A method and system for early warning protection of energy storage facilities

Abstract

The application belongs to the field of energy storage monitoring, and provides a kind of energy storage facility safety early warning protection method and system, including obtaining the temperature, temperature rate of rise, equivalent impedance data of each monitoring unit of battery cluster;Physical adjacency matrix is constructed based on the physical space arrangement relationship of monitoring unit;Physical adjacency matrix is fused with electrical adjacency matrix according to preset weight;Graph structure is constructed based on comprehensive adjacency matrix, and Laplacian matrix of graph is calculated;Based on the temperature and temperature rate of rise, the node neighborhood residual of the Laplacian matrix is calculated, and the temperature consistency residual is obtained;The node neighborhood residual of equivalent impedance is calculated, and the electrical consistency residual is obtained;Temperature prediction residual is calculated under time series prediction model with graph regular constraint;Node early abnormal score is constructed, and quantile threshold value is adaptively set according to health operation data distribution, and the early abnormal score is graded early warning determination.The application can improve the accuracy of energy storage facility safety early warning.

Description

Technical Field

[0001] This invention belongs to the field of energy storage monitoring, and specifically relates to a method and system for safety early warning and protection of energy storage facilities. Background Technology

[0002] With the widespread application of new energy storage systems, battery clusters are playing an increasingly important role in scenarios such as grid peak shaving, backup power, and renewable energy grid connection. However, the operating environment of energy storage facilities is complex. During long-term charging and discharging, individual batteries may experience performance degradation or even safety hazards due to heat accumulation, electrochemical aging, or abnormal connections. When a single cell experiences abnormal temperature rise, sudden increase in internal resistance, or severe polarization, if it cannot be detected and isolated in time, the anomaly will spread rapidly along the thermal diffusion path or electrical coupling path, triggering a chain of failures, and in severe cases, may lead to thermal runaway accidents. Therefore, achieving early anomaly identification and proactive safety protection for localized cells in the battery cluster is crucial for the safe operation of energy storage systems.

[0003] Existing energy storage safety monitoring methods mostly rely on macroscopic parameters such as total voltage, total current, or average module temperature for status assessment. Some systems utilize individual cell voltage or temperature monitoring data and set fixed thresholds to trigger alarms for abnormalities. Some research has also attempted to use machine learning methods, employing time series prediction models or statistical anomaly detection algorithms to determine trends in individual cell parameter changes. However, these methods generally suffer from the following limitations: first, they do not fully consider the spatial and electrical interactions between individual cells within the battery cluster, leading to false alarms or missed alarms due to reliance on single-point data; second, the thresholds are fixed and cannot adaptively adjust to changes in operating conditions or the environment; and third, the prediction models lack topological constraints, making it difficult to reflect the coupling relationship between heat diffusion and current conduction. Summary of the Invention

[0004] To address the problems in the prior art, the present invention provides a safety early warning and protection method for energy storage facilities, comprising the following steps:

[0005] The temperature, heating rate, and equivalent impedance data of each monitoring unit in the battery cluster are acquired, wherein the monitoring unit is a single battery cell or a parallel branch;

[0006] A physical adjacency matrix is ​​constructed based on the physical spatial arrangement of the monitoring units, and an electrical adjacency matrix is ​​constructed based on the electrical series-parallel connection relationship of the monitoring units.

[0007] The physical adjacency matrix and the electrical adjacency matrix are merged according to preset weights to obtain a comprehensive adjacency matrix. The rows and columns of the comprehensive adjacency matrix correspond to each monitoring unit of the battery cluster, and the matrix elements represent the physical or electrical adjacency relationship between the monitoring units.

[0008] A graph structure is constructed based on the comprehensive adjacency matrix. The nodes of the graph structure are the monitoring units of the battery cluster, and the edges of the graph structure are the physical or electrical adjacency relationships between the monitoring units. The Laplace matrix of the graph is then calculated.

[0009] Based on the Laplace matrix, the node neighborhood residuals for the temperature and heating rate are calculated to obtain the temperature uniformity residuals; the node neighborhood residuals for the equivalent impedance are calculated to obtain the electrical uniformity residuals.

[0010] A consistency confidence index is generated based on the temperature consistency residual and the electrical consistency residual, and the temperature prediction residual is calculated under a time series prediction model with graph regularization constraints.

[0011] Based on the temperature prediction residual, temperature consistency residual, and electrical consistency residual, an early node anomaly score is constructed, and a quantile threshold is adaptively set according to the distribution of healthy operation data to perform graded early warning judgment on the early anomaly score.

[0012] When the warning level reaches the intervention level or isolation level, the target unit is disconnected according to the risk propagation relationship between the physical and electrical neighborhoods.

[0013] Furthermore, the matrix used to describe the spatial adjacency relationship of battery cells or parallel branches in the battery cluster, the rows and columns of the physical adjacency matrix all correspond to monitoring units. When two monitoring units are physically adjacent in the actual structure, the corresponding element in the matrix takes the value of one, and when two monitoring units are not adjacent, the corresponding element takes the value of zero.

[0014] The electrical adjacency matrix is ​​used to describe the series or parallel relationship of battery cells or parallel branches in the circuit connection method. The rows and columns of the electrical adjacency matrix also correspond to the monitoring units. When two monitoring units are directly connected in series or parallel in electrical connection, the corresponding element in the matrix takes the value of one, otherwise it takes the value of zero.

[0015] Furthermore, let the composite adjacency matrix be A, the physical adjacency matrix be Ap, and the electrical adjacency matrix be Ae, with weight coefficients α and β respectively. Then the composite adjacency matrix is ​​expressed as A = αAp + βAe, where α is the physical adjacency weight and β is the electrical adjacency weight.

[0016] Furthermore, for each node in the graph structure constructed by the comprehensive adjacency matrix, all non-zero elements in its corresponding row are traversed, the values ​​of these elements are summed and used as the degree value of the node, and filled into the diagonal of the corresponding position in the degree matrix to obtain the degree matrix reflecting the connection strength between each monitoring unit and its neighborhood in the topology.

[0017] After the degree matrix is ​​generated, the Laplacian matrix of the graph is obtained by subtracting the composite adjacency matrix from the degree matrix.

[0018] Furthermore, the time series prediction model with graph regularization constraints is implemented based on a recurrent neural network structure. A graph regularization term is added to the network's loss function, which is calculated by multiplying the Laplacian matrix with the predicted temperature vector to penalize cases where the predicted value differs too much between neighboring nodes.

[0019] The present invention also provides a safety early warning and protection system for energy storage facilities, comprising the following modules:

[0020] The data acquisition module is used to acquire the temperature, heating rate, and equivalent impedance data of each monitoring unit of the battery cluster, wherein the monitoring unit is a single battery cell or a parallel branch.

[0021] A matrix construction module is used to construct a physical adjacency matrix based on the physical spatial arrangement relationship of the monitoring units, and to construct an electrical adjacency matrix based on the electrical series-parallel connection relationship of the monitoring units;

[0022] The matrix fusion module is used to fuse the physical adjacency matrix and the electrical adjacency matrix according to a preset weight to obtain a comprehensive adjacency matrix. The rows and columns of the comprehensive adjacency matrix correspond to each monitoring unit of the battery cluster, and the matrix elements represent the physical or electrical adjacency relationship between the monitoring units.

[0023] The graph structure generation module is used to construct a graph structure based on the comprehensive adjacency matrix. The nodes of the graph structure are the monitoring units of the battery cluster, and the edges of the graph structure are the physical or electrical adjacency relationships between the monitoring units. The Laplace matrix of the graph is calculated accordingly.

[0024] The residual calculation module is used to calculate the node neighborhood residual based on the Laplace matrix for the temperature and heating rate to obtain the temperature-consistent residual, and to calculate the node neighborhood residual for the equivalent impedance to obtain the electrical consistency residual.

[0025] The predictive analysis module is used to generate a consistency confidence index based on the temperature consistency residual and the electrical consistency residual, and to calculate the temperature prediction residual under a time series prediction model with graph regularization constraints.

[0026] The early warning judgment module is used to construct an early node anomaly score based on the temperature prediction residual, temperature consistency residual and electrical consistency residual, and adaptively set a quantile threshold according to the distribution of healthy operation data to perform graded early warning judgment on the early anomaly score.

[0027] The circuit breaker control module is used to perform a circuit breaker operation on the target unit according to the risk propagation relationship between the physical and electrical neighborhoods when the warning level reaches the intervention or isolation level.

[0028] Furthermore, the building module includes a physical adjacency matrix generation unit and an electrical adjacency matrix generation unit.

[0029] The physical adjacency matrix generation unit is used to describe the spatial adjacency relationship of battery cells or parallel branches in the battery cluster. The rows and columns of the physical adjacency matrix correspond to the monitoring units. When two monitoring units are physically adjacent in the actual structure, the corresponding element in the matrix takes the value of one. When two monitoring units are not adjacent, the corresponding element takes the value of zero.

[0030] The electrical adjacency matrix generation unit is used to describe the series or parallel relationship of battery cells or parallel branches in the circuit connection method. The rows and columns of the electrical adjacency matrix also correspond to the monitoring units. When two monitoring units are directly connected in series or parallel in electrical connection, the corresponding element in the matrix takes the value of one, otherwise it takes the value of zero.

[0031] Furthermore, the matrix fusion module is used to generate a comprehensive adjacency matrix according to the formula A=αAp+βAe, where A is the comprehensive adjacency matrix, Ap is the physical adjacency matrix, Ae is the electrical adjacency matrix, α is the physical adjacency weight, and β is the electrical adjacency weight.

[0032] Furthermore, the graph structure generation module includes a degree matrix calculation unit and a Laplacian matrix calculation unit.

[0033] The degree matrix calculation unit is used to traverse the non-zero elements in each row of the comprehensive adjacency matrix, sum the values ​​of these elements and use them as the degree values ​​of the nodes, and fill them into the diagonal elements of the corresponding positions in the degree matrix to obtain a degree matrix that reflects the topological connection strength between each monitoring unit and its neighborhood.

[0034] The Laplacian matrix calculation unit is used to obtain the Laplacian matrix of the graph by subtracting the comprehensive adjacency matrix from the degree matrix after the degree matrix is ​​generated.

[0035] Furthermore, the predictive analysis module includes a time series prediction model unit with graphical regularization constraints.

[0036] The time series prediction model unit is implemented based on a recurrent neural network structure. A graph regularization term is added to the network's loss function. It is calculated by multiplying the Laplacian matrix with the predicted temperature vector. This term is used to penalize situations where the predicted values ​​differ too much between adjacent nodes, thereby enhancing the spatial consistency constraint capability of the prediction results.

[0037] This invention establishes a fusion model of physical adjacency matrix and electrical adjacency matrix at the battery cluster level, achieving a unified expression of the spatial and electrical connections between individual battery cells. This enables the monitoring system to accurately identify the combined influence paths of thermal diffusion and current coupling. By synthesizing the graph structure constructed from the adjacency matrix and the corresponding Laplace matrix, the topological relationships between cell temperature, heating rate, and equivalent impedance are explicitly quantified, providing a reliable structural foundation for subsequent consistency residual analysis and risk propagation modeling.

[0038] This invention introduces a time series prediction model with graph regularization constraints, enabling the temperature prediction process to consider not only the dynamic trends of the time dimension but also the smoothing constraints of spatial topology. This effectively suppresses random fluctuations in individual prediction results and improves the sensitivity and stability of anomaly deviation identification. Simultaneously, the system adaptively sets quantile thresholds based on the distribution of healthy operating data, overcoming the insensitivity of traditional fixed threshold methods to environmental changes and aging characteristics. This makes early warning judgments more flexible, robust, and possesses long-term adaptability.

[0039] Furthermore, upon detecting that the warning level has reached the intervention or isolation condition, this invention performs a circuit breaker operation by combining the risk propagation relationship between the physical and electrical neighborhoods, cutting off the anomaly propagation path at the topological level and achieving dual protection against thermal diffusion and electrical instability. This mechanism significantly reduces the probability of local anomalies evolving into systemic failures, improving the safety, controllability, and reliability of energy storage systems under complex operating conditions. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram illustrating the principle of the present invention. Detailed Implementation

[0042] The embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the following embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0043] In this embodiment, the energy storage facility can be a battery cluster composed of multiple battery cells. Each battery cell or parallel branch is equipped with temperature, voltage, current, and impedance detection devices as a monitoring unit, thereby enabling real-time acquisition of operating parameters. This embodiment establishes physical adjacency relationships and electrical connections within the battery cluster, constructs a comprehensive adjacency matrix, and builds a graph structure based on this matrix to reflect the topological coupling characteristics between monitoring units within the battery cluster. Under the constraints of this graph structure, this embodiment performs neighborhood consistency analysis on the temperature, heating rate, and equivalent impedance of the monitoring units, and introduces a graph regularization term into the time prediction model to achieve cross-neighborhood parameter prediction and anomaly deviation detection.

[0044] refer to Figure 1 As shown, the specific implementation includes the following steps:

[0045] The temperature, heating rate, and equivalent impedance data of each monitoring unit in the battery cluster are acquired, wherein the monitoring unit is a single battery cell or a parallel branch.

[0046] A physical adjacency matrix is ​​constructed based on the physical spatial arrangement of the monitoring units, and an electrical adjacency matrix is ​​constructed based on the electrical series-parallel connection relationship of the monitoring units.

[0047] The physical adjacency matrix and the electrical adjacency matrix are merged according to preset weights to obtain a comprehensive adjacency matrix. The rows and columns of the comprehensive adjacency matrix correspond to each monitoring unit of the battery cluster, and the matrix elements represent the physical or electrical adjacency relationship between the monitoring units.

[0048] A graph structure is constructed based on the comprehensive adjacency matrix. The nodes of the graph structure are the monitoring units of the battery cluster, and the edges of the graph structure are the physical or electrical adjacency relationships between the monitoring units. The Laplace matrix of the graph is then calculated.

[0049] Based on the Laplace matrix, the node neighborhood residuals for the temperature and heating rate are calculated to obtain the temperature consistency residuals; the node neighborhood residuals for the equivalent impedance are calculated to obtain the electrical consistency residuals.

[0050] A consistency confidence index is generated based on the temperature consistency residual and the electrical consistency residual, and the temperature prediction residual is calculated under a time series prediction model with graph regularization constraints.

[0051] Based on the temperature prediction residual, temperature consistency residual, and electrical consistency residual, an early node anomaly score is constructed, and a quantile threshold is adaptively set according to the distribution of healthy operation data to perform graded early warning judgment on the early anomaly score.

[0052] When the warning level reaches the intervention level or isolation level, the target unit is disconnected according to the risk propagation relationship between the physical and electrical neighborhoods.

[0053] Through the above steps, this embodiment enables consistency constraint analysis of individual cells and their neighborhoods at the battery cluster level, avoiding false alarms or missed alarms caused by relying on isolated single-point parameters. Simultaneously, by utilizing a time prediction model with graph regularization constraints, anomaly detection not only depends on historical trends but is also constrained by topological relationships, thus enabling more accurate identification of early signs of localized thermal failure. Therefore, this embodiment effectively improves the safety early warning capability of energy storage facilities under complex operating conditions, ensuring the timeliness and reliability of risk management.

[0054] To further illustrate the technical solution of the present invention, the following detailed explanation and description of each step of the method are provided.

[0055] In the safety early warning of energy storage facilities, the operating status of individual battery cells or parallel branches directly determines the stability of the entire battery cluster. Relying solely on total voltage or total current signals can easily mask anomalies in individual cells, making it impossible to identify early failure characteristics in a timely manner. To achieve early monitoring of potential thermal failures, this embodiment first collects parameters from each monitoring unit of the battery cluster, especially three key parameters: temperature, heating rate, and equivalent impedance, to ensure that the thermal stability and electrochemical state of individual battery cells can be reflected from multiple dimensions.

[0056] In this invention, the monitoring unit refers to the basic detection object that constitutes the battery cluster. It can be a single battery cell or a branch composed of multiple cells connected in parallel. Its feature is that each monitoring unit can independently acquire operating parameters and participate in subsequent analysis as an independent node.

[0057] In this invention, equivalent impedance refers to the AC or DC impedance value of a single cell or branch obtained through specific detection methods. This impedance can reflect the internal polarization of the battery, the integrity of the conductive path, and the stability of the interface.

[0058] In the specific implementation, temperature data is collected by placing temperature sensors on the surface of battery cells or branches. The sensors are preferably thermocouples or high-precision thermistors, and the sampling frequency can be set from once every one to ten seconds, depending on the application scenario. After acquiring the temperature data, the heating rate can be obtained by calculating the difference in temperature change within a continuous sampling period. The heating rate can characterize whether there is abnormal rapid heating in the battery cell. The equivalent impedance can be obtained through the AC impedance testing module integrated in the battery management system, or it can be approximated by measuring the ratio of voltage response to current under static or low-current pulse conditions. In practical applications, the appropriate detection method can be selected according to the energy storage system design.

[0059] By collecting the above parameters, the operating characteristics of the battery cluster at the thermal and electrical levels can be obtained in real time, enabling early identification of local anomalies and avoiding the problem of early warning delay caused by relying solely on macroscopic bus signals. This provides a reliable data foundation for subsequent consistency analysis and early warning models.

[0060] In energy storage facilities, individual cells or parallel branches not only have physical adjacency relationships but also series-parallel connections in their electrical connections. Modeling based solely on a single relationship fails to fully reflect the topological characteristics of the cell cluster, potentially leading to biases in consistency analysis. For example, temperature diffusion and heat dissipation depend on physical adjacency, while current distribution and voltage characteristics are primarily influenced by electrical adjacency. Therefore, to ensure that both spatial thermal coupling and electrical coupling characteristics are considered in subsequent anomaly detection, both physical adjacency matrices and electrical adjacency matrices need to be constructed. Thus, this embodiment constructs a physical adjacency matrix based on the physical spatial arrangement of the monitoring units and an electrical adjacency matrix based on the electrical series-parallel connections of the monitoring units.

[0061] In this invention, the physical adjacency matrix refers to a matrix used to describe the spatial adjacency relationship of battery cells or parallel branches in a battery cluster. The rows and columns of this matrix correspond to monitoring units. When two monitoring units are physically adjacent in the actual structure, the corresponding element in the matrix has a value of one; when two monitoring units are not adjacent, the corresponding element has a value of zero. In this invention, the electrical adjacency matrix refers to a matrix used to describe the series or parallel connection relationship of battery cells or parallel branches in the circuit connection method. The rows and columns of this matrix also correspond to monitoring units. When two monitoring units are directly connected in series or parallel in the electrical connection, the corresponding element in the matrix has a value of one; otherwise, it has a value of zero.

[0062] In the specific implementation process, the first step is to obtain the geometric arrangement information of the battery clusters. This information typically comes from the structural design data of the battery module or the on-site layout diagram. Each monitoring unit has a definite spatial position within the module or cabinet, which can be described by position coordinates. By comparing the relative positional relationship of two monitoring units in spatial coordinates, it can be determined whether they are physically adjacent. When the spatial distance between the center points of two monitoring units is less than a preset threshold, they are determined to be adjacent, and the corresponding element in the physical adjacency matrix is ​​assigned a value of one; when the distance is greater than the threshold, they are determined to be non-adjacent, and the corresponding element is assigned a value of zero. To improve the accuracy of the judgment, the threshold can be set according to the external dimensions of the battery cell and the module gap. For example, when the cell has a rectangular prism structure, the threshold can be equal to the cell width plus the maximum allowable installation gap.

[0063] Subsequently, the series and parallel connections between the monitoring units are determined based on the electrical design topology diagram of the battery cluster. The electrical topology diagram records the connection method and connected objects of each monitoring unit in the circuit. When two monitoring units are electrically directly connected, the corresponding element in the electrical adjacency matrix is ​​assigned a value of one; otherwise, it is assigned a value of zero. Both series and parallel connections can be automatically resolved from the connection information in the topology diagram. For example, if five battery cells are connected in parallel to form a group, then in the electrical adjacency matrix corresponding to these five cells, all elements between them are assigned a value of one; if four groups of parallel cells are connected in series sequentially, then the matrix elements corresponding to adjacent groups are assigned a value of one, while the matrix elements corresponding to non-directly connected groups are assigned a value of zero.

[0064] In terms of implementation, the generation of the physical adjacency matrix and the electrical adjacency matrix can be automatically completed by the control software during the initialization phase of the battery management system. After receiving the module structure data and circuit topology data, the control software can automatically parse the spatial coordinates and circuit connection relationships, and complete the matrix generation. For energy storage systems pre-configured during the design phase, the engineering designers can also directly write the adjacency relationships in the structural parameter file, and the system can call this file during deployment to generate the required matrices.

[0065] When battery clusters are large, such as exceeding a hundred cells, manually marking adjacency relationships one by one becomes inefficient. To ensure efficiency and accuracy, an automated script can be used to read battery layout diagrams and circuit connection tables stored in a database, automatically calculate the spatial adjacency and electrical connectivity between cells, and generate physical and electrical adjacency matrices in batches. The automated script can use geometric calculation methods to determine spatial proximity and directly construct electrical adjacency relationships using the series and parallel connection information identified in the circuit connection table. In this way, adjacency matrices for hundreds to thousands of cells can be generated within seconds to tens of seconds, ensuring high efficiency and accuracy in the construction process.

[0066] For example, in a battery cluster consisting of twenty individual cells, the layout diagram shows the cells arranged in a 4x5 rectangular array. When analyzing the coordinates, the control software automatically identifies adjacent horizontal and vertical cells and assigns them a value of one in the physical adjacency matrix. In the electrical topology, these twenty cells are divided into four series and five parallel groups. Every five cells are assigned a value of one to each other in the electrical adjacency matrix. The matrix elements corresponding to adjacent series-connected groups of five parallel groups are assigned a value of one, while cells not directly connected are assigned a value of zero in the matrix. The physical and electrical adjacency matrices generated in this way can comprehensively and accurately reflect the physical arrangement and electrical connection relationships of the battery cluster, providing a reliable foundation for subsequent consistency residual analysis and graph regularization modeling.

[0067] The physical adjacency matrix and the electrical adjacency matrix are merged according to preset weights to obtain a comprehensive adjacency matrix. The rows and columns of the comprehensive adjacency matrix correspond to each monitoring unit of the battery cluster, and the matrix elements represent the physical or electrical adjacency relationship between the monitoring units.

[0068] In modeling the operational status of a battery cluster, the physical adjacency matrix reflects the spatial thermal conduction and heat dissipation relationships of the batteries, while the electrical adjacency matrix reflects the current distribution and voltage coupling relationships of the batteries within the circuit. Using only a single matrix can only reflect the correlation between batteries in some dimensions, easily leading to incomplete information and affecting the accurate identification of abnormal states. Therefore, this embodiment fuses the physical and electrical adjacency matrices, resulting in a matrix that simultaneously includes both physical spatial features and electrical connection features, thus comprehensively depicting the internal topological relationships of the battery cluster. Therefore, the physical and electrical adjacency matrices are fused according to preset weights to obtain a comprehensive adjacency matrix. The rows and columns of this comprehensive adjacency matrix correspond to the monitoring units of the battery cluster, and the matrix elements represent the physical or electrical adjacency relationships between the monitoring units.

[0069] In this invention, the comprehensive adjacency matrix refers to an adjacency matrix obtained mathematically through weighted superposition. The row and column indices of this matrix correspond to monitoring units within the battery cluster, and the elements in the matrix simultaneously reflect the physical adjacency and electrical connectivity of the monitoring units. In this invention, the preset weight refers to a coefficient manually set during the fusion process. This coefficient is used to balance the importance of physical and electrical adjacency in different application scenarios.

[0070] In the specific implementation process, it is first necessary to ensure that the dimensions of the physical adjacency matrix and the electrical adjacency matrix are consistent, with the number of rows and columns corresponding to the total number of monitoring units in the battery cluster, ensuring that the elements of each matrix can correspond one-to-one during the fusion process. Then, weighting coefficients are set according to different application scenarios and monitoring objectives. In thermal safety risk monitoring scenarios, since heat mainly spreads through conduction and convection diffusion between physically adjacent batteries, the weight of the physical adjacency matrix can be appropriately increased. In current distribution analysis or voltage balance monitoring scenarios, since current is directly conducted in series and parallel paths, the weight of the electrical adjacency matrix can be increased. In this way, the importance of the two types of adjacency relationships can be flexibly adjusted according to different detection objectives.

[0071] In this invention, temperature is the primary external manifestation of thermal failure in a battery cell. When a battery experiences abnormal temperature rise, its heat diffuses to physically adjacent cells, causing the temperature of neighboring cells to also rise. If only the temperature of the cell itself is considered without considering the state of adjacent cells, delayed identification of local anomalies can easily occur. By constructing a physical adjacency matrix, this heat diffusion path can be mathematically represented, enabling subsequent analysis to capture the localized coordinated heating phenomenon caused by heat conduction. On the other hand, the series and parallel connections of batteries in the circuit also affect the propagation of anomalies. When the internal resistance or voltage of a battery cell deviates abnormally, its electrically adjacent cells will be directly affected by the uneven current distribution, potentially leading to additional heating or even performance degradation in these adjacent cells. Therefore, by constructing an electrical adjacency matrix, this electrical coupling relationship can be explicitly represented.

[0072] In particular, the interaction between two monitoring units becomes more significant when a single battery cell is both spatially adjacent to another battery cell and directly connected in electrical topology. In this case, the pair of monitoring units will not only experience coordinated temperature changes through thermal conduction but also interact through electrical stress transmission. Therefore, during the integration process, units that possess both physical and electrical adjacency should be given higher weight to reflect their crucial role in risk propagation.

[0073] In the specific calculation, the physical adjacency matrix and the electrical adjacency matrix are weighted and superimposed according to preset weight coefficients to obtain the comprehensive adjacency matrix. If expressed as a formula, the comprehensive adjacency matrix can be set as A, the physical adjacency matrix as Ap, and the electrical adjacency matrix as Ae, with weight coefficients α and β respectively. The comprehensive adjacency matrix is ​​then expressed as A = αAp + βAe, where α is the physical adjacency weight and β is the electrical adjacency weight. The rows and columns of A, Ap, and Ae correspond to the monitoring units in the battery cluster, and the matrix elements represent the adjacency relationship between two monitoring units. When two units are physically adjacent, the corresponding element in Ap is assigned a value of one; when two units are electrically adjacent, the corresponding element in Ae is assigned a value of one; if two units are both physically and electrically adjacent, the corresponding element in both Ap and Ae is assigned a value of one, and this element will have a higher weight in the fusion matrix A.

[0074] In practice, the weighting coefficients α and β can be set empirically. For example, in scenarios primarily focused on temperature anomaly detection, α can be set to 0.7 and β to 0.3. In scenarios primarily focused on current balance detection, α can be set to 0.4 and β to 0.6. The weighting coefficients can also be adaptively adjusted based on historical operating data. The control software can calculate the contribution of temperature and electrical anomalies to system failures based on long-term monitoring data, dynamically correcting the weight allocation to make the comprehensive adjacency matrix more consistent with actual operating conditions.

[0075] Through the above steps, the resulting comprehensive adjacency matrix can simultaneously reflect the thermal diffusion path and electrical coupling relationship in a unified mathematical structure. In particular, for battery cells that are both physically and electrically adjacent, their potential risk propagation effect can be enhanced in the matrix, thereby enabling more accurate anomaly identification in subsequent residual analysis and risk assessment.

[0076] For example, in the aforementioned battery cluster composed of twenty individual cells, the physical adjacency matrix reflects the spatial adjacency relationships arranged in a four-row, five-column configuration, while the electrical adjacency matrix reflects the electrical connection relationships of four series and five parallel connections. In the actual generation of the comprehensive adjacency matrix, the physical adjacency weight α is set to 0.6, and the electrical adjacency weight β is set to 0.4. For cells numbered 8 and 9, located in the same row and belonging to the same parallel branch, the corresponding elements in Ap and Ae are both 1. After fusion, their matrix element values ​​are 1.0, reflecting a higher adjacency strength. In this way, the comprehensive adjacency matrix can accurately represent the physical and electrical adjacency within the battery cluster, ensuring more precise anomaly detection and risk propagation analysis.

[0077] In energy storage facilities, the interactions between individual battery cells or parallel branches are reflected not only in changes in numerical parameters but also in the topological connections within the overall structure. If only adjacency matrices are used to record single numerical relationships without transforming them into a graph structure for system modeling, the global connections and network characteristics between nodes cannot be fully expressed, making it difficult to support subsequent graph-based residual analysis and anomaly detection. By transforming the comprehensive adjacency matrix into a graph structure, the node and edge relationships of each monitoring unit in the battery cluster can be uniformly described at the mathematical level, thereby constructing a topological network that simultaneously reflects physical and electrical adjacency. Therefore, a graph structure is constructed based on the comprehensive adjacency matrix, where the nodes of the graph structure represent the monitoring units of the battery cluster, and the edges represent the physical or electrical adjacency relationships between the monitoring units, thereby calculating the Laplace matrix of the graph.

[0078] In this invention, a graph structure refers to a mathematical structure composed of nodes and edges, where nodes represent monitoring units in a battery cluster, and edges represent the adjacency relationship between two monitoring units. This adjacency relationship includes physical adjacency and electrical adjacency. In this invention, the Laplace matrix is ​​a matrix constructed based on the degree matrix and adjacency matrix of the graph structure. This matrix can be used to measure the degree of difference between nodes, and its essence is to reflect the overall network coordination by characterizing the relationship between nodes and their neighbors.

[0079] In the specific implementation process, the integrated adjacency matrix is ​​first used as input. The rows and columns of the integrated adjacency matrix correspond to monitoring units in the battery cluster, and each monitoring unit is defined as a node in the graph structure. Based on the element values ​​in the integrated adjacency matrix, it is determined whether there are edges between the nodes. When the element value at a certain position in the integrated adjacency matrix is ​​greater than zero, it indicates that there is a physical or electrical adjacency relationship between the corresponding two monitoring units, and the system establishes an edge connecting these two nodes in the graph structure. If the element value is zero, it indicates that there is no direct adjacency relationship between the two, and no corresponding edge is added to the graph.

[0080] In this invention, the degree matrix refers to a diagonal matrix generated based on a graph structure, where the diagonal elements are the sum of the edge weights between the node and all its neighboring nodes. The degree matrix is ​​constructed as follows: for each node, all non-zero elements in its corresponding row of the integrated adjacency matrix are traversed, and the sum of these elements is used as the node's degree value, which is then filled into the diagonal of the corresponding position in the degree matrix. This process clearly reflects the topological connection strength between each monitoring unit and its neighborhood.

[0081] After the degree matrix is ​​generated, the Laplace matrix of the graph is obtained by subtracting the comprehensive adjacency matrix from the degree matrix. In this invention, the Laplace matrix can be used to measure the consistency deviation between nodes. Its essential function is to transform the relationship between nodes and their neighbors into mathematical constraints, so as to accurately identify local anomalies in subsequent residual calculation and time prediction.

[0082] To ensure the accuracy and real-time performance of matrix calculations, a matrix operation module is integrated into the battery management system. This module automatically maps the adjacency matrix to a graph structure and further calculates the degree matrix and Laplacian matrix. This module employs sparse matrix storage to reduce computational complexity and optimizes computational efficiency through parallel computing. During the initialization phase, the system loads the comprehensive adjacency matrix of the battery clusters. Subsequently, during operation, it periodically calculates or updates the Laplacian matrix to adapt to dynamic changes in the battery cluster's operating environment or connection relationships.

[0083] Through the steps described above, the comprehensive adjacency relationships of battery clusters can be automatically transformed into a graph structure, and a Laplacian matrix can be generated based on this, thus providing a mathematical foundation for subsequent consistency residual analysis. This method not only reduces errors from manual modeling but also improves computational efficiency and real-time performance in large-scale battery cluster scenarios.

[0084] In the safety monitoring of battery clusters, parameter fluctuations in individual battery cells or parallel branches may not directly reflect risk, as some fluctuations may originate from sensor noise or individual differences. Relying solely on single-point values ​​for judgment can easily lead to false alarms or missed alarms. By calculating neighborhood residuals based on the graph Laplace matrix, the differences between a single monitoring unit and its neighboring units can be measured, thereby identifying whether there are abnormal deviations between a particular cell and its neighbors. Temperature and its rate of change can be used to obtain temperature consistency residuals from neighborhood residuals, and equivalent impedance can be used to obtain electrical consistency residuals from neighborhood residuals. This method can more objectively determine whether the parameter anomalies of a particular cell exceed the normal fluctuation range. Therefore, based on the Laplace matrix, the node neighborhood residuals for temperature and heating rate are calculated to obtain temperature consistency residuals, and the node neighborhood residuals for equivalent impedance are calculated to obtain electrical consistency residuals.

[0085] In this invention, the neighborhood residual refers to the difference between the parameter value of a monitoring unit and the weighted average of the parameters of its neighboring nodes. This difference is calculated using a Laplace matrix and is used to measure the degree of deviation between a single unit and its neighborhood. In this invention, the temperature consistency residual refers to the neighborhood residual calculated from temperature and heating rate, used to reflect the deviation of a monitoring unit's thermal behavior from its neighbors. In this invention, the electrical consistency residual refers to the neighborhood residual calculated from equivalent impedance, used to reflect the degree of difference between a monitoring unit's electrical state and its neighbors.

[0086] In the specific implementation process, it is first necessary to acquire the temperature data and heating rate data of each monitoring unit in the battery cluster within the current sampling period, and construct them into a temperature vector and a heating rate vector, respectively. The temperature vector reflects the absolute temperature state of the monitoring unit at that moment, and the heating rate vector reflects the rate of temperature change of the monitoring unit in adjacent sampling periods. In this invention, the temperature vector and heating rate vector are used as inputs to perform operations with the Laplace matrix, and the neighborhood residual of each monitoring unit in the thermal dimension is obtained through matrix multiplication. This neighborhood residual characterizes the degree of difference between the temperature or heating rate of a certain individual unit and the weighted average state of its neighboring units.

[0087] When the temperature consistency residual of a monitoring unit is close to zero, it indicates that its temperature or heating rate is consistent with that of its neighboring nodes, which is normal thermal behavior. When the residual value increases significantly, it indicates that the thermal parameters of that unit differ from those of its neighbors, which may manifest as abnormal heating, uneven heat dissipation, or local hot spots. In this way, abnormal phenomena where a single point deviates from the population trend can be detected in a timely manner, without the overall stability of the neighboring states masking the anomalies of that unit.

[0088] Subsequently, the equivalent impedance data of each monitoring unit is obtained in the electrical dimension and constructed into an impedance vector. Equivalent impedance characterizes the battery's conductivity and polarization characteristics, and varies due to factors such as internal active material degradation and interfacial film deterioration. In this invention, the impedance vector is calculated using a Laplace matrix to obtain the neighborhood residual of each monitoring unit in the electrical dimension. This electrical consistency residual reflects the degree of difference between the impedance value of that individual unit and the average impedance of its neighbors.

[0089] If the electrical consistency residual of a given cell is close to zero or at a low level, it indicates that the impedance state of that cell is consistent with its neighbors and within the normal consistency range. If the residual value increases significantly, it indicates that the impedance of that cell is significantly higher or lower than that of its neighbors, which may mean that the cell has an abnormal electrode interface, internal damage, or a tendency to fail locally. In practical applications, this anomaly usually appears before macroscopic deviations in voltage and current, so residual calculation can enable early identification.

[0090] To ensure real-time computation, the system can periodically perform residual calculations within the battery management unit. The calculation cycle can be set to seconds or minutes depending on application requirements. In battery cluster scenarios with large data volumes, sparse matrix storage and parallel computing strategies can be employed to improve computational efficiency and avoid unnecessary redundant calculations.

[0091] In this way, the temperature consistency residual and the electrical consistency residual can respectively characterize the deviation of the monitoring unit in the thermal dimension and the electrical dimension, thereby achieving cross-dimensional consistency monitoring.

[0092] In the operational monitoring of battery clusters, relying solely on temperature or impedance residuals is insufficient to accurately determine anomalies in individual cells, as data from different dimensions contains noise and random fluctuations, potentially leading to false alarms. To enhance reliability, it is necessary to fuse multidimensional residuals to obtain a comprehensive confidence index that measures whether a monitored cell deviates from the overall state of its neighborhood. Furthermore, the temperature evolution of a single cell is not an isolated change but rather a process that evolves together with time and neighborhood topological relationships. Without considering time trends and neighborhood constraints, it is difficult to detect potential early signs of thermal failure. Therefore, this embodiment fuses temperature consistency residuals and electrical consistency residuals to generate a consistency confidence index, and calculates the temperature prediction residual under a time series prediction model with graph regularization constraints.

[0093] In this invention, the consistency confidence index is a comprehensive quantitative indicator obtained by fusing the temperature consistency residuals of the thermal dimension and the electrical consistency residuals of the electrical dimension. This index is used to characterize the reliability of a monitoring unit maintaining consistency with its neighborhood under overall operating conditions. In this invention, the graph regularization-constrained time series prediction model is a model that introduces adjacency constraints on the basis of traditional time series prediction methods. This constraint incorporates the mutual influence between nodes and their neighbors into the prediction calculation through the graph Laplace matrix, thereby making the prediction results more consistent with the actual topology of the battery cluster.

[0094] In the specific implementation process, the temperature consistency residual and the electrical consistency residual first need to be numerically normalized to eliminate deviations caused by differences in dimensions or inconsistent numerical ranges between different physical quantities. During normalization, minimum-maximum value normalization or standard deviation normalization methods can be used to ensure that the processed residual values ​​are within a uniform range. After normalization, the temperature consistency residual and the electrical consistency residual are weighted and fused according to preset weighting coefficients to obtain a consistency confidence index. The weighting coefficients can be set according to the contribution ratio of thermal anomalies and electrical anomalies to system failures in historical operating data. For example, in scenarios with a high risk of thermal failure, the weight of the temperature residual is increased; in scenarios with a high proportion of internal damage, the weight of the impedance residual is increased. The weighting coefficients can also be dynamically updated by an optimization algorithm to make the consistency confidence index more consistent with the real-time characteristics of the operating conditions. The numerical range of the consistency confidence index is usually limited to between zero and one. The closer the value is to one, the higher the consistency between the monitoring unit and its neighbors; the lower the value, the greater the deviation between the unit and its neighbors.

[0095] In the temperature prediction stage, this embodiment uses a time series prediction model with graph regularization constraints to estimate the future temperature state of each monitoring unit. This model takes historical temperature series as input and not only considers the temporal evolution of the individual unit, but also introduces neighborhood topological constraints provided by the integrated adjacency matrix and Laplace matrix, thereby incorporating spatial correlation into the time prediction framework.

[0096] In this invention, the graph-regularized time series prediction model can be implemented using a recurrent neural network (RNN) structure. For example, a Long Short-Term Memory (LSTM) network can be used as the basic prediction model. The temperature sequence of each monitoring unit within the most recent sampling periods is input into the network to capture the nonlinear evolution trend over time. To introduce neighborhood constraints, a graph regularization term is added to the network's loss function. This term is calculated by multiplying the Laplacian matrix by the predicted temperature vector and is used to penalize situations where the predicted value differs excessively from neighboring nodes. In this way, when updating the weight parameters, the prediction model not only minimizes the difference between the predicted and true values ​​but also automatically adjusts to maintain the consistency of neighborhood prediction results.

[0097] In the specific implementation, assuming the battery cluster contains N monitoring units, the historical temperature sequence at a certain time t is used as input to predict the temperature value at time t+1. After the prediction is completed, the prediction result is compared with the actual collected temperature value to obtain the temperature prediction residual. If the temperature prediction residual remains at a high level over multiple consecutive sampling periods, it indicates that the temperature evolution trend of the cell deviates significantly from the normal evolution trajectory depicted by the model, thus suggesting that the cell may be in the early stage of thermal failure.

[0098] In actual operation, the model can be deployed in the central control module of the battery management system, using real-time collected temperature data and adjacency relationships for prediction and residual calculation. To improve efficiency, the model training process can be completed offline, and parameters can be fine-tuned during operation through a continuous learning mechanism to adapt to changes in operating conditions and environmental conditions.

[0099] Through the above methods, this embodiment not only realizes static anomaly identification based on consistency residuals, but also realizes dynamic anomaly trend identification by combining a time series prediction model with graph regularization constraints. It can provide effective early warning before thermal failure occurs, significantly improving the safety and reliability of energy storage facilities.

[0100] During the operation of battery clusters in energy storage facilities, deviations from a single parameter are often insufficient to accurately identify early risks. This is because temperature prediction residuals, temperature consistency residuals, and electrical consistency residuals each reflect abnormal characteristics in different dimensions. Relying on only one of these residuals may lead to misjudgments due to data noise or local fluctuations. By comprehensively calculating these three types of residuals, the operating status of the monitoring unit can be jointly assessed across multiple dimensions, including thermal, electrical, and temporal evolution, resulting in a quantified anomaly score. This anomaly score comprehensively reflects potential failure signs in individual cells. Furthermore, by using an adaptive quantile threshold setting method based on the distribution of healthy operating data, graded early warning judgments can be achieved without relying on fixed thresholds, thus adapting to different operating environments and individual differences. Therefore, an early node anomaly score is constructed based on the aforementioned temperature prediction residuals, temperature consistency residuals, and electrical consistency residuals. A quantile threshold is adaptively set according to the distribution of healthy operating data to perform graded early warning judgments on the early anomaly score.

[0101] In this invention, the early node anomaly score refers to a numerical value formed by weighted fusion of temperature prediction residuals, temperature consistency residuals, and electrical consistency residuals, used to quantify the degree to which each monitoring unit deviates from its normal state. In this invention, the adaptive quantile threshold refers to a judgment threshold dynamically set according to the statistical quantile method based on the distribution of a large amount of data collected during the healthy operation phase of the battery cluster. This threshold can adaptively adjust with different system and environmental conditions, without relying on a pre-fixed value.

[0102] In the specific implementation process, the temperature prediction residual, temperature consistency residual, and electrical consistency residual first need to be numerically normalized to ensure that the three types of indicators are within a unified numerical range, avoiding uneven weight calculation due to different units of measurement. Normalization can be achieved using linear normalization, mapping the residual values ​​to the interval between zero and one, or using standard deviation normalization, ensuring that the residual values ​​are symmetrically distributed around the mean. This processing step guarantees that outlier indicators from different physical dimensions can be compared and weighted under the same unit of measurement.

[0103] After normalization, weighting coefficients need to be set according to different application scenarios and risk concerns. For example, when battery clusters are in a high-power charge-discharge environment, temperature rise is the main driving factor for potential thermal failure; in this case, the weights of temperature prediction residuals and temperature consistency residuals can be increased. In long-term cycle testing, electrochemical degradation is more sensitive to impedance; in this case, the weight of electrical consistency residuals can be appropriately increased. The node early anomaly score obtained after weighted calculation can comprehensively reflect the degree of deviation of individual battery cells in both thermal and electrical aspects. The higher the score, the greater the degree of deviation of the monitoring unit from normal operating conditions.

[0104] In the threshold determination section, this invention proposes an adaptive setting method based on the distribution of healthy operating data. Healthy operating data refers to a large amount of historical monitoring data collected when the battery cluster is in a stable operating state, free from thermal failure and electrical faults. In this invention, the healthy operating data distribution refers to the statistical distribution characteristics obtained by calculating the residuals from the aforementioned historical monitoring data. This distribution reflects the fluctuation range of each residual index under normal conditions.

[0105] In the specific implementation process, firstly, during the early stages of battery cluster operation or system initialization, parameters such as temperature and impedance are collected over multiple cycles, and their prediction residuals and consistency residuals are calculated respectively. After obtaining the data samples, the residuals of each type are statistically analyzed to form a residual distribution curve. The system calculates the quantiles of this distribution curve, such as the 80th, 90th, and 95th percentiles. The residual values ​​corresponding to these percentile values ​​serve as the adaptive thresholds for different warning levels.

[0106] During operation, the system continuously updates the healthy operating data samples. When the operating status is determined to be normal, the new residual data will be included in the distribution calculation to dynamically adjust the quantile threshold. When an abnormal operating status is detected, the abnormal data will no longer be included in the update of the healthy data distribution, thereby avoiding contamination of the normal distribution by faulty samples and ensuring the accuracy and stability of the threshold setting.

[0107] During the operation of energy storage facilities, when a single battery cell malfunctions, the risk is not limited to itself but can also rapidly propagate to adjacent cells through thermal diffusion or circuit coupling, causing the localized anomaly to escalate into a systemic failure. If the early warning stage only involves providing a notification without taking mandatory isolation measures, delays could lead to thermal runaway or the spread of electrical faults. Therefore, when the warning level reaches the intervention or isolation level, the system needs to perform a circuit breaker operation based on the propagation relationship between the physical and electrical neighborhoods within the battery cluster to sever the coupling path between the target cell and neighboring cells, thereby preventing further risk propagation. Therefore, when the warning level reaches the intervention or isolation level, a circuit breaker operation is performed on the target cell according to the risk propagation relationship between the physical and electrical neighborhoods.

[0108] In this invention, circuit breaking operation refers to isolating a target cell or parallel branch deemed abnormal from the rest of the circuit by controlling electronic switches or relays in the battery management system, thus preventing it from participating in energy transfer and current coupling. In this invention, risk propagation relationship refers to analyzing the range of cells potentially affected by an anomaly based on the physical and electrical adjacency relationships represented by the adjacency matrix, combined with thermal diffusion paths and current distribution paths.

[0109] In practice, the early warning module first determines the anomaly level of a monitoring unit. When the anomaly level reaches the intervention level, the system identifies the range of physically and electrically adjacent units based on the adjacency matrix and calculates possible risk propagation paths. Subsequently, the control module sends a control signal to the circuit breaker, causing the electronic switch corresponding to the target unit to activate, thereby achieving physical isolation. When the anomaly level further reaches the isolation level, the system expands the isolation range, not only cutting off the target unit but also performing graded circuit breaker operations on neighboring units affected by risk propagation, ensuring that potential anomalies do not spread along thermal or electrical paths.

[0110] In engineering implementation, circuit breaking operations are typically achieved through MOSFET switches, relays, or solid-state circuit breakers configured within the battery cluster. To ensure rapid and reliable execution, the control software triggers hardware action within milliseconds upon receiving an isolation command. Simultaneously, the circuit breaking status is recorded in real time and uploaded to the upper-level monitoring platform for subsequent analysis and maintenance.

[0111] This embodiment provides a safety early warning and protection system for energy storage facilities, which consists of the following functional modules:

[0112] The data acquisition module is used to acquire the temperature, heating rate, and equivalent impedance data of each monitoring unit of the battery cluster, wherein the monitoring unit is a single battery cell or a parallel branch.

[0113] A matrix construction module is used to construct a physical adjacency matrix based on the physical spatial arrangement relationship of the monitoring units, and to construct an electrical adjacency matrix based on the electrical series-parallel connection relationship of the monitoring units;

[0114] The matrix fusion module is used to fuse the physical adjacency matrix and the electrical adjacency matrix according to a preset weight to obtain a comprehensive adjacency matrix. The rows and columns of the comprehensive adjacency matrix correspond to each monitoring unit of the battery cluster, and the matrix elements represent the physical or electrical adjacency relationship between the monitoring units.

[0115] The graph structure generation module is used to construct a graph structure based on the comprehensive adjacency matrix. The nodes of the graph structure are the monitoring units of the battery cluster, and the edges of the graph structure are the physical or electrical adjacency relationships between the monitoring units. The Laplace matrix of the graph is calculated accordingly.

[0116] The residual calculation module is used to calculate the node neighborhood residual based on the Laplace matrix for the temperature and heating rate to obtain the temperature-consistent residual, and to calculate the node neighborhood residual for the equivalent impedance to obtain the electrical consistency residual.

[0117] The predictive analysis module is used to generate a consistency confidence index based on the temperature consistency residual and the electrical consistency residual, and to calculate the temperature prediction residual under a time series prediction model with graph regularization constraints.

[0118] The early warning judgment module is used to construct an early node anomaly score based on the temperature prediction residual, temperature consistency residual and electrical consistency residual, and adaptively set a quantile threshold according to the distribution of healthy operation data to perform graded early warning judgment on the early anomaly score.

[0119] The circuit breaker control module is used to perform a circuit breaker operation on the target unit according to the risk propagation relationship between the physical and electrical neighborhoods when the warning level reaches the intervention or isolation level.

[0120] In a further implementation, the matrix construction module includes a physical adjacency matrix generation unit and an electrical adjacency matrix generation unit.

[0121] The physical adjacency matrix generation unit is used to describe the spatial adjacency relationship of battery cells or parallel branches in the battery cluster. The rows and columns of the physical adjacency matrix correspond to the monitoring units. When two monitoring units are physically adjacent in the actual structure, the corresponding element in the matrix takes the value of one. When two monitoring units are not adjacent, the corresponding element takes the value of zero.

[0122] The electrical adjacency matrix generation unit is used to describe the series or parallel relationship of battery cells or parallel branches in the circuit connection method. The rows and columns of the electrical adjacency matrix also correspond to the monitoring units. When two monitoring units are directly connected in series or parallel in electrical connection, the corresponding element in the matrix takes the value of one, otherwise it takes the value of zero.

[0123] In a further implementation, the matrix fusion module is used to generate a comprehensive adjacency matrix according to the formula A=αAp+βAe, where A is the comprehensive adjacency matrix, Ap is the physical adjacency matrix, Ae is the electrical adjacency matrix, α is the physical adjacency weight, and β is the electrical adjacency weight.

[0124] In a further implementation, the graph structure generation module includes a degree matrix calculation unit and a Laplacian matrix calculation unit.

[0125] The degree matrix calculation unit is used to traverse the non-zero elements in each row of the comprehensive adjacency matrix, sum the values ​​of these elements and use them as the degree values ​​of the nodes, and fill them into the diagonal elements of the corresponding positions in the degree matrix to obtain a degree matrix that reflects the topological connection strength between each monitoring unit and its neighborhood.

[0126] The Laplacian matrix calculation unit is used to obtain the Laplacian matrix of the graph by subtracting the comprehensive adjacency matrix from the degree matrix after the degree matrix is ​​generated.

[0127] In a further implementation, the predictive analysis module includes a time series prediction model unit with graphical regularization constraints.

[0128] The time series prediction model unit is implemented based on a recurrent neural network structure. A graph regularization term is added to the network's loss function. It is calculated by multiplying the Laplacian matrix with the predicted temperature vector. This term is used to penalize situations where the predicted values ​​differ too much between adjacent nodes, thereby enhancing the spatial consistency constraint capability of the prediction results.

[0129] It should be noted that the explanation of the aforementioned embodiments of the energy storage facility safety early warning and protection method also applies to the apparatus of the embodiments of this application, and will not be repeated here.

[0130] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0131] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0132] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0133] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. For some module structures not specifically defined in this invention, the content described in the prior art shall prevail. The prior art mentioned in the foregoing background and specific embodiments sections can be considered as part of this invention and used to understand the meaning of some technical features or parameters.

Claims

1. A method for safety early warning and protection of energy storage facilities, characterized in that, The method includes: Acquire temperature, heating rate, and equivalent impedance data of each monitoring unit in the battery cluster, wherein the monitoring unit is a single battery cell or a parallel branch; A physical adjacency matrix is ​​constructed based on the physical spatial arrangement of the monitoring units, and an electrical adjacency matrix is ​​constructed based on the electrical series-parallel connection of the monitoring units. The physical adjacency matrix and the electrical adjacency matrix are merged according to preset weights to obtain a comprehensive adjacency matrix. The rows and columns of the comprehensive adjacency matrix correspond to each monitoring unit of the battery cluster, and the matrix elements represent the physical or electrical adjacency relationship between the monitoring units. A graph structure is constructed based on the comprehensive adjacency matrix. The nodes of the graph structure are the monitoring units of the battery cluster, and the edges of the graph structure are the physical or electrical adjacency relationships between the monitoring units. The Laplace matrix of the graph is then calculated. Based on the Laplace matrix, the node neighborhood residuals for the temperature and heating rate are calculated to obtain the temperature uniformity residuals; the node neighborhood residuals for the equivalent impedance are calculated to obtain the electrical uniformity residuals. The neighborhood residual refers to the difference between the parameter value of the monitoring unit and the weighted average of the parameters of its neighboring nodes. This difference is calculated by the Laplace matrix and is used to measure the degree of deviation between the unit and its neighborhood. Temperature consistency residual refers to the neighborhood residual calculated from temperature and heating rate, which reflects the deviation of the thermal behavior of a monitoring unit from its neighbors; the temperature vector and heating rate vector are used as inputs to perform operations with the Laplace matrix, and the neighborhood residual of each monitoring unit in the thermal dimension is obtained through matrix multiplication. Electrical consistency residuals refer to the neighborhood residuals calculated from equivalent impedances, which reflect the degree of difference between the electrical state of a monitoring unit and its neighbors; by operating the impedance vector and the Laplace matrix, the neighborhood residuals of each monitoring unit in the electrical dimension are obtained. A consistency confidence index is generated based on the temperature consistency residual and the electrical consistency residual, and the temperature prediction residual is calculated under a time series prediction model with graph regularization constraints. Based on the temperature prediction residual, temperature consistency residual, and electrical consistency residual, an early node anomaly score is constructed, and a quantile threshold is adaptively set according to the distribution of healthy operation data to perform graded early warning judgment on the early anomaly score. When the warning level reaches the intervention level or isolation level, the target unit is disconnected according to the risk propagation relationship between the physical and electrical neighborhoods.

2. The energy storage facility safety early warning and protection method according to claim 1, characterized in that, The physical adjacency matrix is ​​used to describe the spatial adjacency relationship of battery cells or parallel branches in battery clusters. The rows and columns of the physical adjacency matrix correspond to monitoring units. When two monitoring units are physically adjacent in the actual structure, the corresponding element in the matrix takes the value of one, and when two monitoring units are not adjacent, the corresponding element takes the value of zero. The electrical adjacency matrix is ​​used to describe the series or parallel relationship of battery cells or parallel branches in the circuit connection method. The rows and columns of the electrical adjacency matrix also correspond to the monitoring units. When two monitoring units are directly connected in series or parallel in electrical connection, the corresponding element in the matrix takes the value of one, otherwise it takes the value of zero.

3. The method for safety early warning and protection of energy storage facilities according to claim 1, characterized in that, Let the composite adjacency matrix be A, the physical adjacency matrix be Ap, and the electrical adjacency matrix be Ae, with weight coefficients α and β respectively. Then the composite adjacency matrix is ​​expressed as A = αAp + βAe, where α is the physical adjacency weight and β is the electrical adjacency weight.

4. The method for safety early warning and protection of energy storage facilities according to claim 1, characterized in that, For each node in the graph structure constructed by the comprehensive adjacency matrix, traverse all non-zero elements in its corresponding row, sum the values ​​of these elements and use them as the degree value of the node, and fill them into the diagonal of the corresponding position in the degree matrix to obtain the degree matrix that reflects the connection strength between each monitoring unit and its neighborhood in the topology. After the degree matrix is ​​generated, the Laplacian matrix of the graph is obtained by subtracting the composite adjacency matrix from the degree matrix.

5. The method for safety early warning and protection of energy storage facilities according to claim 1, characterized in that, The time series prediction model with graph regularization constraint is implemented based on a recurrent neural network structure. A graph regularization term is added to the network's loss function, which is calculated by multiplying the Laplacian matrix with the predicted temperature vector to penalize cases where the predicted value differs too much between neighboring nodes.

6. A safety early warning and protection system for energy storage facilities, characterized in that, The system includes the following modules: The data acquisition module is used to acquire the temperature, heating rate, and equivalent impedance data of each monitoring unit of the battery cluster, wherein the monitoring unit is a single battery cell or a parallel branch. A matrix construction module is used to construct a physical adjacency matrix based on the physical spatial arrangement relationship of the monitoring units, and to construct an electrical adjacency matrix based on the electrical series-parallel connection relationship of the monitoring units; The matrix fusion module is used to fuse the physical adjacency matrix and the electrical adjacency matrix according to a preset weight to obtain a comprehensive adjacency matrix. The rows and columns of the comprehensive adjacency matrix correspond to each monitoring unit of the battery cluster, and the matrix elements represent the physical or electrical adjacency relationship between the monitoring units. The graph structure generation module is used to construct a graph structure based on the comprehensive adjacency matrix. The nodes of the graph structure are the monitoring units of the battery cluster, and the edges of the graph structure are the physical or electrical adjacency relationships between the monitoring units. The Laplace matrix of the graph is calculated accordingly. The residual calculation module is used to calculate the node neighborhood residual based on the Laplace matrix for the temperature and heating rate to obtain the temperature-consistent residual, and to calculate the node neighborhood residual for the equivalent impedance to obtain the electrical consistency residual. The neighborhood residual refers to the difference between the parameter value of the monitoring unit and the weighted average of the parameters of its neighboring nodes. This difference is calculated by the Laplace matrix and is used to measure the degree of deviation between the unit and its neighborhood. Temperature consistency residual refers to the neighborhood residual calculated from temperature and heating rate, which reflects the deviation of the thermal behavior of a monitoring unit from its neighbors; the temperature vector and heating rate vector are used as inputs to perform operations with the Laplace matrix, and the neighborhood residual of each monitoring unit in the thermal dimension is obtained through matrix multiplication. Electrical consistency residuals refer to the neighborhood residuals calculated from equivalent impedances, which reflect the degree of difference between the electrical state of a monitoring unit and its neighbors; by operating the impedance vector and the Laplace matrix, the neighborhood residuals of each monitoring unit in the electrical dimension are obtained. The predictive analysis module is used to generate a consistency confidence index based on the temperature consistency residual and the electrical consistency residual, and to calculate the temperature prediction residual under a time series prediction model with graph regularization constraints. The early warning judgment module is used to construct an early node anomaly score based on the temperature prediction residual, temperature consistency residual and electrical consistency residual, and adaptively set a quantile threshold according to the distribution of healthy operation data to perform graded early warning judgment on the early anomaly score. The circuit breaker control module is used to perform a circuit breaker operation on the target unit according to the risk propagation relationship between the physical and electrical neighborhoods when the warning level reaches the intervention or isolation level.

7. The energy storage facility safety early warning and protection system according to claim 6, characterized in that: The matrix construction module includes a physical adjacency matrix generation unit and an electrical adjacency matrix generation unit. The physical adjacency matrix generation unit is used to describe the spatial adjacency relationship of battery cells or parallel branches in the battery cluster. The rows and columns of the physical adjacency matrix correspond to monitoring units. When two monitoring units are physically adjacent in the actual structure, the corresponding element in the matrix takes a value of one; when two monitoring units are not adjacent, the corresponding element takes a value of zero. The electrical adjacency matrix generation unit is used to describe the series or parallel relationship of battery cells or parallel branches in the circuit connection method. The rows and columns of the electrical adjacency matrix also correspond to monitoring units. When two monitoring units are directly connected in series or parallel in the electrical connection, the corresponding element in the matrix takes a value of one; otherwise, the value is zero.

8. The energy storage facility safety early warning and protection system according to claim 6, characterized in that: The matrix fusion module is used to generate a comprehensive adjacency matrix according to the formula A=αAp+βAe, where A is the comprehensive adjacency matrix, Ap is the physical adjacency matrix, Ae is the electrical adjacency matrix, α is the physical adjacency weight, and β is the electrical adjacency weight.

9. The energy storage facility safety early warning and protection system according to claim 6, characterized in that: The graph structure generation module includes a degree matrix calculation unit and a Laplacian matrix calculation unit. The degree matrix calculation unit is used to traverse the non-zero elements in each row of the comprehensive adjacency matrix, accumulate the values ​​of these elements and use them as the degree values ​​of the nodes, and fill them into the diagonal elements of the corresponding positions in the degree matrix to obtain a degree matrix that reflects the topological connection strength between each monitoring unit and its neighborhood. The Laplacian matrix calculation unit is used to obtain the Laplacian matrix of the graph by subtracting the comprehensive adjacency matrix from the degree matrix after the degree matrix is ​​generated.

10. The energy storage facility safety early warning and protection system according to claim 6, characterized in that: The predictive analysis module includes a time series prediction model unit with graph regularization constraints. The time series prediction model unit is implemented based on a recurrent neural network structure. A graph regularization term is added to the loss function of the network. It is calculated by multiplying the Laplacian matrix and the predicted temperature vector. This term is used to penalize cases where the predicted values ​​differ too much between adjacent nodes, thereby enhancing the spatial consistency constraint capability of the prediction results.

Images