CN113537525B - A fault state adaptive early warning method for battery energy storage system - Google Patents

Abstract

The invention discloses a fault state self-adaptive early warning method of a battery energy storage system. The NJW clustering algorithm is used to reduce the dimension of high-dimensional monitoring data. The only alternative to the original data for cluster analysis, solves the singularity problem that often occurs in traditional methods for non-convex data clustering; uses DTW to dynamically adjust the time axis of asynchronous monitoring data, and maps the two sets of monitoring data to the synchronous time axis to overcome the problem. The observation error caused by asynchronous sampling of monitoring data is solved, and the problem that the time axis of heterologous data of sampling points cannot be one-to-one corresponding to the traditional method is solved. Finally, a sliding window model is constructed to suppress the influence of a small number of outliers in the monitoring data, and clustering is carried out based on the DTM distance. Class analysis, through sparse coefficient LSR and fault threshold FT, objectively select fault clustering points, avoid over-estimation or under-estimation of fault state by traditional fault clustering method, and realize BESS fault state adaptive early warning.

Description

A fault state adaptive early warning method for battery energy storage system

technical field

The invention relates to the technical field of battery energy storage system detection, in particular to a fault state adaptive early warning method for a battery energy storage system.

Background technique

Most of the safety problems encountered by the battery energy storage system (BESS) come from the cell level, mainly including overcharge, overdischarge, internal short circuit and external short circuit. The parameter estimation method realized by establishing an accurate electrothermal simulation model or the threshold limit method realized based on the empirical estimation method only identifies abnormal data during the operation of BESS, and relies heavily on normal data for reference correction, while ignoring normal batteries Observational error and process noise during the cycle. However, large-capacity battery energy storage devices are generally composed of hundreds or thousands of battery cells in series and parallel, and there are dozens of monitoring units working with battery boxes. Different monitoring units perform asynchronous measurements on different battery boxes. There are often large observation errors and process noises between the heterogeneous data, and the sampling time axis and battery cycle process recorded by different monitoring units are also different, so it is difficult to compare the data. These factors affect the BESS fault identification accuracy. The impact cannot be ignored.

Large-capacity BESS usually divides batteries, and coordinates the battery boxes of each partition to adopt different charging and discharging strategies according to different needs. For example, some battery boxes are deeply charged and deeply discharged to meet the peak load and valley filling needs of the power plant. At the same time, the rest of the battery boxes may be It works in the form of shallow charging and shallow discharging to suppress user harmonics and solve power quality problems; the same battery box also has completely different working forms in different stages of charging and discharging, and trickle charging may be used in the early and later stages of charging to protect battery safety, while In the mid-term, high-rate fast charging is used, and the voltage and current curves in different periods are very different. In order to achieve sufficient identification accuracy, the fault identification method based on the measured data of the monitoring unit must jump out of the laboratory environment and try to solve the problem of asynchronous data mismatch between different battery compartments in different cycle stages in actual work.

First of all, when the existing method is used for early warning of BESS fault state, the heterologous data observation error and process noise of different battery box monitoring units are generally ignored directly or processed according to homologous noise. There is no literature on the matching method of heterologous data. Research in depth. The actual BMS (Battery Management System) monitoring data and theoretical research show that the time axis of the data sequence recorded by different monitoring units is often asynchronous, and it is difficult to compare directly. If the data from different sources does not match, the fault warning will likely be misjudged.

Secondly, in the prior art, when the fault identification is realized based on the measured data of the monitoring unit, the charging and discharging strategy and the battery cycle form are usually directly defined according to the battery type. The large-capacity BESS used in the industrial process generally adopts different charging and discharging strategies in the battery boxes of each partition, and the same battery box also has completely different working forms in different stages of charging and discharging. If the monitoring data is not extracted and the time axis is regularized, and the asynchronous data mismatch of the battery boxes of different battery partitions in different cycle stages is not matched in actual work, the fault state early warning will easily be underestimated or overestimated to varying degrees, making The BESS fault state warning cannot reflect the real safety state of the energy storage equipment.

In addition, when identifying abnormal data through fault clustering in existing data, the distance method is often used to directly determine the fault data. However, the clustering distance threshold of fault points is generally given by empirical estimation, and there is a lot of chance. If the set is too large or too small, it is easy to cause false alarms of fault warning.

Terminology Explanation:

Battery energy storage system: an energy storage system that uses batteries as an energy storage carrier and is composed of hundreds or thousands of battery cells to store and supply electrical energy, generally including battery cabinets that play the role of energy storage and monitoring and regulation. two parts of the control cabinet.

Battery management system: The battery management system is a system that monitors and controls the safety and operation status of energy storage batteries. The battery management system saves, processes and feeds back the monitored battery information to users in real time, and adjusts various parameters according to the collected information. , to protect the reliable and stable operation of the battery.

NJW spectral clustering algorithm: a spectral clustering algorithm that obtains the corresponding Laplacian matrix by monitoring the similarity matrix of the data, selects the eigenvectors corresponding to the first several largest eigenvalues as the one-to-one correspondence replacement matrix of the original data, and then calculates the corresponding Laplacian matrix. It is clustered by row.

Dynamic time warping (DTW): an algorithm that compares the similarity of asynchronous time series by bending the time axis of monitoring data, and dynamically warps the asynchronous time series.

SUMMARY OF THE INVENTION

In view of the above problems, the purpose of the present invention is to provide an adaptive early warning method for the fault state of a battery energy storage system, processing heterogeneous data based on the NJW spectrum clustering algorithm, and classifying the battery boxes that can be compared with the fault state under similar working conditions. Class, use dynamic warping algorithm to solve the problem of asynchronous data timeline inconsistency, build a sliding window model to suppress the influence of a small number of outliers in monitoring data, perform cluster analysis based on DTW distance, and objectively select fault clusters through sparse coefficient and fault threshold point to realize the self-adaptive early warning of BESS fault status. The technical solution is as follows:

A fault state adaptive early warning method for a battery energy storage system, comprising the following steps:

Step 1: Use the NJW clustering algorithm to reduce the dimension of the high-dimensional battery box monitoring data, obtain its safety feature standard matrix by constructing a pull matrix of heterogeneous data, and then use the safety feature standard matrix to uniquely replace the battery box target monitoring parameters Cluster analysis of raw data;

Step 2: Regularize the time axis of the security feature standard matrix through the dynamic time warping algorithm, and map the two sets of monitoring data to the synchronous time axis to compare the similarity of the asynchronous monitoring data;

Step 3: Build a sliding window model to suppress the influence of a small number of outliers in the monitoring data, perform cluster analysis based on the DTW distance, objectively select the fault clustering points through the sparse coefficient LSR and the fault threshold FT, and then determine the faulty battery box.

Further, the step 1 specifically includes the following steps:

Step 1.1: Set the target monitoring parameter of the battery management system to monitor the battery box as V, and obtain m groups of battery box monitoring data, each group of data contains n sampling points, and the monitoring data to be clustered is expressed as:

Step 1.2: Extract the battery box monitoring data and construct a similarity matrix W={w _ij ∣i≤m,j≤n}∈V ^n×n as follows:

Among them, u _i and u _j represent two heterogeneous battery box target monitoring data; σ _i and σ _j are adaptive identification parameters, σ _i is the r with the smallest Euclidean distance between the battery box target monitoring parameter _ui and the rest of the monitoring data The average value of the sampling points, σ _j is the average value of the r sampling points with the smallest Euclidean distance between the target monitoring parameter u _j of the battery box and the rest of the monitoring data;

Step 1.3: Calculate the metric matrix D={d _ij ∣i≤m,j≤n}∈V ^n×n according to the similarity matrix W:

Step 1.4: Calculate the Laplace matrix L from the similarity matrix W and the metric matrix D:

Step 1.5: Calculate the eigenvalues and eigenvectors corresponding to the pull matrix L, arrange the eigenvalues and eigenvectors in descending order, and take the first K eigenvectors to form a security feature matrix S=[s ₁ ,s ₂ ,...,s _K ]∈ In V ^n×K , the security feature matrix S is normalized row by row to form the security feature standard matrix Y={y _ij ∣i≤m,j≤n}∈V ^n×K :

In the formula, s _ij is the element of the ith row and j column of the S matrix;

Step 1.6: Each row of the safety feature standard matrix Y corresponds to a battery box target monitoring parameter sequence, which uniquely replaces the original sampling data.

Further, the step 2 specifically includes the following steps:

Step 2.1: In the safety feature standard matrix Y, the monitoring data feature vectors corresponding to the two asynchronous battery boxes a and b are Y _a ={y _a1 ,y _a2 ,...,y _an } and Y _b ={y _b1 , _{y b2} , . _ij ∣i≤n,j≤n}:

MIN=min{r _i-1,j ,r _i,j-1 ,r _i-1,j-1 }

In the formula, d(y _ai , y _bj ) is the Euclidean distance between the sample points y _ai and y _bj ; d(y _ai , y _bj )+MIN is the sum of the minimum Euclidean distances between the current sample point and the neighboring sample points;

Step 2.2: After forming the distance matrix, the DTW distances of battery boxes a and b are:

DTW(Y _a ,Y _b )=r _nn a,b≤m

In the same way, the DTW distance measured by any two battery boxes according to the monitoring parameter V can be calculated. When a=b, the DTW distance is zero.

Further, the step 3 specifically includes the following steps:

Step 3.1: Place a sliding window of length d, d<<n, at the starting position t ₁ of the monitoring data U to be clustered, and move backward with the calculation of the NJW-DTW clustering algorithm, sliding increments is the sampling interval time q; calculate and save the distances of the a and bth battery boxes in each window until the n-d+1th subsequence of length d is formed; and so on, a total of n-d+1 A sliding window and the corresponding distance matrix are expressed as:

DTW _j ={dtw ₁ ,dtw ₂ ,...,dtw _n-d+1 }∈V ^m

The fault condition occurred in one or more of these time windows;

Step 3.2: Calculate the sparse coefficient LSR

The sparsity coefficient LSR(a) of the a-th battery box is defined as the reciprocal of the average distance between this battery box and the remaining battery boxes within the distance k:

In the formula, size (DTW _k (Y _a )) is the number of battery boxes within the distance k of battery box a; DTW (Y _a , Y _b ) is the actual distance of battery boxes a and b within the distance k; no For the loss of generality, the k distance is taken as half of the farthest DTW distance;

Step 3.3: Define the fault threshold FT as the reciprocal of the average distance of all battery boxes to the battery boxes within k distances:

Step 3.4: Arrange the sparse coefficients LSR(a) of all m battery boxes, and use the fault threshold FT as the judgment basis, and identify all battery boxes with sparse coefficients lower than FT as faulty battery boxes.

Further, the target monitoring parameter of the battery box is the terminal voltage of the battery box, the branch current, the battery temperature, the insulation resistance or the battery capacity.

The beneficial effects of the present invention are:

1) The present invention uses the NJW clustering algorithm to reduce the dimension of high-dimensional monitoring data, obtains its eigenvectors by constructing a pull matrix of heterologous data, and then uses the eigenvectors to uniquely replace the original data for cluster analysis. NJW spectral clustering overcomes the limitations of traditional methods on data dimension and sequence length, and solves the singularity problem that often occurs in traditional methods for non-convex data clustering;

2) The present invention uses DTW to dynamically adjust the time axis of asynchronous monitoring data, and maps two sets of monitoring data to the synchronous time axis, which overcomes the observation error caused by the asynchronous sampling of monitoring data, greatly improves the accuracy of clustering results, and solves the problem. The problem that the time axis of heterogeneous data of sampling points in traditional methods cannot be one-to-one correspondence;

3) The present invention constructs a sliding window model to suppress the influence of a small number of outliers in the monitoring data, performs cluster analysis based on the DTW distance, and objectively selects the fault clustering points through the sparse coefficient LSR and the fault threshold FT, thereby avoiding traditional fault clustering. The method can automatically adapt to different BESS parameters and actual working conditions by over-estimating or under-estimating the fault state, so as to realize the self-adaptive early warning of the fault state of the BESS.

Description of drawings

FIG. 1 is a flow chart of a fault state adaptive early warning method of a battery energy storage system according to the present invention.

Figure 2 is a cycle curve of heterologous data on the same time axis.

Figure 3 is a schematic diagram of the sliding window model.

Figure 4 is a schematic diagram of fault clustering.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The technical solution of the present invention is mainly divided into three major steps, namely heterogeneous data clustering, asynchronous sequence dynamic regularization and fault state identification.

1. NJW spectral clustering algorithm for heterogeneous data

The NJW spectral clustering algorithm is a dimensionality reduction clustering algorithm for high-dimensional heterogeneous data. It obtains its eigenvectors by constructing a pull matrix of heterogeneous data, and then uses the eigenvectors to uniquely replace the original data for cluster analysis. NJW spectral clustering has no limit to the data dimension, which effectively avoids the singularity problem that often occurs in heterogeneous non-convex shapes. This embodiment takes the fault state warning based on the terminal voltage of the battery box as an example, and the steps are as follows (the rest of the monitoring data also have completely consistent calculation steps):

1. Set the terminal voltage of the battery box monitored by the battery management system as V, and obtain m groups of battery box monitoring data. Each group of data contains n sampling points, and the monitoring data to be clustered can be expressed as:

U={u ₁ ,u ₂ ,…,u _n }∈V ^m (1)

2. Extract the monitoring data to construct a similarity matrix W={w _ij ∣i≤m,j≤n}∈V ^n×n as follows:

Among them: u _i and u _j represent two heterogeneous battery box terminal voltages, and σ is an adaptive identification parameter. σ _i is the average value of the r sampling points with the smallest Euclidean distance between the target monitoring parameter u _i of the battery box and the rest of the monitoring data, and σ _j is the average value of the r sampling points with the smallest Euclidean distance between the target monitoring parameter u _j of the battery box and the remaining monitoring data. In order to ensure the accuracy of clustering, r is generally taken as 3 to 7. .

3. Calculate the metric matrix D={d _ij ∣i≤m,j≤n}∈V ^n×n according to the similarity matrix W:

4. Calculate the Laplace matrix L from the similarity matrix W and the metric matrix D:

5. Calculate the eigenvalues and eigenvectors corresponding to the pull matrix L, arrange the eigenvalues and eigenvectors in descending order, and take the first K eigenvectors to form a security feature matrix S=[s ₁ ,s ₂ ,...,s _K ]∈V In ^n×K , the S matrix is normalized row by row to form the security feature standard matrix Y={y _ij ∣i≤m,j≤n}∈V ^n×K :

In the formula, s _ij is the element of the ith row and j column of the S matrix.

6. Each row of the safety feature standard matrix Y corresponds to a battery box terminal voltage sequence, which can uniquely replace the original sampling data. It is then subjected to dynamic time warping clustering.

Second, the dynamic time warping algorithm

As shown in Figure 2, the asynchronous monitoring data obtained in different cycle stages of different battery compartments in actual work has the problem of time axis mismatch. The invention improves the K-means clustering algorithm used in the heterologous data NJW spectral clustering algorithm, and improves the Euclidean distance to DTW distance for K-means clustering analysis. The DTW distance compares the similarity of asynchronous monitoring data by regularizing the time axis of the security feature standard matrix Y. The steps are as follows:

1. In the safety feature standard matrix Y, the feature vectors of the monitoring data corresponding to the two asynchronous battery boxes a and b are respectively Y _a ={y _a1 ,y _a2 ,...,y _an } and Y _b ={y _b1 ,y _b2 ,...,y _bn }, n is the number of sample points in each row of the safety feature standard matrix, that is, the number of sampling points included in the monitoring data of each group of battery boxes, the distance matrix R={r of battery boxes a and b can be solved _ij ∣i≤n,j≤n}:

MIN=min{r _i-1,j ,r _i,j-1 ,r _i-1,j-1 } (7)

In the formula, d(y _ai , y _bj ) is the Euclidean distance between the sample points y _ai and y _bj ; d(y _ai , y _bj )+MIN is the sum of the minimum Euclidean distances between the current sample point and the neighboring sample points;

2. After the distance matrix is formed, the DTW distances of battery boxes a and b are

DTW(Y _a ,Y _b )=r _nn a,b≤m (8)

In the same way, the DTW distance measured by any two battery boxes according to the monitoring parameter V can be calculated. When a=b, the DTW distance is zero.

Based on the NJW security feature standard matrix Y, the DTW method dynamically adjusts the time axis of one of the series by finding the numerical similarity of the two sets of monitoring data, avoiding the observation error caused by the asynchronous recording of the data, and can map the two series to the synchronization On the time axis, the similarity of the two sequences can be calculated, which greatly improves the accuracy of the clustering results and avoids the disadvantage that the time axis of the heterologous data of the sampling points cannot be one-to-one corresponding to the traditional method.

3. Fault state identification algorithm

In order to find fault abnormal data in the monitoring data, the present invention proposes a fault state identification method based on the NJW-DTW clustering algorithm. The steps are as follows. In order to simplify the description, the following "distance" refers to the DTW distance:

1. Sliding window model

A sliding window model is used to suppress the influence of a small number of outliers in the monitoring data. BMS monitors the terminal voltage V of the battery box to obtain m groups of battery box monitoring data, each group of data contains n sampling points, and the monitoring data to be clustered can be expressed as: U={u ₁ ,u ₂ ,..., _un }∈ V ^m , 1≤t≤n, place the sliding window of length d (d<<n) at the starting position t ₁ of the monitoring data U to be clustered and move backwards with the calculation of the NJW-DTW clustering algorithm , the sliding increment is the sampling interval time q, as shown in Figure 3. Calculate and save the distances of the a and bth battery boxes in each window until the n-d+1th subsequence of length d is formed; and so on, a total of n-d+1 sliding windows and corresponding The distance matrix of , expressed as: DTW _j ={dtw ₁ ,dtw ₂ ,...,dtw _n-d+1 }∈V ^m , the fault state occurs in one or more time windows.

2. Sparse coefficient LSR (local sparsity ratio, LSR)

The sparsity coefficient LSR(a) of the a-th battery box is defined as the reciprocal of the average distance between this battery box and the remaining battery boxes within the distance k:

In the formula, size (DTW _k (Y _a )) is the number of battery boxes within the distance k of battery box a; DTW (Y _a , Y _b ) is the actual distance of battery boxes a and b within the distance k; no For the loss of generality, the k distance is taken as half of the farthest DTW distance;

The sparse coefficient LSR(i) reflects the safe state distribution density of all battery boxes around battery box i. The smaller the local sparse coefficient, the greater the failure probability of battery box i, and vice versa.

The fault threshold FT (fault threshold, FT) is defined as the reciprocal of the average distance between all battery boxes and the battery boxes within k distances:

If battery box i fails, its sparsity factor should be less than the failure factor FT, because the safe distance of battery box i from other battery boxes is farther than all other battery boxes. All battery boxes whose sparse coefficient LSR is less than the fault factor FT are selected as candidate fault battery boxes, and the fault clustering is shown in Figure 4.

3. Fault status identification

The sparse coefficient LSR(a) of all m battery boxes is arranged, and the fault threshold FT is used as the judgment basis. All battery boxes with a sparse coefficient lower than FT can be regarded as faulty battery boxes.

The invention starts from the heterologous monitoring data, uses the NJW clustering algorithm to reduce the dimension of the high-dimensional data, extracts the characteristic vector of the monitoring data, and eliminates the influence of observation errors and process noise; the DTW method is used to dynamically reorganize the asynchronous data to solve the problem of heterologous data. Data mismatch problem; fault clustering analysis based on objective indicators has strong adaptability to different systems, so that it can objectively reflect the safe operation status of the battery energy storage system and accurately issue fault warnings.

Claims (5)

1. A self-adaptive early warning method for a fault state of a battery energy storage system is characterized by comprising the following steps:

step 1: reducing the dimension of the high-dimensional battery box monitoring data by using an NJW clustering algorithm, acquiring a safety characteristic standard matrix of the high-dimensional battery box monitoring data by constructing a pull-type matrix of heterogeneous data, and performing clustering analysis by using the safety characteristic standard matrix to uniquely replace original data of target monitoring parameters of the battery box;

step 2: regulating the time axis of a safety characteristic standard matrix through a dynamic time regulation algorithm, and mapping two groups of monitoring data onto a synchronous time axis to compare the similarity degree of asynchronous monitoring data;

and 3, step 3: and constructing a sliding window model to inhibit the influence of a small number of outliers in the monitoring data, carrying out cluster analysis based on the DTW distance, objectively selecting fault cluster points through a sparse coefficient LSR and a fault threshold FT, and further determining the fault battery box.

2. The battery energy storage system fault state self-adaptive early warning method according to claim 1, wherein the step 1 specifically comprises the following steps:

step 1.1: setting a target monitoring parameter of a battery management system monitoring battery box as V, acquiring m groups of battery box monitoring data, wherein each group of data comprises n sampling points, and expressing the to-be-clustered monitoring data as follows:

step 1.2: extracting battery box monitoring data to construct similarity matrix W ═ W_ij∣i≤m,j≤n}∈V^n×nThe following were used:

wherein u is_iAnd u_jRepresenting two heterogeneous battery box target monitoring data; sigma_iAnd σ_jTo adaptively identify parameters, [ sigma ]_iIs a target monitoring parameter u of the battery box_iAverage value, sigma, of r sampling points with minimum Euclidean distance in the rest monitoring data_jIs a cell box target monitoring parameter u_jAverage value of r sampling points with minimum Euclidean distance in the rest monitoring data;

step 1.3: calculating to obtain a measurement matrix D ═ { D ═ according to the similarity matrix W_ij∣i≤m,j≤n}∈V^n×n：

Step 1.4: calculating a Laplace matrix L according to the similarity matrix W and the measurement matrix D:

step 1.5: calculating the eigenvalue and eigenvector corresponding to the pull-type matrix L, arranging the eigenvalue and eigenvector in descending order, and taking the first K eigenvectors to form a security eigenvector matrix S ═ S₁,s₂,…,s_K]∈V^n×KIn the method, the security feature matrix S is normalized line by line to form a security feature standard matrix Y ═ Y_ij∣i≤m,j≤n}∈V^n×K：

In the formula s_ijIs the ith row and j columns of elements of the S matrix;

step 1.6: and each row of the safety characteristic standard matrix Y corresponds to a battery box target monitoring parameter sequence and uniquely replaces original sampling data.

3. The battery energy storage system fault state adaptive early warning method according to claim 2, wherein the step 2 specifically comprises the following steps:

step 2.1: in a safety characteristic standard matrix Y, the characteristic vectors of the monitoring data corresponding to two asynchronous battery boxes a and b are respectively Y_a＝{y_a1,y_a2,…,y_an} and Y_b＝{y_b1,y_b2,…,y_bnN is the number of sample points in each row of the safety characteristic standard matrix, namely the number of sample points contained in the monitoring data of each battery box, and the distance matrix R of the battery boxes a and b is obtained by solving (R)_ij∣i≤n,j≤n}：

MIN＝min{r_i-1,j,r_i,j-1,r_i-1,j-1}

In the formula, d (y)_ai,y_bj) Is a sample point y_aiAnd y_bjThe Euclidean distance of; d (y)_ai,y_bj) + MIN is the sum of the minimum Euclidean distance between the current sample point and each adjacent sample point;

step 2.2: after forming the distance matrix, the DTW distance of the battery boxes a, b is:

DTW(Y_a,Y_b)＝r_nn a,b≤m

in the same way, the DTW distance of any two battery boxes measured according to the monitoring parameter V can be calculated, and when a is equal to b, the DTW distance is zero.

4. The battery energy storage system fault state adaptive early warning method according to claim 3, wherein the step 3 specifically comprises the following steps:

step 3.1: sliding window of length d, d<<n, placing the initial position t of the monitoring data U to be clustered₁And moving backwards continuously along with the calculation of the NJW-DTW clustering algorithm, wherein the sliding increment is sampling interval time q; calculating and storing the distance between the a-th battery box and the b-th battery box in each window until an n-d + 1-th subsequence with the length of d is formed; and so on, obtaining n-d +1 sliding windows and corresponding distance matrixes in total, and expressing as:

DTW_j＝{dtw₁,dtw₂,…,dtw_n-d+1}∈V^m

the fault condition occurs in one or more of the time windows;

step 3.2: computing sparse coefficients LSR

Defining the sparse coefficient LSR (a) of the a-th battery box as the reciprocal of the average distance of the battery box within the remaining distance k:

wherein, size (DTW)_k(Y_a) The number of the battery boxes within the distance k of the battery box a is shown in the specification; DTW (Y)_a,Y_b) The actual distance between the battery boxes within the k distance between the battery boxes a and b; the k distance is taken as half of the farthest DTW distance;

step 3.3: the failure threshold FT is defined as the inverse of the average distance of all battery compartments within its k distance:

step 3.4: and arranging the sparse coefficients LSR (a) of all the m battery boxes, and taking the fault threshold value FT as a judgment basis, and identifying all the battery boxes with the sparse coefficients lower than FT as fault battery boxes.

5. The adaptive battery energy storage system fault state early warning method according to any one of claims 1-4, wherein the target battery box monitoring parameter is a battery box terminal voltage, a branch current, a battery temperature, an insulation resistance or a battery capacity.

Images