Energy storage battery health degree evaluation method and system

Abstract

The present application belongs to the technical field of battery, more particularly, relates to a kind of energy storage battery health degree evaluation method and system. The comprehensive relationship network in the battery group to be evaluated is established and the adjacency matrix corresponding to the comprehensive relationship network;The feature vector of the battery group to be evaluated is extracted using the comprehensive relationship network and the adjacency matrix of the battery group to be evaluated;The feature vector and the adjacency matrix are input into the graph neural network model, and the health degree score of each battery in the battery group to be evaluated is obtained. Through the present application, the problem that the existing energy storage battery health degree evaluation method monitors and analyzes each single battery in the battery group as an independent object, and cannot reflect the mutual influence and coupling relationship between single batteries.

Description

Technical Field

[0001] This invention belongs to the field of battery-related technology, and more specifically, relates to a method and system for evaluating the health of energy storage batteries. Background Technology

[0002] In smart grids, energy storage batteries, as core components of large-capacity energy storage devices, must maintain stable operation during high-voltage, high-current charging and discharging processes. To ensure the performance consistency of the battery pack, the voltage, current, and temperature parameters of each individual battery cell must be monitored synchronously in real time. It is estimated that by the end of 2020, the total energy storage capacity of my country's power grid will reach 130 million kilowatts, of which energy storage batteries will account for about 10%, or a total capacity of about 13,000 MW, highlighting their important role in bridging the gap between peak electricity demand and average load.

[0003] However, current energy storage battery health assessments primarily rely on State of Health (SOH) algorithms built into battery management systems (BMS), which have significant limitations. Traditional SOH algorithms are typically based on single parameters such as internal resistance or capacity decay, failing to comprehensively reflect the differences between individual cells within the battery pack, making it difficult to identify aging or failure risks in advance. Existing methods often use fixed thresholds to trigger warnings; when parameters reach critical points, the battery is often already severely aged, accelerating the performance degradation of the entire battery pack. Furthermore, the high-frequency data collected by the BMS is not fully utilized, lacking dynamic analysis and predictive capabilities, and cannot effectively identify batteries in a "sub-healthy" state.

[0004] While some existing machine learning-based battery health assessment methods attempt to incorporate multi-parameter analysis, they suffer from the following drawbacks: low data preprocessing efficiency, reliance on large amounts of labeled data for model training, high computational complexity, and difficulty in achieving real-time assessment. More importantly, existing algorithms fail to effectively model the spatial topological relationships and temporal evolution patterns among individual cells within the battery pack, neglecting the inherent characteristics of the battery pack as a complex network system.

[0005] A battery pack is essentially a complex network system formed by multiple individual cells connected in series and parallel. These individual cells are not only physically adjacent but also electrically coupled and influence each other. When a cell begins to age, it affects the operating state of surrounding cells through mechanisms such as current distribution and temperature conduction, creating a cascading effect. Traditional methods analyze each cell as an independent entity, failing to capture this network effect and resulting in inaccurate evaluation results. Therefore, a method that can solve these problems is urgently needed. Summary of the Invention

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for evaluating the health of energy storage batteries. This solves the problem that existing energy storage battery health assessment methods monitor and analyze each individual battery cell in the battery pack as an independent object, failing to reflect the mutual influence and coupling relationships between individual batteries.

[0007] To achieve the above objectives, according to one aspect of the present invention, a method for evaluating the health of an energy storage battery is provided, the method comprising the following steps: Based on the physical location and electrical connection of the batteries in the battery pack to be evaluated, a spatial physical relationship network between the batteries is established; the operational similarity between time series of different battery parameters is calculated to establish a temporal correlation network between batteries with similar operational characteristics; the spatial physical relationship network and the temporal correlation network are coupled to form a comprehensive relationship network of the battery pack to be evaluated and the adjacency matrix corresponding to the comprehensive relationship network. The degree of parameter difference between each battery and its associated batteries in the battery pack to be evaluated is extracted using the comprehensive relationship network and adjacency matrix of the battery pack to be evaluated, i.e., consistency characteristics; dynamic response characteristics of the rate and pattern of change of each battery parameter over time; electrical coupling characteristics reflecting the battery's internal resistance and power characteristics; and network position characteristics of each battery in the comprehensive relationship network. The consistency characteristics, dynamic response characteristics, electrical coupling characteristics and network position characteristics are combined to form the feature vector of the battery pack to be evaluated. The feature vectors and adjacency matrices are input into a graph neural network model to obtain the health score of each battery in the battery pack to be evaluated.

[0008] More preferably, the adjacency matrix is ​​calculated using the following formula:

[0009]

[0010] in, To integrate weight parameters, This is the adjacency matrix of the comprehensive relational network. This is the adjacency matrix of a spatial physical relation network. This is the adjacency matrix of the time-series performance relation network. For space physical relationship network The average of all non-zero elements in the set. For time-series performance relational network The average of all non-zero elements in the matrix.

[0011] More preferably, the and The calculation formula is as follows:

[0012]

[0013] in, The adjacency matrix of the spatial physical relation network is the first... Line number The element of the column represents the first element. Battery No. 1 and No. 2 The strength of the spatial relationship between the cells To establish the strength of the fixed relationship between battery pairs connected in series. For the strength of the fixed relationship between the parallel battery pairs, The relationship strength weight is calculated based on physical distance. Let be the Euclidean distance between the two batteries. This is the threshold for spatial adjacency determination. The adjacency matrix of the time-series performance relation network is the first... Line 1 The element of the column represents the first element. Battery No. 1 and No. 2 Strength of time-series performance relationship between cells For the first Battery No. 1 and No. 2 Similarity of operating characteristics between battery numbers This is the similarity threshold.

[0014] More preferably, the local feature consistency features include voltage difference values, current difference values, and temperature difference values, each calculated using the following formulas:

[0015] in, For the first The weighted average voltage difference between battery number 1 and its associated batteries. For the first The set of associated batteries for battery number 1, i.e., the adjacency matrix of the comprehensive relationship network. No. The set of battery numbers corresponding to the non-zero elements in the row. The adjacency matrix of the comprehensive relational network is the first... Line number The element of the column represents the first element. Battery No. 1 and No. 2 The strength of the relationship between the number of batteries For the first The voltage value of battery number at the last moment of the current evaluation window. For the first The voltage value of battery number at the last moment of the current evaluation window. This represents the number of times the voltage data was collected.

[0016] Further preferably, the dynamic response characteristics include: voltage change rate statistical characteristics, current change rate statistical characteristics, and temperature change rate statistical characteristics, specifically including: the mean of the voltage change rate sequence, reflecting the average rate of voltage change; the standard deviation of the voltage change rate sequence, reflecting the degree of fluctuation in the rate of voltage change; the maximum and minimum values ​​of the voltage change rate sequence, reflecting extreme cases of voltage change; the mean of the current change rate sequence, reflecting the average rate of current change; the standard deviation of the current change rate sequence, reflecting the degree of fluctuation in the rate of current change; the maximum and minimum values ​​of the current change rate sequence, reflecting extreme cases of current change; the mean of the temperature change rate sequence, reflecting the average rate of temperature change; the standard deviation of the temperature change rate sequence, reflecting the degree of fluctuation in the rate of temperature change; the maximum and minimum values ​​of the temperature change rate sequence, reflecting extreme cases of temperature change; a total of 12 statistical quantities, comprehensively depicting the dynamic change patterns of the battery in the three dimensions of voltage, current, and temperature.

[0017] More preferably, the electrical coupling feature comprises four components, each of which is calculated using the following formula:

[0018]

[0019]

[0020]

[0021] in, For the first Estimated average instantaneous internal resistance of battery number 1. For the selected time pairs, For the first Battery No. 1 in Voltage value at the time of the first acquisition For the first Battery No. 1 in the Voltage value at the time of the first acquisition For the first Battery No. 1 in the Current value at the time of the first acquisition For the first Battery No. 1 in the Current value at the time of the first acquisition To prevent small constants from being divided by zero, For the first Battery No. 1 in the Instantaneous power at the time of the first acquisition For the first Battery No. 1 in the Voltage value at the time of the first acquisition For the first Battery No. 1 in the Current value at the time of the first acquisition The number of times the voltage data was collected. This refers to the sequence number of the time collection point. For the first The average power of battery number 1 For the first The power standard deviation of battery number 1 For the first Battery No. 1 in the Instantaneous power at the time of the first acquisition This represents the number of times the voltage data was collected.

[0022] More preferably, the network topology features include , and The corresponding formulas are as follows:

[0023]

[0024]

[0025] in, For the first The degree centrality of cell A represents the number of its neighbors; To meet Battery serial number A set; The cardinality (radix) represents the number of elements in the set. The adjacency matrix of the comprehensive relational network is the first... Line number Column elements, For the first The weighted degree centrality of cell A represents the sum of the relationship strengths between cell A and all its neighbors. This represents the total number of individual cells in the battery pack. For the first The clustering coefficient of cell number indicates the degree of connectivity between its neighbors. For the first The actual number of connections between the neighbors of battery number 1; For the first The number of neighbors of battery number 1, if Then define .

[0026] More preferably, the graph neural network includes four graph convolutional layers, and the formula for calculating the health score in the graph neural network is as follows:

[0027] The loss function for a graph neural network is:

[0028] in, and These are the weight matrix and bias term of the output layer, respectively. The fourth convolutional layer is the first... The feature vector of battery number 1 The training samples are numbered. This represents the total number of training samples. To predict health status, For true health status labeling, For the first The health score of each battery.

[0029] According to another aspect of the present invention, a battery health evaluation system is provided, the system including an actuator for performing the battery health evaluation method described above.

[0030] According to another aspect of the present invention, a computer storage medium is provided having a computer program stored thereon, which, when executed, is used to implement the above-described method for evaluating the health of an energy storage battery.

[0031] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: 1. This method constructs a two-layer battery relationship network that integrates physical spatial relationships and operational performance correlations. Based on this network, it extracts multi-dimensional features reflecting the battery health status, modeling the battery pack as a relationship network system. While analyzing the state of individual cells, it fully considers their interactions with adjacent and electrically connected cells. By introducing network modeling concepts, it can comprehensively reflect the overall characteristics of the battery pack as a system. By analyzing the temporal evolution patterns of parameters, it can detect early signs of deteriorating health status. By considering the importance of the cell's position in the network, it can accurately identify the key cells that have the greatest impact on overall performance. These technical features give this method significant advantages over existing technologies in terms of assessment accuracy, early warning timeliness, and maintenance decision support capabilities. It also has dynamic trend analysis capabilities, enabling the extraction of early warning information from the temporal evolution patterns of parameters, thus achieving early identification of battery health deterioration.

[0032] 2. This invention utilizes a comprehensive relationship network and adjacency matrix to extract multidimensional features reflecting battery health status. Consistency features describe the degree of parameter difference between each battery and its associated batteries; a greater difference indicates that the battery's operating state deviates more from the normal mode. Dynamic response features describe the rate and pattern of change of battery parameters over time; abnormal change patterns often foreshadow deterioration in health status. Network location features describe the importance of the battery in the relationship network; batteries in key positions have a greater impact on overall performance. Through the comprehensive extraction of multidimensional features, this invention can comprehensively characterize battery health status from three different perspectives: consistency, dynamism, and network characteristics. This avoids the one-sidedness caused by traditional methods relying on a single parameter, making the evaluation results more accurate and reliable. It can detect abnormal trends before battery parameters exceed alarm thresholds, providing sufficient response time for maintenance decisions.

[0033] 3. This invention employs graph neural network (Graph Neural Network) technology for deep learning and analysis of the extracted multidimensional features. Graph Neural Networks can directly propagate information and aggregate features within the network structure, automatically learning the interaction patterns between batteries. The network progressively extracts health status features from local to global levels through a multi-layered structure, ultimately outputting a health score for each battery. The score ranges from zero to one hundred, with higher scores indicating better health. The introduction of Graph Neural Networks enables the model to automatically learn complex interaction patterns between batteries without requiring manual rule setting, significantly improving the intelligence level of the assessment. Through end-to-end learning, assessment errors are greatly reduced, the accuracy of health status classification is improved, and rapid assessment capabilities are provided, meeting the requirements for real-time monitoring of energy storage systems. Attached Figure Description

[0034] Figure 1 This is an overall flowchart of the energy storage battery health evaluation method constructed according to a preferred embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram illustrating the construction and fusion of a double-layer battery relationship network according to a preferred embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0037] This invention provides a method for evaluating the health of energy storage batteries based on battery relationship network analysis. This method models the battery pack as a network system with individual batteries as nodes and inter-battery relationships as connections, and assesses the battery health status by analyzing network structure characteristics and node evolution patterns.

[0038] The implementation of this technical solution consists of three main stages. The first stage is data acquisition and network construction, which involves collecting battery pack operating data and establishing a network model describing the relationships between batteries. The second stage is feature extraction and network learning, which involves extracting health status features from the network model and learning them using a graph neural network. The third stage is health calculation and evaluation, which involves mapping the learned features to a health score and generating an evaluation report.

[0039] like Figure 1 As shown, this invention constructs a two-layer battery relationship network that integrates physical spatial relationships and operational performance correlations, and extracts multi-dimensional features reflecting the battery's health status based on this network. The technical solution includes four steps: data acquisition, two-layer relationship network construction, network feature extraction, and health scoring. Details are as follows: S1 Data Acquisition Data acquisition employs a multi-parameter synchronous acquisition strategy. The voltage data acquisition period is 200 milliseconds, the current data acquisition period is 200 milliseconds, the temperature data acquisition period is 1 second, and the data acquisition duration for a single evaluation is 3 minutes. For data including... The data collected from the battery pack containing individual cells is organized into three matrices: Voltage data matrix (dimension) ),element Indicates the first Battery No. 1 in the Voltage value at the time of the first acquisition.

[0040] Current data matrix (dimension) ),element Indicates the first Battery No. 1 in the The current value during the first data acquisition.

[0041] Temperature data matrix (dimension) ),element Indicates the first Battery No. 1 in the Temperature value at the time of the first collection.

[0042] in (3 minutes / 200 milliseconds) (3 minutes / 1 second) represents the number of times each parameter was collected. After data collection, outlier removal and noise filtering preprocessing were performed.

[0043] S2 Dual-Layer Battery Relationship Network Construction A battery relationship network is constructed, incorporating both spatial and temporal dimensions of association information. Spatially, a spatial association network is established based on the physical location and electrical connections of the batteries, establishing connections between adjacent batteries and electrically connected batteries. Temporally, a temporal association network is established by calculating the similarity between time series of different battery parameters, establishing connections between batteries with similar operating characteristics. These two dimensions of association networks together constitute a complete network model describing the characteristics of the battery pack system.

[0044] like Figure 2 As shown, the dual-layer battery relationship network comprises two sub-networks: a spatial-physical relationship network and a temporal performance relationship network. The spatial-physical relationship network reflects the physical location and electrical connections of the batteries, while the temporal performance relationship network reflects the similarity of battery operating characteristics. The two sub-networks are weighted and fused to form a comprehensive relationship network.

[0045] (1) Construction of spatial physical relationship network Adjacency matrix for spatial physical relation networks This indicates that the matrix dimension is Matrix elements Indicates the first Battery No. 1 and No. 2 The strength of the spatial relationship between batteries is determined by considering both physical distance and electrical connection.

[0046] First, a battery location coordinate table is established to record the three-dimensional spatial coordinates of each battery within the battery pack. The coordinates of battery number 1 are as follows , No. The coordinates of battery number 1 are as follows Calculate the Euclidean distance between the two batteries:

[0047] Set spatial adjacency determination threshold ,when At that time, it is assumed that the two batteries are spatially adjacent. Threshold Determined based on the physical layout of the battery pack, it is preferably set to 1.5 times the center-to-center distance between adjacent batteries.

[0048] For battery pairs that are spatially adjacent, the strength of their relationship is calculated using a Gaussian distance decay function:

[0049] Among them, attenuation parameter Preferably, it is set to half the distance between adjacent batteries, which gives greater weight to battery pairs that are closer together.

[0050] In addition to physical distance, electrical connections are also included in the spatial relationship network. Examining the circuit connection topology of the battery pack identifies battery pairs with direct electrical connections. For two batteries connected in series, their relationship strength is set to a fixed value because they share the same current path. For two batteries connected in parallel, due to the current redistribution effect, their relationship strength is set to a fixed value. Preferably, .

[0051] Considering both physical distance and electrical connection factors, the adjacency matrix elements of the spatial physical relationship network Determined according to the following rules:

[0052] (2) Construction of time-series performance relationship network Adjacency matrix for time-series performance relational networks This indicates that the matrix dimension is Matrix elements Indicates the first Battery No. 1 and No. 2 The similarity of the operating characteristics of the batteries was determined by comparing their voltage-current joint time series.

[0053] Extracting voltage data matrix The As the first The voltage-time series of battery number 1 is denoted as vector. Extracting the current data matrix The As the first The current-time series of battery number 1 is denoted as vector. The voltage and current sequences are concatenated to form a joint feature sequence. This sequence can comprehensively reflect the voltage and current response characteristics of the battery. Similarly, the first... The joint characteristic sequence of battery number 1 is denoted as: .

[0054] The Dynamic Time Warping (DTW) algorithm is used to calculate the first... Battery No. 1 and No. 2 The combined sequence of battery number 1 and The distance between them. Dynamic Time Warping (DTW) algorithms, by constructing a cumulative distance matrix, find the optimal time alignment path between two sequences, handling scaling and offset of time series on the time axis. The DTW distance output by the algorithm is denoted as... The smaller the distance, the more similar the operating characteristics of the two batteries.

[0055] Based on the dynamic time warping distance, a Gaussian kernel function is used to convert the distance into similarity:

[0056] scale parameter Preferably, the value is a comprehensive measure of the normal fluctuation range of battery voltage and current, and this conversion makes the similarity value range from 0 to 1.

[0057] To control the sparsity of the temporal performance relation network, a similarity threshold is set for filtering. The similarity threshold is then defined. Preferably The adjacency matrix elements of the time-series performance relation network are determined according to the following rules:

[0058] Two-layer network fusion A comprehensive relational network is formed by weighted fusion of the spatial physical relational network and the temporal performance relational network. The comprehensive relational network uses an adjacency matrix. The fusion formula is as follows:

[0059] in This is the fusion weight parameter, with a value ranging from 0 to 1. Fusion Weight This reflects the relative importance of spatial and temporal relationships in health assessment.

[0060] Fusion weights An adaptive method is used to determine the value. The average connection strength (average of non-zero elements) of the spatial relation network and the temporal relation network is calculated and denoted as . and The fusion weights are set as follows:

[0061] This setting allows subnetworks with higher connection strength to receive greater fusion weight.

[0062] S3 Network Feature Extraction Based on the constructed two-layer relationship network and voltage, current, and temperature data, a multi-dimensional feature vector characterizing the battery's health state is extracted. Feature extraction includes four categories: local consistency features, dynamic response features, electrical coupling features, and network topology features.

[0063] (1) Local consistency feature Local consistency features measure the degree of parameter differences between an individual cell and its associated cells. For the first... Battery number, from the adjacency matrix of the comprehensive relational network. Extract the first The row contains non-zero elements corresponding to the first row. The set of associated batteries for battery number 1 is denoted as . .

[0064] Calculate the first Voltage difference between battery number 1 and its associated battery :

[0065] in For the first The voltage value of battery number at the last moment of the current evaluation window. For the first The voltage value corresponding to battery number 1.

[0066] Similarly, calculate the current difference. and temperature difference The formulas for calculating current difference and temperature difference are the same as those for voltage difference, except that the voltage values ​​in the formulas are replaced with current values. and ) and temperature value ( and ).

[0067] Local consistency features include , and The three components reflect the degree to which the battery deviates from its neighbors in terms of voltage, current, and temperature, respectively.

[0068] (2) Dynamic response characteristics Dynamic response characteristics measure the changes in battery parameters over time. Extract the voltage data matrix. The As the first The complete voltage-time series of battery number [number] is obtained. The differences between adjacent time points in this series are calculated to obtain the voltage rate of change series. , its first The elements are:

[0069] Calculate statistical characteristics of the voltage change rate sequence, including the mean. Standard deviation Maximum value and minimum value The mean reflects the average rate of change of voltage, the standard deviation reflects the degree of fluctuation in the rate of change, and the maximum and minimum values ​​reflect extreme cases of change.

[0070] Similarly, based on the current data matrix The Row and temperature data matrix The Okay, calculate the current rate of change sequence and the temperature rate of change sequence respectively. The elements of the current rate of change sequence are... The elements of the temperature change rate sequence are Statistical characteristics were calculated for both sequences. The statistical characteristic of the rate of change of current is as follows: , , and The statistical characteristics of the rate of temperature change are: , , and .

[0071] The dynamic response characteristics include 12 statistical measures, specifically: mean of voltage change rate sequence, standard deviation of voltage change rate sequence, maximum value of voltage change rate sequence, minimum value of voltage change rate sequence, mean of current change rate sequence, standard deviation of current change rate sequence, maximum value of current change rate sequence, minimum value of current change rate sequence, mean of temperature change rate sequence, standard deviation of temperature change rate sequence, maximum value of temperature change rate sequence, and minimum value of temperature change rate sequence. Among them, the mean reflects the average rate of change of the parameter, the standard deviation reflects the degree of fluctuation of the rate of change, and the maximum and minimum values ​​reflect the extreme cases of change, comprehensively depicting the dynamic change law of the battery in the three dimensions of voltage, current, and temperature.

[0072] (3) Electrical coupling characteristics Electrical coupling characteristics utilize voltage and current data to extract physical quantities reflecting the battery's internal resistance and power characteristics. These characteristics directly indicate the battery's health status.

[0073] First, calculate the estimated instantaneous internal resistance. Select several time pairs within the evaluation window where the current change amplitude is large, and calculate the corresponding ratio of voltage change to current change to obtain multiple estimated instantaneous internal resistance values. Average these estimates to obtain the first estimated value. The average instantaneous internal resistance characteristics of the battery :

[0074] in For the selected time pairs, To prevent division by zero, a small constant (preferably taken as...) This characteristic reflects the relative level of the battery's internal resistance, and increased internal resistance is an important indicator of battery aging.

[0075] Next, the power characteristics are calculated. The voltage time series and current time series are multiplied point-by-point to obtain the power time series. Battery No. 1 in The instantaneous power at the time of the first acquisition is:

[0076] Calculate the mean and standard deviation of the power time series, denoted as , ... and :

[0077] Average power reflects the average power output level of the battery, while power standard deviation reflects power fluctuations.

[0078] The voltage-current response correlation characteristics were recalculated. The Pearson correlation coefficient was used to calculate the voltage time series. With current time series The degree of linear correlation between them is denoted as This coefficient reflects the degree of synchronization between voltage and current changes, and its value ranges from -1 to 1. Healthy batteries exhibit a good linear relationship between voltage and current responses, while aging batteries may show response misalignment, causing the correlation coefficient to deviate from the normal range.

[0079] Electrical coupling characteristics include , , and These four components directly reflect the battery's health status from an electrophysical perspective.

[0080] (4) Network topology characteristics Network topology features measure the importance of a battery's position within a relational network. Calculate the... Degree centrality of battery number adjacency matrix No. Number of non-zero elements in a row:

[0081] Calculate weighted degree centrality adjacency matrix No. The sum of all elements in the row:

[0082] Calculate clustering coefficient This reflects the tightness of the connection between the battery's neighbors. For the first... The clustering coefficient of battery number 1 is defined as:

[0083] in For the first The number of neighbors of battery number 1 This represents the actual number of connections between these neighbors. If... Then define A higher clustering coefficient indicates a denser local network where the battery resides.

[0084] Network topology features include , and Three components.

[0085] Feature vector combination Combining the above four types of features forms the first... Complete feature vector of battery number 1 :

[0086] The feature vector contains 22 components, which characterize the battery health status from four dimensions: consistency, dynamism, physicality, and network properties. The feature vector is normalized using a minimum-maximum normalization method to unify the value range of each component to between 0 and 1.

[0087] S4 Health Score A graph neural network model is used to map the extracted feature vectors into a health score. The health score in this invention comprehensively reflects the degree of degradation in three aspects: capacity decay, internal resistance increase, and consistency degradation of the battery. The score range is from 0 to 100. A score between 90 and 100 indicates that the battery is in excellent condition, with a capacity retention rate higher than 95%, internal resistance increase less than 10%, and good consistency with neighboring batteries. A score between 80 and 90 indicates that the battery is in good condition, with a capacity retention rate between 90% and 95%, internal resistance increase between 10% and 20%, and slight deviations in consistency beginning to appear. A score between 70 and 80 indicates that the battery is in average condition, with a capacity retention rate between 85% and 90%, internal resistance increase between 20% and 30%, and consistency deviations requiring attention. A score between 60 and 70 indicates that the battery is in poor condition, with a capacity retention rate between 80% and 85%, internal resistance increase between 30% and 50%, and significantly deteriorated consistency, requiring prompt maintenance. A score below 60 indicates that the battery is in poor condition, with a capacity retention rate of less than 80% or an internal resistance increase of more than 50%, posing a safety risk and requiring immediate handling or replacement.

[0088] The input to the graph neural network is the feature vector matrix of all batteries. (dimension is) Adjacency matrix of comprehensive relational networks The output is a health score vector for each battery. .

[0089] The graph neural network consists of four graph convolutional layers, with feature dimensions of 64, 128, 64, and 32 respectively. Let... Let be the initial characteristic matrix, and be the th . The formula for calculating layer graph convolution is:

[0090] in For the first The feature matrix of the layer ( ), The weight matrix is ​​learnable. Exponential linear unit activation function. Defined as:

[0091] in These are the default parameters. Graph convolution uses the adjacency matrix. Enables information propagation and feature aggregation between nodes.

[0092] The graph convolutional layer is followed by two fully connected neural network layers with 16 and 1 neurons respectively. Let the fourth graph convolutional layer be the... The feature vector of battery number 1 is (i.e., matrix) The (Line). The last layer uses the sigmoid activation function to map the output value to the range of 0 to 1, and multiplies it by 100 to get a health score of 0 to 100.

[0093] in, and These are the weight matrix and bias term of the output layer, respectively.

[0094] The graph neural network model is trained using historical data. The training data includes operational data and corresponding health labels for batteries in different health states. Training employs the backpropagation algorithm and the Adam optimizer, with a learning rate of 0.001, a batch size of 32, and 200 training epochs. The loss function is the mean squared error.

[0095] in The number of training samples. To predict health status, These are the actual health status labels. The trained model is used for real-time health status assessment.

[0096] The present invention will be further described below with reference to specific embodiments.

[0097] This embodiment uses a lithium iron phosphate battery pack, containing 200 individual cells, each with a rated voltage of 3.2V and a rated capacity of 150Ah. The cells are arranged in a 10x20 array, with a row and column spacing of 5cm. The battery pack uses a hybrid series-parallel connection, with 10 cells connected in series to form one battery string, for a total of 20 series-parallel strings.

[0098] The data acquisition system employs a high-precision multi-channel synchronous acquisition card, with voltage measurement accuracy of ±0.01% and temperature measurement accuracy of ±0.1°C. Voltage sensors are connected to the positive and negative terminals of each individual cell, while temperature sensors are positioned at the center of each cell's surface. The central processing unit utilizes the NVIDIA Jetson AGX Xavier edge computing platform, equipped with a 512-core Volta GPU and 32GB of memory, with the software environment based on the PyTorch Geometric framework.

[0099] (1) Construction of training dataset Training the graph neural network model requires a labeled historical dataset. The training data comes from actual operating data of 15 different energy storage power stations, covering battery samples from different manufacturers, batches, operating conditions, and aging levels.

[0100] (2) Data collection and labeling Data acquisition is conducted during the normal operation of the energy storage power station and does not affect system operation. For the battery packs participating in data acquisition, voltage, temperature, and current data are recorded within a 3-minute window according to the data acquisition strategy of this invention, forming an operational data sample. Each sample contains complete operational status data of a battery pack at a specific moment.

[0101] Health labels are obtained through offline testing. During the same period of collecting operational data, offline capacity and internal resistance tests are performed on the battery pack. Capacity testing uses a standard constant current charge-discharge method, charging the battery from an empty state at a 0.5C rate, then discharging it to the cutoff voltage at a 0.5C rate. The ratio of the measured discharge capacity to the rated capacity is the capacity retention rate. Internal resistance testing uses the AC impedance method; the ratio of the current internal resistance to the initial internal resistance minus 1 is the internal resistance growth rate.

[0102] The health rating is calculated based on a combination of capacity retention and internal resistance growth rate. The formula is: Health Rating = 60 multiplied by Capacity Retention Rate + 40 multiplied by (1) minus Internal Resistance Growth Rate divided by 1.5 (1) multiplied by 100. In this formula, capacity decay has a weight of 60%, and internal resistance growth has a weight of 40%. For example, a battery with a capacity retention rate of 0.85 and an internal resistance growth rate of 0.3 has a health rating of 83.

[0103] (3) Data set size and distribution The training dataset contains 800,000 samples. The samples come from 15 energy storage power stations ranging in size from 500kW to 10MW. The battery manufacturers include five mainstream companies such as CATL, BYD, and Guoxuan High-Tech. The battery life ranges from brand new to 5 years, and the charge / discharge rate ranges from 0.2C to 1C.

[0104] The health scores of the samples ranged from 20 to 100. Samples with a health score between 90 and 100 accounted for 18%, primarily from new battery packs less than one year old. Samples with a health score between 80 and 90 accounted for 22%, primarily from battery packs one to two years old. Samples with a health score between 70 and 80 accounted for 28%, primarily from battery packs two to three years old. Samples with a health score between 60 and 70 accounted for 20%, primarily from battery packs three to four years old. Samples with a health score below 60 accounted for 12%, primarily from battery packs more than four years old or showing significant degradation.

[0105] The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The training set (640,000 samples) was used for model parameter learning, the validation set (80,000 samples) for hyperparameter tuning and early stopping detection, and the test set (80,000 samples) for final performance evaluation. This partitioning process ensured that samples from the same battery pack within the same time period would not appear in both the training and test sets simultaneously, preventing data leakage.

[0106] (4) Model training and evaluation The graph neural network model was trained on the training set. Training employed the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and a total of 200 training epochs. The loss function was the mean squared error, calculated as the sum of squared differences between the predicted health score and the labeled health score. An early stopping mechanism was used during training; training ceased when the validation set loss did not decrease for 10 consecutive epochs.

[0107] The model performance was evaluated on an independent test set. The mean absolute error was 2.1 points, meaning the average deviation between the predicted and actual health levels was 2.1 points. The root mean square error was 3.4 points. The correlation coefficient was 0.96, indicating a high correlation between the predicted and actual values. Health levels were divided into 5 grades of 20 points each, with a classification accuracy of 94.8%, meaning the consistency between the system's assessed health level and the actual health level was 94.8%.

[0108] (5) Verification of the construction of a two-layer relationship network The construction process of the two-layer relationship network is verified using a real-world evaluation as an example. The evaluation object is a battery pack that has been running for 18 months and has accumulated 2500 charge-discharge cycles. During the evaluation, it is in a 0.5C constant current discharge state at an ambient temperature of 25°C.

[0109] A spatial physical relationship network was constructed according to the detailed technical solution. Based on the 10x20 battery arrangement and 5cm spacing, the Euclidean distance between all battery pairs was calculated. An adjacency threshold of 7.5cm was set, identifying an average of 3.8 physical neighbors per battery. Based on the circuit connection topology, series and parallel connections were identified, with an average of 2.4 electrically connected neighbors per battery. The resulting spatial physical relationship network shows an average of 6.2 spatial neighbors per battery.

[0110] A time-series performance relationship network was constructed according to the detailed technical solution. Voltage time series of all batteries within a 3-minute window were extracted, and the similarity between sequences was calculated using a dynamic time warping algorithm. A similarity threshold of 0.7 was set, identifying an average of 11.6 time-series neighbors for each battery. These neighbors were mainly distributed among batteries with similar performance in parallel branches, verifying that the time-series relationship network can capture the similarity of operating characteristics.

[0111] Calculate the fusion weights. The average connection strength of the spatial relationship network is 0.458, and the average connection strength of the temporal relationship network is 0.850. The adaptive fusion weights are 0.35 to 0.65. The fused integrated relationship network retains the constraints of physical space and electrical connections while highlighting the correlation of performance similarity, with an average of 15.3 integrated neighbors per cell.

[0112] (6) Examples of health assessment Based on the constructed comprehensive relationship network, feature vectors of all 200 batteries are extracted. Each feature vector contains 13 components: 2 components of local consistency features, 8 components of dynamic response features, and 3 components of network topology features. The normalized feature vectors are then input into the trained graph neural network model.

[0113] The model outputs health scores for 200 batteries. The scores range from 62.5 to 85.3, with an average score of 76.8. Battery #156 scored 65.2 and battery #189 scored 67.8, both below 70, and were marked as high-priority targets by the system. Offline testing was performed on these two batteries. Battery #156 showed a capacity retention rate of 0.82 and an internal resistance growth rate of 0.35, resulting in an actual health score of 66.1. Battery #189 showed a capacity retention rate of 0.84 and an internal resistance growth rate of 0.32, resulting in an actual health score of 69.2. The deviations between the system evaluation results and the offline testing results were 0.9 and 1.4 points, respectively, validating the accuracy of the evaluation.

[0114] (7) System deployment and long-term verification The trained model was deployed to three energy storage power plants of different sizes for long-term validation. The first power plant had a scale of 1MW / 2MWh, the second had a scale of 5MW / 10MWh, and the third had a scale of 500kW / 1MWh, with a cumulative deployment and operation time of more than 18 months.

[0115] In the application of the 1MW / 2MWh power plant, the system performs a health assessment every hour, with each assessment taking 8.5 seconds. The system operated continuously for 6 months without any failures. During this period, 12 manual spot checks were conducted, and batteries with low health identified by the system were tested offline, with a verification consistency rate of 96.2%.

[0116] The system maintains stable performance under various environmental conditions. Within an ambient temperature range of -15°C to 45°C, the system's assessment accuracy varies by less than 5%. The system's assessment accuracy remains consistent across different charge / discharge rates from 0.2C to 1C. The system can accurately identify differences in health status in battery packs with varying degrees of aging.

[0117] Long-term verification shows that the two-layer relationship network method of the present invention can accurately assess the health of energy storage batteries. Compared with the traditional method, the assessment accuracy is reduced from 8.3% error to 2.1%, and the early warning time is 15 to 20 minutes earlier, providing effective support for the health management and preventive maintenance of energy storage systems.

[0118] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the health of energy storage batteries, characterized in that, The method includes the following steps: Based on the physical location and electrical connection of the batteries in the battery pack to be evaluated, a spatial physical relationship network between the batteries is established; the operational similarity between time series of different battery parameters is calculated to establish a temporal correlation network between batteries with similar operational characteristics; the spatial physical relationship network and the temporal correlation network are coupled to form a comprehensive relationship network of the battery pack to be evaluated and the adjacency matrix corresponding to the comprehensive relationship network. The comprehensive relationship network and adjacency matrix of the battery pack under evaluation are used to extract the degree of parameter difference between each battery and its associated batteries, i.e., local consistency characteristics; and the dynamic response characteristics of the rate and pattern of change of each battery parameter over time. Electrical coupling characteristics reflecting the battery's internal resistance and power characteristics; and network position characteristics reflecting the importance of each battery in the comprehensive relationship network; these local consistency characteristics, dynamic response characteristics, electrical coupling characteristics, and network position characteristics are combined to form the feature vector of the battery pack to be evaluated. The feature vectors and adjacency matrices are input into a graph neural network model to obtain the health score of each battery in the battery pack to be evaluated.

2. The method for evaluating the health of an energy storage battery as described in claim 1, characterized in that, The adjacency matrix is ​​calculated using the following formula: in, To integrate weight parameters, This is the adjacency matrix of the comprehensive relational network. This is the adjacency matrix of a spatial physical relation network. This is the adjacency matrix of the time-series performance relation network. For space physical relationship network The average of all non-zero elements in the set. For time-series performance relational network The average of all non-zero elements in the set.

3. The method for evaluating the health of an energy storage battery as described in claim 2, characterized in that, The and The calculation formula is as follows: in, The adjacency matrix of the spatial physical relation network is the first... Line 1 The element of the column represents the first element. Battery No. 1 and No. 2 The strength of the spatial relationship between battery numbers To establish the strength of the fixed relationship between battery pairs connected in series. For the strength of the fixed relationship between the parallel battery pairs, The relationship strength weight is calculated based on physical distance. Let be the Euclidean distance between the two batteries. This is the threshold for spatial adjacency determination. The adjacency matrix of the time-series performance relation network is the first... Line 1 The element of the column represents the first element. Battery No. 1 and No. 2 Strength of time-series performance relationship between cells For the first Battery No. 1 and No. 2 Similarity of operating characteristics between battery numbers This is the similarity threshold.

4. A method for evaluating the health of an energy storage battery as described in claim 1 or 3, characterized in that, The local consistency features include voltage difference values, current difference values, and temperature difference values, wherein the voltage difference value is calculated using the following formula: in, For the first The weighted average voltage difference between battery number 1 and its associated batteries. For the first The set of associated batteries for battery number 1, i.e., the adjacency matrix of the comprehensive relationship network. No. The set of battery numbers corresponding to the non-zero elements in the row. The adjacency matrix of the comprehensive relational network is the first... Line 1 The element of the column represents the first element. Battery No. 1 and No. 2 The strength of the relationship between the number of batteries For the first The voltage value of battery number at the last moment of the current evaluation window. For the first The voltage value of battery number at the last moment of the current evaluation window. This represents the number of times the voltage data was collected.

5. A method for evaluating the health of an energy storage battery as described in claim 1 or 3, characterized in that, The dynamic response characteristics include: the mean of the voltage rate of change sequence, the standard deviation of the voltage rate of change sequence, the maximum value of the voltage rate of change sequence, the minimum value of the voltage rate of change sequence, the mean of the current rate of change sequence, the standard deviation of the current rate of change sequence, the maximum value of the current rate of change sequence, the minimum value of the current rate of change sequence, the mean of the temperature rate of change sequence, the standard deviation of the temperature rate of change sequence, the maximum value of the temperature rate of change sequence, and the minimum value of the temperature rate of change sequence.

6. A method for evaluating the health of an energy storage battery as described in claim 1 or 3, characterized in that, The electrical coupling characteristic comprises four components, each of which is calculated using the following formula: in, For the first Estimated average instantaneous internal resistance of battery number 1. For the selected time pairs, For the first Battery No. 1 in Voltage value at the time of the first acquisition For the first Battery No. 1 in Voltage value at the time of the first acquisition For the first Battery No. 1 in Current value at the time of the first acquisition For the first Battery No. 1 in Current value at the time of the first acquisition To prevent small constants from being divided by zero, For the first Battery No. 1 in Instantaneous power at the time of the first acquisition For the first Battery No. 1 in Voltage value at the time of the first acquisition For the first Battery No. 1 in Current value at the time of the first acquisition The number of times the voltage data was collected. This refers to the sequence number of the time collection point. For the first The average power of battery number 1 For the first The power standard deviation of battery number 1 For the first Battery No. 1 in Instantaneous power at the time of the first acquisition This represents the number of times the voltage data was collected.

7. A method for evaluating the health of an energy storage battery as described in claim 1 or 3, characterized in that, The network location features include Their respective formulas are as follows: in, For the first The degree centrality of cell A represents the number of its neighbors; To meet Battery serial number A set; The cardinality (radix) represents the number of elements in the set. The adjacency matrix of the comprehensive relational network is the first... Line 1 Column elements, For the first The weighted degree centrality of cell A represents the sum of the relationship strengths between cell A and all its neighbors. This represents the total number of individual cells in the battery pack. For the first The clustering coefficient of cell number indicates the degree of connectivity between its neighbors. For the first The actual number of connections between the neighbors of battery number 1; For the first The number of neighbors of battery number 1, if Then define .

8. The method for evaluating the health of an energy storage battery as described in claim 1, characterized in that, The graph neural network comprises four graph convolutional layers, and the formula for calculating the health score in this graph neural network is as follows: The loss function for a graph neural network is: in, and These are the weight matrix and bias term of the output layer, respectively. The fourth convolutional layer is the first... The feature vector of battery number 1 The training sample is numbered. This represents the total number of training samples. To predict health status, For true health status labeling, For the first The health score of each battery.

9. A health evaluation system for energy storage batteries, characterized in that, The system includes an actuator for performing a method for evaluating the health of an energy storage battery as described in any one of claims 1-8.

10. A computer storage medium having a computer program stored thereon, characterized in that, When executed, the computer program is used to implement the energy storage battery health evaluation method according to any one of claims 1-8.

Images