Adaptive DBSCAN abnormal battery identification method based on euclidean distance without prior weight

Abstract

The application provides an adaptive DBSCAN abnormal battery identification method based on a Euclidean distance without prior weight, belongs to the technical field of battery anomaly detection of energy storage power stations under data driving, and comprises the following steps: preprocessing original data; taking the preprocessed data as input, obtaining the no-prior weight w of each attribute of the original input data of the battery of the energy storage power station; obtaining the multivariate time series Euclidean distance matrix dist with the no-prior weight w; based on dist, respectively performing MDS low-dimensional embedding and adaptive selection of DBSCAN clustering algorithm parameters; performing DBSCAN clustering abnormal battery identification; and finally outputting a sequence composed of the clustering labels of each battery, wherein the label value of-1 indicates that the battery corresponding to the serial number is an abnormal battery. The application can effectively mine the potential information of the distribution of original battery data, better measure the differences between the input attributes of the multivariate time series, solve the problem of unsatisfactory clustering outlier identification caused by small differences between input parameters, and improve the accuracy of abnormal battery identification of the energy storage power station.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of abnormal battery detection of energy storage power station under data driving, and particularly relates to an adaptive DBSCAN abnormal battery identification method based on a non-prior weight Euclidean distance. BACKGROUND

[0002] With the continuous advancement of energy cleaning and grid intelligence, batteries have developed from single application to group application, and from small-scale single system to large-scale multi-system. As one of the main forms of large-scale application of battery systems, energy storage battery systems play an important role in regulating voltage fluctuations in the power grid. At the same time, as a battery system with high capacity and high power, it can store short-term off-grid power energy on the power generation side, saving resource loss; on the power transmission side, it can buffer voltage and frequency mutations through peak shaving and frequency modulation; on the power consumption side, it can also use peak-valley time-of-use electricity prices and low-storage-high-discharge methods to reduce user electricity costs.

[0003] The energy storage power station thermal management system usually adopts the method of configuring air conditioners in containers, which cannot cover every battery plug-in box, but each battery plug-in box is equipped with multiple temperature sensors for real-time temperature monitoring. Battery production differences, operating environment differences and other reasons will cause inconsistencies in the performance degradation speed and health level of energy storage power station batteries in actual use. The internal and external characteristic parameters of poor performance batteries often deviate from other normal batteries in the group. Through the identification of abnormal batteries and potential abnormal batteries, users can pay more attention to battery abnormalities or potential abnormal information in a timely manner, so as to replace the batteries in time before the battery life reaches the limit.

[0004] The identification of abnormal batteries can be converted into the problem of outlier identification of the DBSCAN algorithm for solution. The conventional DBSCAN clustering algorithm needs to input Eps and MinPts parameters manually, and has the defect of being sensitive to input parameters. Improper selection of parameters may cause clustering errors or even clustering failure, and small differences in parameter input may also lead to low clustering accuracy and unsatisfactory identification of outliers. The running state of energy storage power station batteries may deviate, but the deviation is usually not large, which will further limit the application of the conventional DBSCAN algorithm, and many improvements of the DBSCAN algorithm are focused on improving the clustering effect in some specific application scenarios, which has certain limitations. SUMMARY

[0005] Therefore, the application aims to provide an adaptive DBSCAN abnormal battery identification method based on a priori weight-free Euclidean distance, which can effectively mine the potential information of the distribution of original battery data and better measure the differences between multi-time series input attributes.

[0006] To achieve the above purpose, the technical scheme of the application is as follows: the adaptive DBSCAN abnormal battery identification method based on a priori weight-free Euclidean distance comprises the following steps:

[0007] Step one, pre-process the original data;

[0008] Step two, take the pre-processed data as input to obtain the a priori weight-free w of each attribute of the original input data of the battery of the energy storage power station;

[0009] Step three, obtain the multi-time series Euclidean distance matrix dist with the a priori weight-free w;

[0010] Step four, based on dist, perform MDS low-dimensional embedding and adaptive selection of DBSCAN clustering algorithm parameters respectively;

[0011] Step five, perform DBSCAN clustering abnormal battery identification;

[0012] Step six, the algorithm finally outputs a sequence composed of the clustering labels of each battery, and the label value of-1 indicates that the battery corresponding to the serial number is an abnormal battery.

[0013] Further, the step one comprises supplementing missing data and transforming the dimension and shape of the input data to meet the requirements of a priori weight-free calculation and multi-time series Euclidean distance matrix input.

[0014] Further, the battery data of the energy storage power station is in the form of a 3D tensor of (s, t, f), wherein s is the number of samples, here the number of battery plug-in boxes, t is the time step length, and f is the number of battery attributes;

[0015] Assuming that P is the total sample of battery data, P i is the i-th sample of battery input data, P i =[P i1 ,P i2 ,...P iD ], then each element in P i is a univariate time series of each attribute of the battery, and P i1 is taken as an example,

[0016]

[0017] Where D represents the total number of attributes, tN represents the total number of time steps, and T represents the matrix transpose.

[0018] Furthermore, step two specifically involves: combining the sample and the time step together as the object to be evaluated, that is, transforming the original data size from (s,t,f) to (s*t,f) as the input for this step;

[0019] By using an interval-type indicator U to positively shape the original data, it can be specifically represented as follows:

[0020]

[0021] Among them, P kd For the univariate time series corresponding to the d-th attribute of the k-th sample, the optimal interval for the d-th attribute is set to [a...]. d ,b d ], P' kd This is the sequence after forward orientation;

[0022] Standardize the normalization matrix, denoted as the X matrix; calculate the proportion of the k-th sample under the d-th attribute, and use it as an element y in the probability matrix Y. kd Specifically, it is expressed as follows:

[0023]

[0024] Where K = s*t is the total number of samples to be evaluated, x kd The corresponding element in matrix X;

[0025] The information entropy of the d-th attribute is e d Information utility value f d and entropy weight w d It can be represented as follows:

[0026]

[0027] f d =1-e d

[0028]

[0029] Based on the above calculations, the weight vector without prior knowledge can be obtained as follows:

[0030] w = {w1, w2, ... w} d ,...,w D}

[0031] Further, the step three is specifically: the distance matrix is composed of the "distance" between each battery plug-in box and all other battery plug-in boxes, in order to more clearly describe the distance between different samples and give a general solution method of multivariate time series distance matrix, different symbols P and Q are used to represent the data set here;

[0032] The corresponding dth attribute distance between different samples of the multivariate time series can be expressed as follows:

[0033]

[0034] Wherein, P id is the dth attribute of the ith sample in the data set, Q jd is the dth attribute of the jth sample in the data set, is each element inside the univariate time series of the dth attribute of the ith sample, is each element inside the univariate time series of the dth attribute of the jth sample;

[0035] For solving the Euclidean distance matrix without prior weight, from a point to a row and then to the whole matrix, it is expressed as follows:

[0036]

[0037] Wherein, I and J are the total number of samples of the data sets P and Q respectively, and I = J in the current problem.

[0038] Further, the dimensionality of dist is reduced by MDS low-dimensional embedding to improve the sample density, and the low-dimensional inner product matrix B is solved, as shown in the following formula:

[0039]

[0040] Wherein, and are the row mean, column mean and matrix mean of the square of the original distance matrix respectively; b ij is each element of the matrix B;

[0041] Eigenvalue decomposition is performed on B, and the diagonal matrix ∧ composed of the first D eigenvalues and the corresponding eigenvector matrix dist is restored, and the low-dimensional equivalent matrix D' is calculated as follows:

[0042]

[0043] Further, the adaptive selection of DBSCAN clustering algorithm parameters includes two aspects, which are the Eps and MinPts estimation interval determination based on kernel density estimation and the final selection of parameters based on the consistency and stability of the number of outliers.

[0044] Further, the kernel density estimation determines the parameter range according to the data distribution characteristics; the Gaussian kernel is selected as the kernel function, and the optimal bandwidth is selected through the grid search, and the kernel density estimation based on the Gaussian kernel function can be expressed as follows:

[0045]

[0046] wherein n is the sample capacity, h is the bandwidth, exp is the natural constant e,

[0047] (x1,x2,...,x n ) is the inter-bay distance value calculated by the population of random variables, and x is the current input distance value;

[0048] The original distance matrix dist is converted into a vector in the unit of the bay as the input of the above formula, and the kernel density curve estimation can be obtained; it is generally considered that the first peak value peak1 is the intra-cluster distance of the energy storage battery, and the second peak value peak2 is the inter-cluster distance;

[0049] For the DBSCAN clustering algorithm, the estimation range of the algorithm Eps parameter can be set based on peak1, that is,

[0050] Eps∈(peak1-Δ1,peak1+Δ2)

[0051] wherein Δ2 and Δ1 are the upper and lower fluctuation amounts of peak1, which can be flexibly selected in actual application, but should meet the following constraints:

[0052]

[0053] Based on the estimation range of the Eps parameter and the dist obtained above, the estimation of all bay objects MinPts is obtained:

[0054]

[0055] wherein num j represents the number of bays contained in the Eps neighborhood of bay j;

[0056] After removing the duplicates and sorting the above estimation, the optional value set {MinP1, MinP2,..., MinP mp} composed of mp MinPts parameters can be obtained.

[0057] Further, after determining the estimation intervals of Eps and MinPts, the final parameter values of Eps and MinPts are adaptively selected based on the outlier quantity consistency degree and stability degree indexes, and the specific implementation is as follows:

[0058] Assume the current energy storage power station has J battery cells, and the possible outlier cases are defined as s∈(0,1,2,...,J), with a total of J+1 cases; the discrete level of the Eps parameter is set to L, that is, the interval (peak1-Δ1,peak1+Δ2) is uniformly discretized into L points, and the discrete unit is defined. Then the discrete set of possible points e for the Eps parameter is {peak1-Δ1, peak1-Δ1+u,...,peak1+Δ2};

[0059] Each point in the set s is considered as an actual outlier for each iteration. During this process, all combinations of the estimated intervals of Eps and MinPts are iterated, which means a total of (J+1)*mp*L experiments are required.

[0060] The experiment first records the cases where outliers *s* in the cluster match the actual outliers. Given these matches, it further records the number of experiments (l) in L trials for each of the *mp* optional *MinPts* parameters that satisfy the condition of matching cluster outliers with actual outliers. i Where (i = 1, 2, ..., mp), l is selected. i MinP corresponding to the maximum value i As the final parameter value of MinPts;

[0061] After selecting MinP i Given that each point in the input set e is used as an experimental parameter, the set consisting of the corresponding DBSCAN algorithm input parameters (MinPts, Eps) can be represented as {(MinPts, Eps)} i ,peak1-Δ1),(MinP i ,peak1-Δ1+u),...(MinP i ,peak1+Δ2)};

[0062] We sequentially count the number of outliers in the DBSCAN clustering under the above L parameter combinations, denoted as set s', where s' = {s'1, s'2, ..., s'} L Based on the outlier stability index, the final parameter value of Eps is selected, which is the largest continuous repetition interval in the statistical set s', denoted as (s'). seq_d ,s' seq_u Simultaneously, record the positions of the upper and lower bounds of the interval in the set s', denoted as seq_d and seq_u, where seq_d, seq_u ∈ (1, 2, ..., L). Then, the final selected Eps i It can be represented as follows:

[0063]

[0064] Further, the low-dimensional equivalent matrix D' obtained by the MDS low-dimensional embedding is taken as the battery feature data finally input into the algorithm, and (MinP i , Eps i ) is taken as the parameter combination finally selected by the algorithm, and then the DBSCAN clustering is performed to identify the abnormal battery, and the sequence composed of the clustering labels of the batteries is finally output, and the label value of-1 indicates that the battery corresponding to the serial number is an abnormal battery.

[0065] Compared with the prior art, the adaptive DBSCAN abnormal battery identification method based on the Euclidean distance without prior weight has the following advantages:

[0066] (1) The adaptive weight update based on the entropy weight method can effectively mine the potential information of the original battery data distribution, and better measure the differences between the multi-time series input attributes;

[0067] (2) The adaptive DBSCAN algorithm based on the kernel density estimation and the consistency and stability of the outlier number can solve the problem of unsatisfactory clustering of outliers caused by small differences in input parameters, and improve the accuracy of the abnormal battery identification of the energy storage power station. BRIEF DESCRIPTION OF DRAWINGS

[0068] The accompanying drawings, which form a part of the disclosure, are intended to provide further understanding of the present disclosure, and the schematic embodiments of the present disclosure and the description thereof are used to explain the present disclosure, and do not constitute improper limitations on the present disclosure. In the drawings:

[0069] Figure 1 is the algorithm flowchart of the adaptive DBSCAN abnormal battery identification method based on the Euclidean distance without prior weight provided by the present disclosure. DETAILED DESCRIPTION

[0070] It should be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0071] The present disclosure provides an adaptive DBSCAN abnormal battery identification method based on the Euclidean distance without prior weight. The adaptive DBSCAN abnormal battery identification method based on the Euclidean distance without prior weight comprises solving the Euclidean distance matrix without prior weight of the multi-time series data of the battery of the energy storage power station based on the entropy weight method, and performing low-dimensional embedding on the distance matrix to improve the sample density, and then taking the distance matrix as the input of the clustering algorithm; the reasonable interval of the Eps and MinPts parameters in the DBSCAN algorithm is determined by using the kernel density estimation based on the distribution characteristics of the battery data, and finally the parameters are adaptively selected in the interval based on the consistency and stability of the outlier number.

[0072] The application aims at the data characteristics of the battery group and the problem of low clustering accuracy caused by unreasonable artificial selection of Eps and MinPts parameters of the traditional DBSCAN algorithm, and provides an adaptive DBSCAN abnormal battery identification method based on a priori weight-free Euclidean distance.

[0073] The application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0074] As shown in the drawings, Figure 1 The application is an adaptive DBSCAN abnormal battery identification method based on a priori weight-free Euclidean distance, which comprises the following steps:

[0075] Step one, pre-process the original data, supplement the missing data, and transform the dimension and shape of the input data to meet the requirements of a priori weight-free calculation and multivariate time series Euclidean distance matrix input;

[0076] When the method is applied, the battery data of the energy storage power station is in the form of 3D tensor (s, t, f) as the input of distance matrix calculation, wherein s is the number of samples, here is the number of battery plug-in boxes, t is the time step length, and f is the number of battery attributes;

[0077] Suppose P is the total sample of battery data, P i is the i-th sample of battery input data, P i =[P i1 ,P i2 ,...P iD ], then each element in P i is a univariate time series of each attribute of the battery, and P i1 is taken as an example,

[0078]

[0079] Wherein D represents the total number of attributes, tN represents the total number of time steps, and T represents the matrix transpose.

[0080] Step two, taking the pretreated data as input, the original input data of the energy storage power station battery attributes of the priori weight w is calculated, specifically: the update of the priori weight based on the entropy weight method is applied to each attribute of the original input data of the energy storage power station battery, therefore, the sample and time step are integrated together as the object to be evaluated, that is, the size of the original data is changed from (s, t, f) to (s*t, f) as the input of this step;

[0081] By using the interval type index U to normalize the original data, it can be expressed as follows:

[0082]

[0083] Where, P kd is the one-dimensional time series corresponding to the dth attribute of the kth sample, and the best interval of the dth attribute is set as [a d ,b d ], P' kd is the sequence after normalization;

[0084] The normalized matrix is denoted as X matrix; the proportion of the kth sample under the dth attribute is calculated as the element y kd in the probability matrix Y, which is expressed as follows:

[0085]

[0086] Where, K=s*t is the total number of samples to be evaluated, and x kd is the corresponding position element of the X matrix;

[0087] The information entropy e d , information utility value f d and entropy weight w d of the dth attribute can be expressed as follows:

[0088]

[0089] f d =1-e d

[0090]

[0091] After the above calculation, the priori weight vector w={w1,w2,...w d ,...,w D} can be obtained.

[0092] Step 3: Calculate the multivariate time series Euclidean distance matrix dist with no prior weights w. Specifically, the distance matrix is ​​composed of the "distance" between each battery box and all other battery boxes (including itself). That is, P and Q are essentially the same input. However, to more clearly describe the distance between different samples and to give a general solution method for the multivariate time series distance matrix, different symbols P and Q are used here to represent the dataset.

[0093] The distance between different samples of a multivariate time series corresponding to the d-th attribute can be represented as follows:

[0094]

[0095] Among them, P id Let Q be the d-th attribute of the i-th sample in the dataset. jd Let d be the attribute of the j-th sample in the dataset. Let each element in the univariate time series of the i-th sample and its d-th attribute be an element within the series. For each element within the univariate time series of the j-th sample and the d-th attribute;

[0096] For solving the Euclidean distance matrix without prior weights, the generalization from a point in the distance matrix to a row, and then to the entire matrix, can be represented as follows:

[0097]

[0098] Where I and J are the total number of samples in datasets P and Q, respectively, and in the current problem, I = J.

[0099] Step four: Based on dist, perform MDS low-dimensional embedding (obtaining the low-dimensional equivalent matrix D') and adaptive selection of DBSCAN clustering algorithm parameters (MinP). i , Eps i );

[0100] Specifically, the steps to obtain the low-dimensional equivalent matrix D' are as follows: The dist is reduced in dimensionality through MDS low-dimensional embedding to increase sample density, and the low-dimensional inner product matrix B is solved as follows:

[0101]

[0102] in, and These are the row mean, column mean, and matrix mean of the squared original distance matrix, respectively; b ij Let B be the elements of matrix B;

[0103] Eigenvalue decomposition of B can be performed by selecting the diagonal matrix ∧ formed by the first D eigenvalues ​​and the corresponding eigenvector matrix. The dist is reduced to obtain a low-dimensional equivalent matrix D' as follows:

[0104]

[0105] The adaptive selection of the DBSCAN clustering algorithm parameters includes Eps and MinPts estimation interval determination based on kernel density estimation and parameter final selection based on the consistency of the number of outliers and stability.

[0106] The kernel density estimation determines the parameter range according to the data distribution characteristics, which can effectively determine the parameters in a relatively reasonable range; the selection of the bandwidth has a greater impact on the kernel density estimation result than the kernel density function, a common Gaussian kernel is selected as the kernel function, and the best bandwidth is selected through grid search, and the kernel density estimation based on the Gaussian kernel function can be expressed as follows:

[0107]

[0108] Where n is the sample size, h is the bandwidth, exp is the natural constant e,

[0109] (x1,x2,...,x n ) is the inter-cabinet distance value calculated for the population of random variables, x is the current input distance value;

[0110] The original distance matrix dist is converted into a vector in the unit of cabinet as input to the above formula to obtain the kernel density curve estimation; it is generally considered that the first peak peak1 is the intra-cluster distance of the energy storage battery, and the second peak peak2 is the inter-cluster distance;

[0111] For the DBSCAN clustering algorithm, the estimation range of the algorithm Eps parameter can be set based on peak1, that is, Eps∈(peak1-Δ1, peak1+Δ2), where Δ2 and Δ1 are the upper and lower fluctuations of peak1, which can be flexibly selected in actual application, but should satisfy the following constraints:

[0112]

[0113] Based on the estimation range of the Eps parameter and the dist obtained above, the estimation of all cabinet objects MinPts is obtained:

[0114]

[0115] Where num j represents the number of cabinets included in the Eps neighborhood of cabinet j; the above estimation is removed and sorted to obtain a selectable value set {MinP1, MinP2,..., MinPmp}。

[0116] After determining the estimation interval of Eps and MinPts, the final parameter values of Eps and MinPts are adaptively selected based on the outlier quantity coincidence degree and stability index. The specific implementation is as follows:

[0117] Assume that the current energy storage power station battery has J plug-in boxes, set the possible situation of outliers s e (0, 1, 2,..., J), a total of J+1 situations; set the discrete level of Eps parameter as L, that is, uniformly disperse the interval (peak1-Δ1, peak1+Δ2) into L points, and set the discrete unit Then the Eps parameter discrete point set e is {peak1-Δ1, peak1-Δ1+u,..., peak1+Δ2};

[0118] In view of the fact that the abnormal degree of the energy storage power station battery cannot be known in actual engineering, and the purpose of DBSCAN clustering should be to ensure that the clustered outliers coincide with the actual outliers as much as possible, each point in the s set is regarded as the actual outlier each time, and all combinations of the estimation interval of Eps and MinPts are traversed each time, that is, a total of (J+1)*mp*L experiments are required;

[0119] The experiment first records the situation of outliers s that coincide with the clustered outliers and the actual outliers, and under the premise of coincidence, further records the number of experiments l that meet the coincidence condition of the clustered outliers and the actual outliers in the L experiments of the mp selectable MinPts parameters i , (i=1, 2,..., mp), and selects l i with the maximum value as the final parameter value of MinPts; i .

[0120] Under the premise of selecting MinP i , traverse each point in the input set e as an experimental parameter, and the set formed by the input parameters (MinPts, Eps) of the DBSCAN algorithm can be represented as {(MinP i , peak1-Δ1), (MinP i , peak1-Δ1+u),..., (MinP i , peak1+Δ2)};

[0121] The number of DBSCAN clustered outliers under the above L parameter combinations is sequentially counted and recorded as a set s', s'={s'1, s'2,..., s'J+1}. L},based on the outlier stability index to select the final parameter value of Eps, that is, the maximum continuous repeated interval in the statistical set s', denoted as (s' seq_d ,s' seq_u ), while recording the upper and lower bounds of the interval in the set s', denoted as seq_d and seq_u, wherein seq_d, seq_u (1, 2,..., L), then the final selected Eps i can be expressed as follows:

[0122]

[0123] Step five, taking D' as the final input battery feature data of the algorithm, taking (MinP i , Eps i ) as the final selected parameter combination of the algorithm, and then performing DBSCAN clustering to identify abnormal batteries;

[0124] Step six, the final output of the algorithm is a sequence composed of battery cluster labels, and a label value of -1 indicates that the corresponding serial number battery is an abnormal battery.

[0125] Taking the low-dimensional equivalent matrix D' obtained by MDS low-dimensional embedding as the final input battery feature data of the algorithm, taking (MinP i , Eps i ) as the final selected parameter combination of the algorithm, and then performing DBSCAN clustering to identify abnormal batteries, the final output of the algorithm is a sequence composed of battery cluster labels, and a label value of -1 indicates that the corresponding serial number battery is an abnormal battery.

[0126] In the above process of identifying abnormal batteries, the selection of algorithm parameters is completely based on the actual operation data of energy storage station batteries, and related information such as the potential distribution characteristics of data is mined through kernel density estimation and non-prior weight Euclidean distance, and the parameters are adaptively selected under the premise of meeting the consistency and stability of the number of outliers. Since it is a data-driven adaptive clustering to identify abnormal batteries, in actual application, it is not necessary to pay attention to the differences in battery topology and capacity size of different energy storage stations, and it is not necessary to pay attention to the size of battery attribute dimension (voltage, temperature, etc.), and the method has a certain generality.

[0127] The above only describes the preferred embodiments of the present application and should not be used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An adaptive DBSCAN abnormal battery identification method based on a pre- weight-free Euclidean distance, characterized in that, Comprising the following steps: Step one, pre-processing the original data; Step two, taking the pre-processed data as input, and calculating the no-priori weight w of each attribute of the battery original input data of the energy storage power station; Integrate the sample and time step together as the object to be evaluated, that is, change the size of the original data from (s, t, f) to (s*t, f) as the input of this step; Where s is the number of samples, here is the number of battery plug-in boxes, t is the time step length, and f is the number of battery attributes; By using the interval type index U to normalize the original data, which can be specifically expressed as follows: wherein P kd is the dth attribute of the kth sample, and the optimal interval of the dth attribute is set as [a d ,b d ], P ' kd is the sequence after being normalized. The normalized forward matrix is denoted as X matrix; the proportion of the kth sample under the dth attribute is calculated as an element y in the probability matrix Y kd , which is specifically represented as follows: Where K = s*t is the total number of samples to be evaluated, x kd The corresponding element in matrix X; the information entropy e of the d-th attribute. d Information utility value f d and entropy weight w d It can be represented as follows: f d = 1 - e d After the above calculation, the no-priori weight vector is as follows: w = {w1, w2,... w d ,...,w D}; Step three, calculate the multivariate time series Euclidean distance matrix dist with no-priori weight w; Step four, based on dist, respectively perform MDS low-dimensional embedding and adaptive selection of DBSCAN clustering algorithm parameters; Among them, the adaptive selection of DBSCAN clustering algorithm parameters includes two aspects, which are respectively the Eps and MinPts estimation interval determination based on kernel density estimation and the final selection of parameters based on the conformity degree and stability of outliers; The kernel density estimation determines the parameter range according to the data distribution characteristics; The Gaussian kernel is selected as the kernel function, and the best bandwidth is selected by grid search. The kernel density estimation based on Gaussian kernel function can be expressed as follows: wherein n is the sample size, h is the bandwidth, exp is the natural constant e, (x1, x2,..., x n ) is the inter-cassette distance value calculated from the population of random variable pairs, and x is the current input distance value. Convert the original distance matrix dist into a vector in the unit of plug-in box as the input of the above formula, and the kernel density curve estimation can be obtained; The first peak peak1 is the intra-cluster distance of the battery of the energy storage power station, and the second peak peak2 is the inter-cluster distance; For the DBSCAN clustering algorithm, set the estimation range of the Eps parameter based on peak1, that is Eps∈(peak1-△1,peak1+△2) Where △2 and △1 are the upper and lower fluctuation of peak1, which can be flexibly selected in practical application, but should meet the following constraints: Based on the estimation range of Eps parameter and the calculated dist, the estimation of MinPts of all plug-in box objects is obtained: num j represents the number of the patch boxes contained in the Eps neighborhood of the patch box j; The above estimated values are de-duplicated and sorted, and an optional value set {MinP1, MinP2,..., MinPmp} of MinPts parameters can be obtained mp} After determining the estimation interval of Eps and MinPts, the final parameter values of Eps and MinPts are adaptively selected based on the outlier quantity conformity degree and stability index, which can be specifically implemented as follows: Assume that the current energy storage power station battery has J plug-in boxes, and set the possible situation of outliers s∈ (0, 1, 2,..., J), a total of J+1 cases; set the discrete level of the Eps parameter as L, that is, uniformly disperse the (peak1-△1, peak1+△2) interval into L points, set the discrete unit Then the Eps parameter discrete selectable point set e is {peak1-△1, peak1-△1+u,..., peak1+△2}. Traverse each point in the s set as the actual outlier each time, and traverse all combinations of the estimation interval of Eps and MinPts during each time, that is, a total of (J+1)*mp*L experiments are required; The experiment first records the case that the clustering outliers exist and are consistent with the actual outliers s, and under the premise of existing consistency, further records the number of experiments l in which the clustering outliers meet the condition of being consistent with the actual outliers in each of the l experiments of the mp selectable MinPts parameters i where (i = 1, 2, …, mp), and l is selected i The maximum value corresponds to MinP i as the final parameter value of MinPts; After MinP i is selected, the set of input parameters of DBSCAN algorithm can be represented as { (MinP i , peak1-△1), (MinP i , peak1-△1+u),..., (MinP i , peak1+△2)} by traversing each point in the input set e as the experimental parameter. Count the number of DBSCAN clustering outliers under the above L parameter combinations in sequence, denoted as set s ' , ' s ' ={s ' 1,s ' 2,...,s L} ' Based on the outlier stability index, the final Eps parameter value is selected, that is, the maximum continuous repeated interval in the statistical set s ' seq_d , ' seq_u The upper and lower boundaries of the interval are recorded in the set s ' , denoted as seq_d and seq_u, where seq_d, seq_u∈(1,2...,L), and the final selected Eps i can be expressed as follows: Step five, DBSCAN clustering abnormal battery identification; Step six, the final output of the algorithm is a sequence composed of the clustering labels of each battery, and the label value-1 indicates that the corresponding serial number battery is an abnormal battery. 2.The adaptive DBSCAN-based outlier battery identification method based on the Euclidean distance without prior weights according to claim 1, wherein, The step one includes supplementing missing data and transforming the dimension and shape of the input data to meet the requirements of no-priori weight calculation and multivariate time series Euclidean distance matrix input. 3.The adaptive DBSCAN-based outlier battery identification method based on the Euclidean distance without prior weights according to claim 2, characterized in that: The energy storage power station battery data is in the form of a 3D tensor of (s, t, f) as the input of distance matrix calculation, where s is the number of samples, here is the number of battery plug-in boxes, t is the time step length, and f is the number of battery attributes; Assume P is the total sample of battery data, P i is the i-th sample of battery input data, P i = [P i1 , P i2 ,... P iD ], then P i is a univariate time series of each attribute of the battery, and P i1 is an example, Where D represents the total number of attributes, tN represents the total number of time steps, and T represents the matrix transpose. 4.The adaptive DBSCAN-based outlier battery identification method based on the Euclidean distance without prior weights according to claim 1, wherein, The third step is specifically: the distance matrix is composed of the distance between each battery plug-in box and all other battery plug-in boxes, in order to more clearly describe the distance between different samples and give a general solving method of multivariate time series distance matrix, here different symbols P and Q are used to represent the data set; The corresponding dth attribute distance between different samples of the multivariate time series can be expressed as follows: where P id is the dth attribute of the ith sample in the dataset, Q jd is the dth attribute of the jth sample in the dataset, is each element within the unary time series of the dth attribute of the ith sample, is each element within the unary time series of the dth attribute of the jth sample; For solving the Euclidean distance matrix without prior weight, from a point to a row and then to the whole matrix in the distance matrix, it is expressed as follows: Where I and J are the total number of samples of data sets P and Q respectively, and in the current problem I=J. 5.The adaptive DBSCAN-based outlier battery identification method based on the Euclidean distance without prior weights according to claim 1, wherein: The dimensionality of dist is reduced by MDS low-dimensional embedding to improve the sample density, and the low-dimensional inner product matrix B is solved, as shown in the following formula: wherein, and are the row mean, column mean and matrix mean of the original distance matrix squared, respectively; b ij are the elements of matrix B. Eigenvalue decomposition is performed on B to select the first D eigenvalues to form a diagonal matrix Λ and the corresponding eigenvector matrix V Reduction is performed on dist to find the low-dimensional equivalent matrix D ' : 6.The adaptive DBSCAN-based outlier battery identification method based on the Euclidean distance without prior weights according to claim 1, wherein: The low-dimensional equivalent matrix D obtained by the MDS low-dimensional embedding ' The battery feature data as the final input of the algorithm, take (MinP i , Eps i ) as the final selected parameter combination of the algorithm, then perform DBSCAN clustering to identify abnormal batteries, and the final output of the algorithm is a sequence composed of the clustering labels of each battery. If the label value is -1, it means that the battery corresponding to the serial number is an abnormal battery.

Images