Table 2

[0150] User Number Industry Type 1 B&B 2 Charging Station 3 Restaurants 4 barbecue 5 Residential electricity consumption 6 bar

[0151] like Figure 1 As shown, this embodiment discloses a method for identifying abnormal user load based on a stereo image library and dual-channel weighting, the specific situation of which is as follows:

[0152] 1) Collect user load data and industry labels, and perform data cleaning on the load data, including outlier removal, missing value filling and standardization. The cleaned load data constitutes a load data set, which is divided into a training data set and a test data set in a ratio of 7:3. The specific steps are as follows:

[0153] 1.1) Obtain the load data of each user in the form of daily load series;

[0154] 1.2) Traverse the daily load sequence of each user one by one, and judge the missing situation of the user's daily load sequence. If the daily load sequence is missing more than 30% of the sampling data continuously, the daily load sequence will be abandoned; for the retained user daily load sequence, the average value and standard deviation of the daily load sequence are calculated using the following formula:

[0155]

[0156]

[0157] In the formula, is the average value of the daily load sequence, x(i) is the load data at the i-th sampling time point in the daily load sequence, i=1,2,…,n, n is the total number of sampling points in the daily load sequence, n=1440 in the embodiment, and s is the standard deviation of the daily load sequence.

[0158] 1.3) Traverse the load data of each sampling time point in the daily load sequence. If the difference between the load data corresponding to the sampling time point and the average value of the daily load sequence is greater than 3 times the standard deviation of the load sequence of the day, it will be regarded as abnormal data and eliminated. After elimination, the load data corresponding to the sampling time point will be treated as missing values.

[0159] 1.4) Traverse the user's daily load sequence and determine whether the load data corresponding to the adjacent sampling time points of the missing value exists. If the load data corresponding to the adjacent sampling time points before and after are all missing values, the missing value is assigned to 0; if the sampling time point corresponding to the missing value is the first sampling time point or the last sampling time point in the daily load sequence, the load data corresponding to the only adjacent sampling time point is used to replace the missing value; if the load data corresponding to the adjacent sampling time points are all normal values, the average value of the load data corresponding to the adjacent sampling time points before and after is used to replace the missing value;

[0160] 1.5) Traverse each sampling time point in the user's daily load sequence and map the corresponding load data to the interval [0,1]. The calculation formula is as follows:

[0161]

[0162] Where, L max , L min They represent the maximum and minimum values ​​of the user's daily load sequence, respectively, and x i ′ is the load data at the i-th sampling time point after standardization.

[0163] 2) Construct a user load characteristic analysis module including a statistical module and a clustering module, calculate the imbalance degree between different users and different industries in the training data set, cluster and extract the typical power consumption curve of users in multiple time scales, and then use the typical power consumption curve of users in multiple time scales to calculate the multi-dimensional load characteristic index of users, and obtain the analysis results of user load characteristics. The specific steps are as follows:

[0164] 2.1) Construct a user load characteristic analysis module, which consists of a statistical module and a clustering module, which are used to count and divide load data and cluster to obtain typical power consumption curves of users at multiple time scales.

[0165] 2.2) The statistical module first counts the imbalance between different users and between different industries in the training data set. The imbalance between different users includes the number of users included in the user load data and the number of daily load sequences contained in each user; the imbalance between different industries includes the number of industry types included in the user load data and the total number of users in each industry.

[0166] Then, the user industry label data is counted to generate the user industry label sequence H Y , as shown below:

[0167] H Y ={H1,H2,…,H H}

[0168] In the formula, H H Represents the industry label corresponding to the Hth industry.

[0169] Finally, based on the time index in the user load data and combined with the statutory holiday information, the user load data is divided into four time scales: year, month, day and holiday;

[0170] 2.3) The clustering module first calculates the expected daily load sequence E of the user within the time scale T x 、Entropy E n and super entropy H e , T′∈{year, month, day, holiday}, generating cloud droplets Cloud(E X ,E n ,H e ), which is used to calculate the shape similarity between daily load sequences, as shown in the following formula:

[0171]

[0172]

[0173]

[0174]

[0175] Where, L k and L j denote the kth and jth daily load sequences respectively, k = 1, 2, ..., N, j = 1, 2, ..., N, N is the total number of daily load sequences contained in the user time scale T′, D(L k ,L j ) represents the comprehensive similarity between the cloud droplets generated by the kth and jth daily load sequences, En k and En j denote the entropy of cloud droplets generated by the kth and jth daily load sequences, He k and He j denote the super entropy of cloud droplets generated by the kth and jth daily load sequences, S(E xk -E xj |) represents the overlap area between the expected cloud droplet curves generated by the kth and jth daily load sequences, and S(0) represents the overlap area when the distance between the expected cloud droplet curves generated by the kth and jth daily load sequences is 0.

[0176] The density peak clustering algorithm is used for clustering. For the user daily load sequences within the same time range, the local density and relative distance of each daily load sequence within the corresponding time range are calculated. The calculation formula is as follows:

[0177]

[0178]

[0179] In the formula, ρ k and δ k denote the local density and relative distance of the kth daily load sequence, ρ j , i denote the local density of the jth and ith daily load series, respectively, d c represents the cutoff distance, which is 0.8 in the embodiment, and χ(·) represents a piecewise function with a value of 0 or 1. If the value in the brackets is negative, the function value is 1; if the value in the brackets is positive, the function value is 0.

[0180] After obtaining the local density and relative distance of the daily load sequence, the central value and outlier degree of the daily load sequence are calculated. The calculation formula is as follows:

[0181] λ k =ρ k 2 +δ k 3

[0182]

[0183] In the formula, λ k and k They represent the central value and outlier degree of the k-th daily load series respectively.

[0184] All daily load sequences within the time scale T′ are arranged from large to small according to the outlier degree, and the first q daily load sequences are excluded as outlier daily load sequences; then all the remaining daily load sequences are arranged from large to small according to the center value, and the first w daily load sequences are taken as the initial clustering centers, and d initial clusters are created. The remaining daily load sequences are sequentially assigned to the cluster to which the clustering center with the smallest relative distance belongs to obtain the initial clustering result. In the embodiment, the values ​​of q and w are 10 and 3 respectively.

[0185] Traverse the clusters in the initial clustering results and use the following formula to calculate the distance between clusters:

[0186]

[0187] Where D c (C a ,C b ) represents cluster C a and cluster C b The distance between and Represent cluster C a and cluster C b The total number of daily load sequences within .

[0188] Merge the two clusters with the smallest inter-cluster distance, traverse the remaining d-1 clusters, and recalculate the inter-cluster distance; if all the calculated inter-cluster distances are greater than the set threshold D c,max When , the merging stops and the clustering results are saved. In the k clusters finally saved, the daily load sequence with the smallest distance from other daily load sequences in the cluster is selected and output as the typical power consumption curve of the user in multiple time scales within the time scale T′. In the embodiment, D c,max Take 1.4.

[0189] 2.4) After obtaining the typical power consumption curve of the user at multiple time scales, calculate the user's multi-dimensional load characteristic index to form a user's multi-dimensional load characteristic index sequence, including daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient, and normal period power consumption percentage. The definition and specific calculation formula of each index are as follows:

[0190] The daily minimum load rate C1 is the ratio of the minimum value of the daily load sequence to the maximum value of the daily load sequence:

[0191]

[0192] The daily average load factor C2 is the ratio of the average value of the daily load sequence to the maximum value of the daily load sequence:

[0193]

[0194] Where, L a Represents the average value of daily load series.

[0195] The daily peak-to-valley difference rate C3 is the ratio of the peak-to-valley difference of the daily load sequence to the maximum value of the daily load sequence:

[0196]

[0197] The valley power coefficient C4 is the ratio of the power consumption during the valley period to the total power consumption of the daily load sequence:

[0198]

[0199] Where, L G Indicates the power consumption during off-peak hours, L T Represents the total electricity consumption of the daily load sequence.

[0200] The normal period electricity consumption percentage C5 is the ratio of normal period electricity consumption to the total electricity consumption of the daily load sequence:

[0201]

[0202] Where, L p Indicates the power consumption during normal periods.

[0203] 3) Combine the user load characteristic analysis results with the industry label, calculate the user and industry weight coefficients according to the imbalance degree between different users and different industries, and improve the density peak clustering algorithm by weight. Then use the weighted improved density peak clustering algorithm to perform secondary clustering to obtain the typical power consumption curve of the industry at multiple time scales, and calculate the multi-dimensional load characteristic index of the industry. Figure 2 The three-dimensional image library is constructed by using the load characteristic index (as shown in Figure 1) and multi-dimensional load characteristic indexes. The specific steps are as follows:

[0204] 3.1) Combine the user load characteristic analysis results with the industry label, calculate the user and industry weighting coefficients according to the imbalance degree between different users and different industry types, and save them to the compensation module in the stereo image library;

[0205] The calculation formula of user weight coefficient is as follows:

[0206]

[0207] In the formula, v Yy represents the user weight coefficient of the yth user, Y y represents the number of daily load sequences of the yth user, y = 1, 2, ..., pz , p z represents the total number of users, Y max and Y min They represent the maximum and minimum values ​​of the user's daily load sequence quantity respectively.

[0208] The industry weighting coefficient includes the industry curve weighting coefficient and the industry characteristic weighting coefficient. The calculation formula is as follows:

[0209]

[0210]

[0211] In the formula, v Ch and ν fh They represent the industry curve weighting coefficient and industry characteristic weighting coefficient of the hth industry, respectively. h represents the total number of users in the hth industry, h = 1, 2, ..., H, H is the number of industry types included in the user load data, p max and p min Respectively represent the maximum and minimum values ​​of the total number of industry users.

[0212] 3.2) Obtain the typical power consumption curve of multiple time scales of each user in the time scale T′ of the training data set, divide the users with the same industry label into the same range, and perform secondary clustering using the improved density peak clustering algorithm. When calculating the local density and relative distance of the typical power consumption curve of multiple time scales of the user, the density peak clustering algorithm is improved using the weighted coefficient of the industry curve, as shown in the following formula:

[0213]

[0214]

[0215] In the formula, ρ' y and δ' y Respectively represent the local density and relative distance of the typical power consumption curve of the yth user, L' y and L' z They represent the typical electricity consumption curves of the y-th user and the z-th user respectively.

[0216] After secondary clustering using the improved density peak clustering algorithm, the typical electricity consumption curves of all industries in multiple time scales within the time scale T′ are obtained and saved in the curve module in the stereo image library.

[0217] Figure 2This is a graph of typical electricity consumption curves of multiple time scales in various industries on a daily time scale in an embodiment of the method of the present invention, wherein the darker curves are the typical electricity consumption curves of various industries on a daily time scale obtained after secondary clustering, and the lighter curve set is a set of typical electricity consumption curves of all users in the corresponding industry on a daily time scale.

[0218] For each industry, the typical electricity consumption curve of the industry at multiple time scales within the daily time scale is used as the reference sequence, and the mean value of the correlation coefficient between the typical electricity consumption curve of the user at multiple time scales and the reference sequence is calculated as the threshold of the correlation coefficient of the industry curve, as shown in the following formula:

[0219]

[0220] In the formula, ξ ch represents the curve correlation coefficient threshold of the h-th industry, x h0 (i) is the load data at the i-th sampling time point in the typical electricity consumption curve of the h-th industry, x hy (i) is the load data of the ith sampling time point in the daily load sequence of the yth user in the hth industry, ρ is the resolution coefficient, and They represent the minimum difference between two levels and the maximum difference between two levels respectively.

[0221] 3.3) Obtain the multi-dimensional load characteristic index of each user in the training data set, weight it using the user weighting coefficient, calculate the multi-dimensional load characteristic index of each industry, and save it as the industry multi-dimensional load characteristic index sequence to the feature module of the stereoscopic image library, as shown in the following formula:

[0222]

[0223] In the formula, C hg represents the g-th multi-dimensional load characteristic index of the h-th industry, g = 1, 2, 3, 4, 5, corresponding to the five load characteristic indexes of daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and normal period power consumption percentage respectively; C yg Represents the g-th multi-dimensional load characteristic indicator of the y-th user.

[0224] For each industry, the industry multi-dimensional load characteristic index sequence is used as the reference sequence, and the mean value of the correlation coefficient between the user multi-dimensional load characteristic index sequence and the reference sequence is calculated as the industry characteristic correlation coefficient threshold, as shown in the following formula:

[0225]

[0226] In the formula, ξ fh Indicates the characteristic correlation coefficient threshold of the h-th industry.

[0227] 3.4) Add the industry curve correlation coefficient threshold to the industry characteristic correlation coefficient threshold to obtain the industry correlation coefficient threshold, as shown in the following formula:

[0228] ξ h =ξ ch +ξ fh

[0229] In the formula, ξ h Represents the correlation coefficient threshold of the h-th industry.

[0230] 3.5) According to the changes in user load data in different time scales and the electricity consumption characteristics of different industries, set the curve weight matrix θ C and the feature weight matrix θ f , where the curve weight matrix θ C =(θ C1 ,θ C2 ,θ C3 ,θ C4 ) represent the attention weights of the typical electricity consumption curves of each industry in the year, month, day and holiday time scales; the feature weight matrix θ f =(θ f1 ,θ f2 ,θ f3 ,θ f4 ,θ f5 ) respectively represent the attention weights of the five multi-dimensional load characteristic indicators, namely, daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient, and percentage of electricity consumption during normal hours. The two weight matrices are saved in the weight module of the stereo image library. In the embodiment, θ C =(4,3,1,2),θ f =(0.2,0.3,0.4,0.1,0.1).

[0231] 4) Design a load curve-feature dual-channel weighted identification strategy. After performing load curve-feature dual-channel weighting on each industry in the stereo portrait library, calculate the total correlation coefficient between the unknown users in the test data set and each industry in the stereo portrait library, and obtain the identification results of the unknown user load type and abnormal power consumption behavior. The specific steps are as follows:

[0232] 4.1) Obtain the unknown user daily load sequence L from the test dataset N , calculate the multi-dimensional load characteristic index of unknown users and form a multi-dimensional load characteristic index sequence C N , as shown below:

[0233] C N ={C N1 ,C N2 ,C N3 ,C N4 ,C N5}

[0234] In the formula, C N1 , C N2 , C N3 , C N4 and C N5 They respectively represent the daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient, and normal period electricity consumption percentage of unknown users.

[0235] 4.2) Extract the typical electricity consumption curves of all industries in years, months, days and holidays from the stereoscopic image library, as shown in the following formula:

[0236] L Yh ={x Yh (1),x Yh (2),…,x Yh (n)}

[0237] L Mh ={x Mh (1),x Mh (2),…,x Mh (n)}

[0238] L Dh ={x Dh (1),x Dh (2),…,x Dh (n)}

[0239] L Jh ={x Jh (1),x Jh (2),…,x Jh (n)}

[0240] Where, L Yh , L Mh , L Dh and L Jh represents the typical electricity consumption curves of the hth industry in the year, month, day and holiday, x Yh (n), x Mh (n), x Dh (n) and x Jh (n) represents the load data at the nth sampling time point in the typical electricity consumption curve of the year, month, day and holiday of the hth industry.

[0241] Take the unknown user daily load sequence L N The curve correlation coefficient ζ between the reference sequence and the typical electricity consumption curve of the hth industry in the year, month, day and holiday at the i-th sampling time point is calculated. Yh (i), ξ Mh (i), ξ Dh (i) and ξ Jh(i), as shown below:

[0242]

[0243]

[0244]

[0245]

[0246] In the formula, ρ is the resolution coefficient, and Respectively represent the minimum difference between the two levels and the maximum difference between the two levels, x Yh (i), x Mh (i), x Dh (i) and x Jh (i) represents the load data of the typical electricity consumption curve of the hth industry at the i-th sampling time point.

[0247] Combining the curve correlation coefficients of n sampling time points, the curve correlation coefficient R between the reference sequence and the typical electricity consumption curves of the hth industry in the year, month, day and holiday can be obtained. Yh , R Mh , R Dh and R Jh :

[0248]

[0249]

[0250]

[0251]

[0252] Traverse the typical electricity consumption curves of all industries in the stereoscopic image library by year, month, day and holiday, and obtain the 4×H curve correlation coefficient matrix R1:

[0253]

[0254] In the formula, R YH , R MH , R DH and R JH They represent the curve correlation coefficients between the reference sequence and the typical electricity consumption curves of the Hth industry in years, months, days and holidays.

[0255] With the curve weight matrix θ C =(θ C1 ,θ C2 ,θ C3 ,θ C4) to obtain the 1×H weighted curve correlation coefficient matrix R of unknown users and industries. C :

[0256] R c =θ c R1

[0257] 4.3) Extract multi-dimensional load characteristic index sequences of various industries from the stereoscopic image library:

[0258] C h ={C h1 ,C h2 ,C h3 ,C h4 ,C h5}

[0259] In the formula, C h represents the multi-dimensional load characteristic index sequence of the h-th industry, C h1 , C h2 , C h3 , C h4 and C h5 They are the daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and percentage of electricity consumption during normal hours for the hth industry.

[0260] Taking the unknown user multidimensional load characteristic index sequence as the reference sequence, calculate the characteristic correlation coefficient ζ of the g-th multidimensional load characteristic index between the reference sequence and the multidimensional load characteristic index sequence of the h-th industry hg , as shown below:

[0261]

[0262] In the formula, C Ng Represents the g-th multi-dimensional load characteristic indicator of the unknown user.

[0263] Traversing the characteristic sequences of each industry, we obtain the 5×H characteristic correlation coefficient matrix R2:

[0264]

[0265] In the formula, ξ H1 , H2 , H3 , H4 and H5 They are the characteristic correlation coefficients between the reference sequence and the multi-dimensional load characteristic index sequence of the Hth industry in terms of daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and percentage of electricity consumption during normal periods.

[0266] Using the feature weight matrix θ f =(θ f1 ,θf2 ,θ f3 ,θ f4 ,θ f5 ) weights the feature correlation coefficients to obtain a 1×H weighted feature correlation coefficient matrix R f :

[0267] R f =θ f R2

[0268] 4.4) Read the h×1 industry curve weighting coefficient matrix v from the stereo image library C and industry characteristic weighted coefficient matrix v f , for R C and R f Perform dual-channel weighting as shown below:

[0269] v C =(v C1 ,v C2 ,…,v CH ) T

[0270] v f =(v f1 ,ν f2 ,…,ν fH ) T

[0271] R C =v C R c

[0272] R F =v f R f

[0273] In the formula, R C and R F They represent the curve correlation coefficient matrix and feature correlation coefficient matrix after dual-channel weighting respectively.

[0274] 4.5) Using the curve correlation coefficient matrix and feature correlation coefficient matrix after dual-channel weighting, calculate the total correlation coefficient matrix between unknown users and each industry in the stereoscopic portrait library, and take the index value corresponding to the maximum value as the index value corresponding to the user load type, as shown in the following formula:

[0275] R = argmax(diag(R C +R F ))

[0276] Where R is the index value corresponding to the unknown user load type, and the argmax(·) function represents the index value corresponding to the maximum value in the matrix.

[0277] According to the index value corresponding to the user load type, the corresponding industry label is read from the industry label sequence to obtain the user load type identification result.

[0278] 4.6) After obtaining the unknown user load type identification result, the corresponding industry correlation coefficient threshold is obtained from the constructed three-dimensional portrait library. If the correlation coefficient between the unknown user and the industry is less than the corresponding industry correlation coefficient threshold, it is considered that the unknown user's electricity consumption behavior deviates from the normal electricity consumption behavior habits of the corresponding industry, and there may be abnormal electricity consumption behavior. An abnormal alarm signal is output, which is further used to guide the work of troubleshooting electricity hazards.

[0279] In order to verify the effectiveness of the user abnormal load identification method based on the stereo image library and dual-channel weighting of the present invention, the load data of 1500 users in a tourist area were collected as a load data set for experiments. The load data set was divided into a training data set and a test data set in a ratio of 7:3. Abnormal load identification was performed on 450 users in the test data set. The calculation formulas for the load type identification accuracy and the abnormal power consumption behavior identification accuracy are as follows:

[0280]

[0281]

[0282] In the formula, α h and β h They represent the industry identification accuracy and anomaly identification accuracy of the h-th industry, n h and n Y They represent the number of users whose load type and abnormal power consumption behavior of the hth industry are correctly identified, n z Represents the total number of users in the test dataset.

[0283] Table 3 below shows the accuracy of the recognition results of an embodiment of the user abnormal load recognition method based on the stereo image library and dual-channel weighting of the present invention:

[0284] Table 3

[0285]

[0286]

[0287] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for identifying abnormal user load based on a stereo image library and dual-channel weighting, characterized in that: The following steps are involved: 1) Collect users' load data and industry labels, and clean the load data, including outlier removal, missing value filling and standardization. The cleaned load data constitutes a load data set, which is divided into a training data set and a test data set according to the proportion; 2) Construct a user load characteristic analysis module that includes a statistical module and a clustering module, calculate the imbalance degree between different users and different industries in the training data set, cluster and extract the user's multi-time scale typical electricity consumption curve, and then use the user's multi-time scale typical electricity consumption curve to calculate the user's multi-dimensional load characteristic index to obtain the user load characteristic analysis result; 3) Combine the user load characteristic analysis results with industry labels, calculate the user and industry weight coefficients according to the imbalance degree between different users and different industries, and improve the density peak clustering algorithm by weight. Then, use the weighted improved density peak clustering algorithm to perform secondary clustering to obtain the typical power consumption curve of the industry at multiple time scales, and calculate the multi-dimensional load characteristic indicators of the industry. Use the typical power consumption curves of multiple time scales and multi-dimensional load characteristic indicators of each industry to build a three-dimensional portrait library; 4) Design a load curve-feature dual-channel weighted identification strategy. After performing load curve-feature dual-channel weighting on each industry in the stereoscopic portrait library, calculate the total correlation coefficient between the unknown users in the test data set and each industry in the stereoscopic portrait library, and obtain the identification results of the unknown user's load type and abnormal power consumption behavior.

2. The method for identifying abnormal user load based on a stereo image library and dual-channel weighting according to claim 1 is characterized in that: The specific operation steps of step 1) are as follows: 1.1) Obtain the load data of each user in the form of daily load series; 1.2) Traverse the daily load sequence of each user one by one, and judge the missing situation of the user's daily load sequence. If the daily load sequence is missing more than 30% of the sampling data continuously, the daily load sequence will be abandoned; for the retained user daily load sequence, the average value and standard deviation of the daily load sequence are calculated using the following formula: In the formula, is the average value of the daily load sequence, x(i) is the load data at the i-th sampling time point in the daily load sequence, i = 1, 2, …, n, n is the total number of sampling points in the daily load sequence, and s is the standard deviation of the daily load sequence; 1.3) Traverse the load data of each sampling time point in the daily load sequence. If the difference between the load data corresponding to the sampling time point and the average value of the daily load sequence is greater than 3 times the standard deviation of the load sequence of the day, it will be regarded as abnormal data and removed. After removal, the load data corresponding to the sampling time point will be treated as missing values; 1.4) Traverse the user's daily load sequence and determine whether the load data corresponding to the adjacent sampling time points of the missing value exists. If the load data corresponding to the adjacent sampling time points before and after are all missing values, the missing value is assigned to 0; if the sampling time point corresponding to the missing value is the first sampling time point or the last sampling time point in the daily load sequence, the load data corresponding to the only adjacent sampling time point is used to replace the missing value; if the load data corresponding to the adjacent sampling time points are all normal values, the average value of the load data corresponding to the adjacent sampling time points before and after is used to replace the missing value; 1.5) Traverse each sampling time point in the user's daily load sequence and map the corresponding load data to the interval [0,1]. The calculation formula is as follows: Where, L max , L min They represent the maximum and minimum values ​​of the user's daily load sequence, respectively, and x i ′ is the load data at the i-th sampling time point after standardization.

3. The method for identifying abnormal user load based on a stereo image library and dual-channel weighting according to claim 2 is characterized in that: The specific operation steps of step 2) are as follows: 2.1) Construct a user load characteristic analysis module, which consists of a statistical module and a clustering module, which are used to count and divide load data and cluster to obtain typical power consumption curves of users at multiple time scales; 2.2) The statistical module first counts the imbalance between different users and different industries in the training data set. The imbalance between different users includes the number of users included in the user load data and the number of daily load sequences included in each user; the imbalance between different industries includes the number of industry types included in the user load data and the total number of users in each industry; Then, the user industry label data is counted to generate the user industry label sequence H Y , as shown below: H Y ={H1,H2,…,H H } In the formula, H H Represents the industry label corresponding to the Hth industry; Finally, based on the time index in the user load data and combined with the statutory holiday information, the user load data is divided into four time scales: year, month, day and holiday; 2.3) The clustering module first calculates the expected daily load sequence E of the user within the time scale T x 、Entropy E n and super entropy H e , T′∈{year, month, day, holiday}, generating cloud droplets Cloud(E X ,E n ,H e ), which is used to calculate the shape similarity between daily load sequences, as shown in the following formula: Where, L k and L j denote the kth and jth daily load sequences respectively, k = 1, 2, ..., N, j = 1, 2, ..., N, N is the total number of daily load sequences contained in the user time scale T′, D(L k ,L j ) represents the comprehensive similarity between the cloud droplets generated by the kth and jth daily load sequences, En k and En j denote the entropy of cloud droplets generated by the kth and jth daily load sequences, He k and He j denote the super entropy of cloud droplets generated by the kth and jth daily load sequences, S(|E xk -E xj |) represents the overlap area between the expected cloud droplet curves generated by the kth and jth daily load sequences, and S(0) represents the overlap area when the distance between the expected cloud droplet curves generated by the kth and jth daily load sequences is 0; The density peak clustering algorithm is used for clustering. For the user daily load sequences within the same time range, the local density and relative distance of each daily load sequence within the corresponding time range are calculated. The calculation formula is as follows: In the formula, ρ k and δ k denote the local density and relative distance of the kth daily load sequence, ρ j , i denote the local density of the jth and ith daily load series, respectively, d c represents the cutoff distance, χ(·) represents a piecewise function with a value of 0 or 1. If the value in the brackets is negative, the function takes a value of 1; if the value in the brackets is positive, the function takes a value of 0; After obtaining the local density and relative distance of the daily load sequence, the central value and outlier degree of the daily load sequence are calculated. The calculation formula is as follows: l k =ρ k 2 +d k 3 In the formula, λ k and k They represent the central value and outlier degree of the k-th daily load series respectively; Arrange all daily load sequences within the time scale T′ from large to small according to the degree of outliers, and exclude the first q daily load sequences as outlier daily load sequences; then arrange all the remaining daily load sequences from large to small according to the center value, take the first w daily load sequences as the initial clustering centers, create d initial clusters, and assign the remaining daily load sequences to the clusters to which the clustering centers with the smallest relative distances belong in turn to obtain the initial clustering results; Traverse the clusters in the initial clustering results and use the following formula to calculate the distance between clusters: Where D c (C a ,C b ) represents cluster C a and cluster C b The distance between a C and N b C Represent cluster C a and cluster C b The total number of daily load sequences within ; Merge the two clusters with the smallest inter-cluster distance, traverse the remaining d-1 clusters, and recalculate the inter-cluster distance; if all the calculated inter-cluster distances are greater than the set threshold D c,max When , the merging stops and the clustering results are saved; among the k clusters finally saved, the daily load sequence with the smallest distance from other daily load sequences in the cluster is selected and output as the typical power consumption curve of multiple time scales of the user in the time scale T′; 2.4) After obtaining the typical power consumption curve of the user at multiple time scales, calculate the user's multi-dimensional load characteristic index to form a user's multi-dimensional load characteristic index sequence, including daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient, and normal period power consumption percentage. The definition and specific calculation formula of each index are as follows: The daily minimum load rate C1 is the ratio of the minimum value of the daily load sequence to the maximum value of the daily load sequence: The daily average load factor C2 is the ratio of the average value of the daily load sequence to the maximum value of the daily load sequence: Where, L a represents the average value of daily load series; The daily peak-to-valley difference rate C3 is the ratio of the peak-to-valley difference of the daily load sequence to the maximum value of the daily load sequence: The valley power coefficient C4 is the ratio of the power consumption during the valley period to the total power consumption of the daily load sequence: Where, L G Indicates the power consumption during off-peak hours, L T Indicates the total power consumption of the daily load sequence; The normal period electricity consumption percentage C5 is the ratio of normal period electricity consumption to the total electricity consumption of the daily load sequence: Where, L p Indicates the power consumption during normal period.

4. The method for identifying abnormal user load based on a stereo image library and dual-channel weighting according to claim 3 is characterized in that: The specific operation steps of step 3) are as follows: 3.1) Combine the user load characteristic analysis results with the industry label, calculate the user and industry weighting coefficients according to the imbalance degree between different users and different industry types, and save them to the compensation module in the stereo image library; The calculation formula of user weight coefficient is as follows: In the formula, v Yy represents the user weight coefficient of the yth user, Y y represents the number of daily load sequences of the yth user, y = 1, 2, ..., p z , p z represents the total number of users, Y max and Y min They represent the maximum and minimum values ​​of the number of user daily load sequences respectively; The industry weighting coefficient includes the industry curve weighting coefficient and the industry characteristic weighting coefficient. The calculation formula is as follows: In the formula, v Ch and v fh They represent the industry curve weighting coefficient and industry characteristic weighting coefficient of the hth industry, respectively. h represents the total number of users in the hth industry, h = 1, 2, ..., H, H is the number of industry types included in the user load data, p max and p min Respectively represent the maximum and minimum values ​​of the total number of industry users; 3.2) Obtain the typical power consumption curve of multiple time scales of each user in the time scale T′ of the training data set, divide the users with the same industry label into the same range, and perform secondary clustering using the improved density peak clustering algorithm. When calculating the local density and relative distance of the typical power consumption curve of multiple time scales of the user, the industry curve weighting coefficient is used to perform weighted improvement on the density peak clustering algorithm, as shown in the following formula: In the formula, ρ' y and δ' y Respectively represent the local density and relative distance of the typical power consumption curve of the yth user, L' y and L' z They represent the typical electricity consumption curves of the y-th user and the z-th user respectively; After secondary clustering using the improved density peak clustering algorithm, the typical power consumption curves of multiple time scales of all industries within the time scale T′ are obtained and saved in the curve module in the stereoscopic image library. For each industry, the typical power consumption curves of multiple time scales of the industry within the daily time scale are used as the reference sequence. The mean value of the correlation coefficient between the typical power consumption curves of multiple time scales of users and the reference sequence is calculated as the threshold value of the correlation coefficient of the industry curve, as shown in the following formula: In the formula, ξ ch represents the curve correlation coefficient threshold of the h-th industry, x h0 (i) is the load data at the i-th sampling time point in the typical electricity consumption curve of the h-th industry, x hy (i) is the load data of the ith sampling time point in the daily load sequence of the yth user in the hth industry, ρ is the resolution coefficient, and Respectively represent the minimum difference between two levels and the maximum difference between two levels; 3.3) Obtain the multi-dimensional load characteristic index of each user in the training data set, weight it using the user weighting coefficient, calculate the multi-dimensional load characteristic index of each industry, and save it as the industry multi-dimensional load characteristic index sequence to the feature module of the stereoscopic image library, as shown in the following formula: In the formula, C hg represents the g-th multi-dimensional load characteristic index of the h-th industry, g = 1, 2, 3, 4, 5, corresponding to the five load characteristic indexes of daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and normal period power consumption percentage respectively; C yg represents the g-th multi-dimensional load characteristic index of the y-th user; For each industry, the industry multi-dimensional load characteristic index sequence is used as the reference sequence, and the mean value of the correlation coefficient between the user multi-dimensional load characteristic index sequence and the reference sequence is calculated as the industry characteristic correlation coefficient threshold, as shown in the following formula: In the formula, ξ fh Indicates the characteristic correlation coefficient threshold of the h-th industry; 3.4) Add the industry curve correlation coefficient threshold to the industry characteristic correlation coefficient threshold to obtain the industry correlation coefficient threshold, as shown in the following formula: x h =ξ ch +ξ fh In the formula, ξ h represents the correlation coefficient threshold of the h-th industry; 3.5) According to the changes in user load data in different time scales and the electricity consumption characteristics of different industries, set the curve weight matrix θ C and the feature weight matrix θ f , where the curve weight matrix θ C =(θ C1 ,θ C2 ,θ C3 ,θ C4 ) represent the attention weights of the typical electricity consumption curves of each industry in the year, month, day and holiday time scales; the feature weight matrix θ f =(θ f1 ,θ f2 ,θ f3 ,θ f4 ,θ f5 ) respectively represent the attention weights of the five multi-dimensional load characteristic indicators, namely, daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and percentage of electricity consumption during normal hours. The two weight matrices are saved in the weight module in the stereo image library.

5. The method for identifying abnormal user load based on a stereo image library and dual-channel weighting according to claim 4 is characterized in that: The specific operation steps of step 4) are as follows: 4.1) Obtain the unknown user daily load sequence L from the test dataset N , calculate the multi-dimensional load characteristic index of unknown users and form a multi-dimensional load characteristic index sequence C N , as shown below: C N ={C N1 ,C N2 ,C N3 ,C N4 ,C N5 } In the formula, C N1 , C N2 , C N3 , C N4 and C N5 They represent the daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient, and normal-time power consumption percentage of unknown users respectively; 4.2) Extract the typical electricity consumption curves of all industries in years, months, days and holidays from the stereoscopic image library, as shown in the following formula: L Yh ={x Yh (1),x Yh (2),…,x Yh (n)} L Mh ={x Mh (1),x Mh (2),…,x Mh (n)} L Dh ={x Dh (1),x Dh (2),…,x Dh (n)} L Jh ={x Jh (1),x Jh (2),…,x Jh (n)} Where, L Yh , L Mh , L Dh and L Jh represents the typical electricity consumption curves of the hth industry in the year, month, day and holiday, x Yh (n), x Mh (n), x Dh (n) and x Jh (n) represents the load data of the nth sampling time point in the typical electricity consumption curve of the year, month, day and holiday of the hth industry; Take the unknown user daily load sequence L N The curve correlation coefficient ζ between the reference sequence and the typical electricity consumption curve of the hth industry in the year, month, day and holiday at the i-th sampling time point is calculated. Yh (i), ξ Mh (i), ξ Dh (i) and ξ Jh (i), as shown below: In the formula, and Respectively represent the minimum difference between the two levels and the maximum difference between the two levels, x Yh (i), x Mh (i), x Dh (i) and x Jh (i) represents the load data of the typical electricity consumption curve of the hth industry at the i-th sampling time point in the year, month, day and holiday respectively; Combining the curve correlation coefficients of n sampling time points, we can get the curve correlation coefficient R between the reference sequence and the typical electricity consumption curves of the hth industry in the year, month, day and holiday. Yh , R Mh , R Dh and R Jh : Traverse the typical electricity consumption curves of all industries in the stereoscopic image library by year, month, day and holiday, and obtain the 4×H curve correlation coefficient matrix R1: In the formula, R YH , R MH , R DH and R JH They represent the curve correlation coefficients between the reference sequence and the typical electricity consumption curves of the Hth industry in the year, month, day and holiday; With the curve weight matrix θ C =(θ C1 ,θ C2 ,θ C3 ,θ C4 ) to obtain the 1×H weighted curve correlation coefficient matrix R of unknown users and industries. C : R c =θ c R1 4.3) Extract multi-dimensional load characteristic index sequences of various industries from the stereoscopic image library: C h ={C h1 ,C h2 ,C h3 ,C h4 ,C h5 } In the formula, C h represents the multi-dimensional load characteristic index sequence of the h-th industry, C h1 , C h2 , C h3 , C h4 and C h5 They are the daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and percentage of electricity consumption during normal hours for the hth industry; Taking the unknown user multidimensional load characteristic index sequence as the reference sequence, calculate the characteristic correlation coefficient ζ of the g-th multidimensional load characteristic index between the reference sequence and the multidimensional load characteristic index sequence of the h-th industry hg , as shown below: In the formula, C Ng represents the g-th multi-dimensional load characteristic index of the unknown user; Traversing the characteristic sequences of each industry, we obtain the 5×H characteristic correlation coefficient matrix R2: In the formula, ξ H1 , H2 , H3 , H4 and H5 They are the characteristic correlation coefficients between the reference sequence and the multi-dimensional load characteristic index sequence of the Hth industry in terms of daily minimum load rate, daily average load rate, daily peak-to-valley difference rate, valley power coefficient and percentage of electricity consumption during normal periods; Using the feature weight matrix θ f =(θ f1 ,θ f2 ,θ f3 ,θ f4 ,θ f5 ) weights the feature correlation coefficients to obtain a 1×H weighted feature correlation coefficient matrix R f : R f =θ f R2 4.4) Read the h×1 industry curve weighting coefficient matrix v from the stereo image library C and industry characteristic weighted coefficient matrix v f , for R C and R f Perform dual-channel weighting as shown below: v C =(v C1 ,v C2 ,…,v CH ) T v f =(v f1 ,v f2 ,…,v fH ) T R C =v C R c R F =v f R f In the formula, R C and R F They represent the curve correlation coefficient matrix and feature correlation coefficient matrix after dual-channel weighting respectively; 4.5) Using the curve correlation coefficient matrix and feature correlation coefficient matrix after dual-channel weighting, calculate the total correlation coefficient matrix between unknown users and each industry in the stereoscopic portrait library, and take the index value corresponding to the maximum value as the index value corresponding to the user load type, as shown in the following formula: R=argmax(diag(R C +R F )) Where R is the index value corresponding to the unknown user load type, and the argmax(·) function represents the index value corresponding to the maximum value in the matrix; According to the index value corresponding to the user load type, the corresponding industry label is read from the industry label sequence to obtain the user load type identification result; 4.6) After obtaining the unknown user load type identification result, the corresponding industry correlation coefficient threshold is obtained from the constructed three-dimensional portrait library. If the correlation coefficient between the unknown user and the industry is less than the corresponding industry correlation coefficient threshold, it is considered that the unknown user's electricity consumption behavior deviates from the normal electricity consumption behavior habits of the corresponding industry, and there may be abnormal electricity consumption behavior. An abnormal alarm signal is output, which is further used to guide the work of troubleshooting electricity hazards.

Images