Container energy storage system health state prediction method, device and program product

Abstract

The present application relates to the technical field of energy storage system, and discloses a container energy storage system health state prediction method, device and program product, comprising: performing physical topology modeling on the container energy storage system to obtain a multi-dimensional weight adjacency matrix; based on the multi-dimensional weight adjacency matrix and historical temperature data set of the container energy storage system, performing multi-objective optimization by using an improved non-dominated sorting genetic algorithm, and determining master cell groups and slave cell groups of the container energy storage system; according to the master cell groups and the slave cell groups, obtaining original signals of the container energy storage system; based on the original signals, processing by an anti-aliasing filtering method and a signal preprocessing method, and establishing a target master cell feature matrix; based on the target master cell feature matrix and the multi-dimensional weight adjacency matrix, performing space-time graph neural network prediction to obtain slave cell health state prediction results of the container energy storage system, improve the prediction error stability and accuracy, and reduce the cost.

Description

Technical Field

[0001] This invention relates to the field of energy storage system technology, specifically to a method, device, and program product for predicting the health status of a containerized energy storage system. Background Technology

[0002] In the field of cell-level health status monitoring, the most widely adopted approach is the all-sensor monitoring scheme, which deploys current, voltage, and temperature sensors for each cell to predict state of health (SOH). However, this approach leads to an exponential increase in hardware costs: more than 9,600 sensors are required in a 3,200-cell system, accounting for more than 35% of the total cost. More seriously, this approach generates tens of thousands of data channels, resulting in communication delays in the battery management system exceeding 200 milliseconds, far exceeding real-time control requirements. The root cause of these shortcomings is that this method fails to consider the physical interrelationships of cell clusters—in fact, in containerized energy storage systems, there is significant spatial thermal conduction coupling between cells (temperature gradient less than 8 ℃ / m) and consistent current distribution in the electrical topology (parallel branch current deviation less than 3%), but these interrelationships are completely ignored.

[0003] Another common solution is a sampling monitoring approach, which involves randomly deploying sensors on approximately 10% of the battery cells. However, this approach suffers from severe spatial bias: because the sampling does not cover high-risk areas for thermal runaway (e.g., corner cells age 15-20% faster than central cells), prediction errors exceed 7%. Furthermore, this method ignores the current interactions between parallel branches, causing a SOC estimation offset greater than 8%, and the error accumulates over time—after 200 charge-discharge cycles, the prediction error expands to 3.2 times the initial value. The root cause lies in the use of a static independent sampling model, which fails to address the propagation effects of cell aging in the spatiotemporal dimensions and the state migration mechanism of sensor data to non-monitoring points.

[0004] The third technical approach is a purely data-driven solution. This method utilizes deep learning networks such as LSTM to predict State of Health (SOH), but it requires input of all battery cell data. This leads to a data dimensionality disaster—when the number of cells exceeds 500, the number of model parameters explodes to the billion level, making it difficult to deploy in engineering practice. More specifically, this method has extremely high hardware requirements, requiring GPU servers to run, exceeding the capacity of embedded devices in battery management systems; the prediction error can reach 12% under complex conditions such as low temperatures, because it completely deviates from the constraints of electrochemical aging mechanisms (such as overpotential rise, SEI film thickening, and other physical processes); at the same time, this approach has weak transferability, requiring model retraining every time a new batch of batteries is replaced. These problems essentially stem from simply treating the prediction problem as a pure mathematical mapping, ignoring the physical interpretability of electrochemical aging mechanisms and the state correlation characteristics under the three-dimensional spatial constraints of energy storage systems.

[0005] In summary, the core contradiction facing existing technologies is that achieving accurate prediction requires deploying a massive number of sensors, leading to a sharp increase in costs; while reducing costs would significantly sacrifice prediction accuracy. This contradiction has become a key technological bottleneck restricting the commercialization of large-scale containerized energy storage systems. Summary of the Invention

[0006] This invention provides a method, device, and program product for predicting the health status of a containerized energy storage system, in order to solve the problem that existing cell-level health status monitoring technologies cannot balance accurate prediction and cost reduction.

[0007] In a first aspect, the present invention provides a method for predicting the health status of a containerized energy storage system, the method comprising:

[0008] A physical topology model of the containerized energy storage system is performed to obtain a multidimensional weighted adjacency matrix. This matrix encodes the spatial and electrical coupling relationships of the cell clusters within the system using three weighted terms: thermal conduction, electrical impedance, and current distribution. Based on the multidimensional weighted adjacency matrix and the historical temperature dataset of the system, an improved non-dominated sorting genetic algorithm is used for multi-objective optimization to determine the master and slave cell groups. The original signals of the system are then obtained based on these groups. These signals are processed using anti-aliasing filtering and signal preprocessing methods to establish a target master cell feature matrix. Finally, based on the target master cell feature matrix and the multidimensional weighted adjacency matrix, a spatiotemporal graph neural network is used to predict the health status of the slave cells, incorporating electrochemical mechanism constraints.

[0009] The health status prediction method for containerized energy storage systems provided by this invention encodes the spatial and electrical coupling relationship of the cell clusters in the containerized energy storage system and determines a multi-dimensional weighted adjacency matrix through three weighted terms: thermal conduction, electrical impedance, and current distribution. This overcomes the limitation of traditional all-sensor schemes that treat cells as independent units. Furthermore, by combining the multi-dimensional weighted adjacency matrix and using an improved non-dominated sorting genetic algorithm for multi-objective optimization to determine the master-slave cell group, the spatial bias of traditional sampling monitoring and the cost increase of all-sensor schemes are solved. Furthermore, by acquiring raw signals according to the master-slave cell group, differentiated data acquisition is achieved, reducing unnecessary sensor deployment, lowering hardware costs and data transmission pressure, while ensuring the integrity of core status data. Furthermore, by combining anti-aliasing filtering methods and signal preprocessing methods to establish the target master cell feature matrix, signal noise and dimensional influences are eliminated, enabling the extraction of core features strongly correlated with the cell health status. This provides high-quality input for neural network prediction and helps improve prediction accuracy. Furthermore, by incorporating a spatiotemporal graph neural network constrained by electrochemical mechanisms for prediction, the shortcomings of purely data-driven models that are detached from electrochemical mechanisms are overcome. This enables high-precision predictions on resource-constrained BMS embedded devices and solves the problem of excessive prediction errors under complex operating conditions. Therefore, by implementing this invention, the stability and accuracy of prediction errors in large-scale energy storage systems are improved, costs are reduced, and a balance between accurate prediction and cost reduction is achieved.

[0010] In one optional implementation, a physical topology model of the containerized energy storage system is performed to obtain a multidimensional weighted adjacency matrix, including:

[0011] The system obtains the absolute position coordinates of the center point of each cell in the containerized energy storage system and establishes a distance matrix; it obtains multiple distribution ratios of the current in multiple branches under the initial static conditions in the parallel branch connection topology diagram of the containerized energy storage system, which reflects the circuit connection between cells in the containerized energy storage system; it applies a preset AC excitation signal to the current in each branch in the parallel branch connection topology diagram and obtains multiple AC impedance phase differences between different adjacent cells; and it constructs a multi-dimensional weighted adjacency matrix based on the distance matrix, multiple distribution ratios, and multiple AC impedance phase differences.

[0012] The health status prediction method for containerized energy storage systems provided by this invention accurately quantifies the spatial distance relationship between battery cells by acquiring the absolute position coordinates of the cells and establishing a distance matrix. Furthermore, by acquiring the branch current distribution ratio, the current balance characteristics of the electrical topology between cells are clarified, providing a basis for calculating the current distribution weight term and helping to reduce prediction deviations caused by differences in current distribution. Furthermore, by applying AC excitation and acquiring the AC impedance phase difference, the degree of electrical characteristic coupling between cells can be captured, supplementing the electrical dimension correlation information and helping to improve the comprehensiveness of the multi-dimensional weighted adjacency matrix. Furthermore, by integrating the distance matrix, multiple distribution ratios, and multiple AC impedance phase differences to construct a multi-dimensional weighted adjacency matrix, the quantitative encoding of the physical relationship between the cells is realized, laying the foundation for subsequent main cell selection and spatiotemporal feature learning of the neural network.

[0013] In one optional implementation, based on a multidimensional weighted adjacency matrix and the historical temperature dataset of the containerized energy storage system, an improved non-dominated sorting genetic algorithm is used for multi-objective optimization to determine the master and slave cell groups of the containerized energy storage system, including:

[0014] Based on the multidimensional weighted adjacency matrix and the historical temperature dataset of the container energy storage system, a multi-objective optimization function is determined. The multi-objective optimization function is used to minimize the thermal characterization error and the topological similarity error. An improved non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function and determine the master cell group and slave cell group of the container energy storage system.

[0015] The health status prediction method for containerized energy storage systems provided by this invention ensures that the selected main battery cell group can accurately reflect the temperature change law of the entire system and retain the core physical correlation characteristics by determining a multi-objective optimization function that minimizes thermal characterization error and topological similarity error, thus avoiding information loss caused by sampling. Furthermore, by using an improved non-dominated sorting genetic algorithm, a Pareto optimal subset of main battery cells can be efficiently selected, taking into account both the optimization objective and constraints, solving the randomness and static defects of traditional sampling, and improving the representativeness of the main battery cells.

[0016] In one optional implementation, based on the original signal, after processing with anti-aliasing filtering and signal preprocessing methods, a target main cell feature matrix is ​​established, including:

[0017] Based on preset differential filtering parameters, anti-aliasing filtering is performed on the main cell group signal and the slave cell group signal in the original signal to obtain the initial main cell group signal and the initial slave cell group signal. Based on the initial main cell group signal and the initial slave cell group signal, the signal preprocessing method is used to establish the initial main cell feature matrix. The initial main cell feature matrix is ​​normalized to obtain the target main cell feature matrix.

[0018] The health status prediction method for containerized energy storage systems provided by this invention applies anti-aliasing filtering to the master and slave cell signals separately according to preset differential parameters. This adapts to the different sampling frequencies and signal characteristics of the master and slave cells, effectively eliminating frequency aliasing distortion and ensuring the purity of both types of signals. Furthermore, signal preprocessing is used to process the signals and extract key physical features related to cell aging to establish an initial master cell feature matrix, transforming the original signal into effective features. Additionally, normalization eliminates dimensional differences between features, which helps accelerate the convergence speed of the neural network and avoids prediction bias caused by inconsistent feature scales.

[0019] In one optional implementation, based on the initial master cell group signal and the initial slave cell group signal, an initial master cell feature matrix is ​​established after signal preprocessing, including:

[0020] Wavelet denoising is performed on the initial main cell group signal to obtain the target main cell group signal; moving average filtering is performed on the initial slave cell group signal to obtain the target slave cell group signal; feature extraction is performed based on the target main cell group signal and the target slave cell group signal to obtain multiple feature parameters; and an initial main cell feature matrix is ​​established based on the multiple feature parameters.

[0021] The health status prediction method for containerized energy storage systems provided by this invention, through wavelet denoising of the initial main cell group signal, can accurately retain the key features of the main cells while eliminating noise interference, ensuring the accuracy of core features. Furthermore, by applying moving average filtering to the initial slave cell group signal, noise in the slave cell voltage signal is smoothed in a lightweight manner, balancing processing efficiency and signal stability, meeting the basic requirements for deviation detection. Moreover, by accurately capturing key characterization indicators of cell health status based on the target signal and constructing an initial main cell feature matrix, the correlation between features and state of health (SOH) is improved.

[0022] In one optional implementation, based on the target main cell feature matrix and the multidimensional weighted adjacency matrix, the pre-prediction result of the main cell health status of the containerized energy storage system is obtained through spatiotemporal graph neural network prediction, including:

[0023] The target main cell feature matrix is ​​input into the physical feature embedding layer of the spatiotemporal graph neural network for processing to obtain the physical embedding feature matrix. The physical embedding feature matrix and the preset heat dissipation system parameter set are input into the thermal constraint 3D convolutional layer of the spatiotemporal graph neural network for processing to obtain the spatial feature matrix that fuses the thermal distribution correlation of the main cell. The spatial feature moments and the multidimensional weighted adjacency matrix are input into the current balance graph attention layer of the spatiotemporal graph neural network for processing to obtain the electrical feature matrix that fuses the spatial thermal features and electrical balance constraints. The physical embedding feature matrix, the spatial feature matrix, and the electrical feature matrix are input into the feature fusion layer of the spatiotemporal graph neural network for processing to obtain the fused feature matrix. The fused feature moments are input into the output layer of the spatiotemporal graph neural network and processed using the state transition operator to obtain the prediction result of the cell health status of the container energy storage system.

[0024] The health status prediction method for containerized energy storage systems provided by this invention, through physical feature embedding layer processing, can transform electrochemical mechanism parameters into features that can be processed by neural networks, injecting physical interpretability and avoiding the mechanism detachment problem of purely data-driven models. Furthermore, through thermally constrained 3D convolutional layer processing to obtain a spatial feature matrix, it simulates the three-dimensional spatial heat conduction law, strengthens the learning of heat dissipation path-related features, and solves the defect of traditional models ignoring the influence of heat distribution. Furthermore, through current balance graph attention layer processing, it can capture the influence of uneven current distribution in parallel branches, suppress the interference of electrical imbalance on prediction, and improve the accuracy of electrical topology-related predictions. Furthermore, through feature fusion layer, it integrates physical, spatial, and electrical multi-dimensional features and extracts cross-dimensional key correlation information, providing comprehensive and integrated feature support for SOH prediction. Finally, the output layer uses state transition operators to establish a precise mapping between main cell features and slave cell health status, thereby enabling efficient inference from a small amount of main cell data to a large amount of slave cell SOH, reducing dependence on full data.

[0025] In one alternative implementation, the method further includes:

[0026] Based on the predicted health status of the battery cells, multiple predicted voltage values ​​of the battery cells are calculated. Based on the multiple predicted voltage values ​​and the multiple measured voltage values ​​of the battery cells, multiple voltage prediction deviation rates are determined. Based on the multiple voltage prediction deviation rates, it is determined whether to trigger correction. When correction is triggered, the key aging parameters of the main battery cells are measured by AC impedance spectroscopy, and the impedance fitting residual is determined. Based on the multiple voltage prediction deviation rates and the impedance fitting residual, the state transition operator is updated.

[0027] The health status prediction method for containerized energy storage systems provided by this invention calculates the predicted cell voltage value based on the SOH prediction result, establishing a correlation between SOH and voltage. This provides a quantifiable comparative indicator for deviation detection, facilitating real-time monitoring of prediction accuracy. Furthermore, by determining the voltage prediction deviation rate, the difference between the predicted and measured values ​​is precisely quantified, providing a clear basis for correction triggering and helping to promptly detect prediction drift issues. Moreover, by combining the voltage prediction deviation rate to determine whether correction should be triggered, dynamic triggering of correction can be achieved, avoiding unnecessary calibration operations while ensuring that the prediction error does not exceed the allowable range. Furthermore, after triggering correction, key aging parameters of the main cell are measured using AC impedance spectroscopy, and the impedance fitting residual is determined. The state transition operator is updated by combining the voltage prediction deviation rate and the impedance fitting residual, realizing dynamic correction of the master-slave cell state mapping relationship. This avoids the blindness of pure data correction, solves the problem of prediction error accumulation during long-term operation, ensures prediction stability throughout the entire life cycle, and does not require interruption of system operation.

[0028] In a second aspect, the present invention provides a health status prediction device for a containerized energy storage system, the device comprising:

[0029] The modeling module performs physical topology modeling of the containerized energy storage system, obtaining a multi-dimensional weighted adjacency matrix. This matrix encodes the spatial and electrical coupling relationships of the cell clusters within the system using three weighted terms: thermal conduction, electrical impedance, and current distribution. The optimization module performs multi-objective optimization based on the multi-dimensional weighted adjacency matrix and the historical temperature dataset of the containerized energy storage system, utilizing an improved non-dominated sorting genetic algorithm to determine the master and slave cell groups. The acquisition module acquires the raw signals of the containerized energy storage system based on the master and slave cell groups. The processing module, based on the raw signals, processes them using anti-aliasing filtering and signal preprocessing methods to establish the target master cell feature matrix. The prediction module, based on the target master cell feature matrix and the multi-dimensional weighted adjacency matrix, uses a spatiotemporal graph neural network to predict the health status of the slave cells in the containerized energy storage system. The spatiotemporal graph neural network incorporates electrochemical mechanism constraints.

[0030] Thirdly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the container energy storage system health status prediction method of the first aspect or any corresponding embodiment described above.

[0031] Fourthly, the present invention provides a computer program product, including computer instructions, which are used to cause a computer to execute the container energy storage system health status prediction method of the first aspect or any corresponding embodiment described above. Attached Figure Description

[0032] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of an application scenario according to an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the first process of the health status prediction method for a containerized energy storage system according to an embodiment of the present invention;

[0035] Figure 3 This is a flowchart of the main battery cell selection process according to an embodiment of the present invention;

[0036] Figure 4 This is a flowchart of signal preprocessing according to an embodiment of the present invention;

[0037] Figure 5 This is a spatiotemporal graph neural network architecture diagram according to an embodiment of the present invention;

[0038] Figure 6 This is a flowchart of graph attention layer calculation according to an embodiment of the present invention;

[0039] Figure 7 This is a diagram of the calibration system architecture according to an embodiment of the present invention;

[0040] Figure 8 This is a timing diagram of the operation flow according to an embodiment of the present invention;

[0041] Figure 9 This is a structural block diagram of a container energy storage system health status prediction device according to an embodiment of the present invention;

[0042] Figure 10 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0045] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0046] As an optional application scenario of this invention, the specific application environment architecture or specific hardware architecture on which the health status prediction method for containerized energy storage systems depends is described herein. For example... Figure 1 As shown, the architecture system may include at least one terminal device and at least one server. Figure 1 The system is illustrated in the example, which includes a computer 101, a mobile terminal 102, and a server 103, and the terminal devices such as the computer 101 and the mobile terminal 102 are connected to the server 103 through a network 110.

[0047] Specifically, the terminal device can be a smartphone, tablet, laptop, PDA, desktop computer, game console, smart TV, smart wearable device, in-vehicle terminal, VR (Virtual Reality) device, AR (Augmented Reality) device, etc. Server 103 can be a standalone physical server, a server cluster, a distributed system, or a cloud server providing cloud services. Network 110 can be a wired or wireless network, examples of which include, but are not limited to, the Internet, corporate intranet, local area network, wide area network, mobile communication network, and combinations thereof.

[0048] Based on the shortcomings of existing technologies pointed out in the background section, this application aims to systematically solve the following technical problems through a master-slave fusion prediction architecture:

[0049] 1. How to overcome the fundamental contradiction between cost and precision.

[0050] This solution addresses the technical challenge of reducing the number of sensors deployed from 100% to <5% in a 3200-cell-level containerized energy storage system while achieving a prediction error of ≤2.3%. It also eliminates the exponential increase in hardware costs (sensor costs account for >35% of existing solutions) and communication latency (>200 ms) caused by a full sensor monitoring scheme.

[0051] 2. How to overcome the accumulation of spatial bias in sampling monitoring.

[0052] To address the error amplification phenomenon (ΔSOH≥7%) that occurs after 200 cycles of random sampling schemes, a spatiotemporal correlation modeling method based on electrical-thermal coupling is proposed to solve the problem of prediction bias accumulation caused by neglecting spatial thermal conduction coupling (temperature gradient <8℃ / m) and current distribution differences (parallel branch deviation <3%).

[0053] 3. How to achieve high-dimensional prediction for resource-constrained edge devices.

[0054] Overcoming the computational complexity barrier (model parameter count > 10⁹) faced by pure data-driven solutions in >500 cell systems, we constructed a lightweight algorithm that can run on BMS embedded systems, eliminating dependence on GPU servers, while ensuring prediction errors < 5% under complex operating conditions such as low temperatures.

[0055] 4. How to establish a portable physical-data fusion mechanism.

[0056] Overcoming the shortcomings of existing methods in being unable to analyze the physical processes of cell aging (such as SEI film thickening and lithium deposition), a prediction framework that combines the interpretability of electrochemical mechanisms with the fitting capabilities of deep learning is established, reducing the cost of algorithm transfer and retraining across battery batches by more than 80%.

[0057] According to an embodiment of the present invention, a method for predicting the health status of a containerized energy storage system is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0058] This embodiment provides a method for predicting the health status of a containerized energy storage system, which can be used on the aforementioned mobile terminals, such as mobile phones and tablets. Figure 2 This is a flowchart of a health status prediction method for a containerized energy storage system according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps:

[0059] Step S201: Perform physical topology modeling on the container energy storage system to obtain a multi-dimensional weighted adjacency matrix.

[0060] In one optional embodiment, the containerized energy storage system refers to an energy storage device that integrates a large number of lithium-ion cells in a specific spatial layout and electrical topology within a container, and is equipped with a heat dissipation system and a battery management system (BMS) for energy storage and dispatch, and has the characteristics of large scale and modularity.

[0061] In one alternative embodiment, physical topology modeling represents the modeling process for quantifying the physical relationships between cells in a containerized energy storage system.

[0062] In an optional embodiment, the multidimensional weighted adjacency matrix is ​​used to encode the spatial and electrical coupling relationship of the cell cluster in the container energy storage system through three types of weight terms: thermal conduction, electrical impedance, and current distribution. It is represented as an N×N dimensional mathematical matrix (N is the total number of cells), as shown in the following relationship (1):

[0063] (1)

[0064] In the formula: Represents a multidimensional weighted adjacency matrix; Represents the Euclidean distance from the center point of the battery cell ( The spatial heat conduction correlation terms were calculated using an exponential decay function. Modeling thermal coupling effects, It is the distance characteristic constant; Based on high-frequency (≥1 kHz) AC impedance phase difference Electrical characteristic coupling terms; Based on the current distribution ratio of parallel branches Electrical topology associations.

[0065] Furthermore, through weighting coefficients ( , , Dynamically balancing the three types of correlations to generate a topological mapping matrix. .

[0066] Furthermore, this technology breaks through the limitations of traditional all-sensor solutions that treat battery cells as independent units. It explicitly encodes the thermal conduction and current interaction network between battery cells within the container, allowing the system to cover more than 95% of the cluster status information by selecting only the <5% battery cells with the highest system association information entropy as the main battery cells.

[0067] In one optional embodiment, by integrating physical information such as the spatial location, electrical topology (branch current distribution, impedance phase difference), and thermodynamic parameters of the cells in the containerized energy storage system, a multi-dimensional weighted adjacency matrix is ​​constructed. This matrix can transform the three types of coupling relationships between cells—thermal conduction, electrical impedance, and current distribution—into quantifiable matrix elements, laying a physical foundation for subsequent master-slave cell selection and high-precision health status prediction.

[0068] Step S202: Based on the multidimensional weighted adjacency matrix and the historical temperature dataset of the container energy storage system, multi-objective optimization is performed using an improved non-dominated sorting genetic algorithm to determine the master cell group and slave cell group of the container energy storage system.

[0069] In one optional embodiment, the improved non-dominated sorting genetic algorithm represents a multi-objective evolutionary algorithm optimized on the basis of the traditional non-dominated sorting genetic algorithm. It is used to dynamically repair the population through evolutionary operations such as binary encoding, selection, simulated binary crossover, and bit flip mutation, combined with constraints, to generate a Pareto front solution set, and finally select the optimal solution that meets the multi-objective optimization requirements. It can be adapted to the specific scenario of main cell selection in container energy storage systems.

[0070] In one optional embodiment, the main cell group represents a representative subset of cells selected from all cells in the containerized energy storage system, accounting for less than 5% of the total number of cells. This subset possesses high information entropy and can cover more than 95% of the system state information. Further, the secondary cell group represents the set of remaining cells in the total number of cells excluding the main cell group.

[0071] In one optional embodiment, the historical temperature dataset represents the set of temperature data of the entire system and each cell at different times during the operation of the containerized energy storage system, recorded by means of thermal imaging equipment, etc., to reflect the temperature field distribution law of the system, the temperature rise characteristics of the cells, and to locate the highest temperature area of ​​the system and the corresponding cell location.

[0072] In one alternative embodiment, thermal imaging data during system operation can be collected using devices integrated into the cooling system within the container system. This data is then combined with the airflow direction vector of the cooling system. (Typical Z-axis direction), heat dissipation coefficient Parameters such as (optimized values ​​for lithium iron phosphate batteries) are used to process thermal imaging data and establish a temperature field distribution model. Finally, the model outputs data on the spatial distribution of temperature, which is then compiled into a historical temperature dataset covering both the time dimension (continuous operation period) and the spatial dimension (location of all battery cells in the system).

[0073] In one optional embodiment, based on a multidimensional weighted adjacency matrix and a historical temperature dataset, an improved non-dominated sorting genetic algorithm is used to solve a multi-objective optimization problem that minimizes thermal characterization error and topological similarity error. Under the constraint that the number of main cells is ≤ 5% of the total number of cells, the main cell group that can best represent the state of the entire system is selected, and the remaining cells constitute the slave cell group.

[0074] Step S203: Obtain the raw signals of the container energy storage system based on the main cell group and the slave cell group.

[0075] In one optional embodiment, the original signal represents the initial electrical signal directly acquired by the corresponding sensor according to the master and slave cell roles, without filtering, noise reduction, or feature extraction, and may include the master cell original signal and the slave cell original signal.

[0076] In one optional embodiment, based on the determined master-slave cell division result, the process of acquiring the original signals of the full parameters of the master cell and the original signals of the base voltage of the slave cell through differentiated sensor configuration and targeted data acquisition is carried out, thereby realizing on-demand acquisition.

[0077] In one alternative embodiment, differentiated hardware configurations are employed based on the different roles of the master and slave cells. Taking Table 1 below as an example, the master cell employs high-density full-parameter acquisition (voltage / current / temperature); the slave cell only requires basic voltage monitoring (no current / temperature sensors).

[0078] Table 1. Hardware Configuration of Master-Slave Cells

[0079]

[0080] Furthermore, the sensors are synchronously controlled by the battery management system (BMS), and raw signals are acquired according to the hardware configuration parameters shown in Table 1 above.

[0081] Step S204: Based on the original signal, the target main cell feature matrix is ​​established after processing by anti-aliasing filtering and signal preprocessing methods.

[0082] In an optional embodiment, the anti-aliasing filtering method represents a signal purification method for eliminating aliasing caused by insufficient sampling during analog-to-digital conversion. In this embodiment, based on the Nyquist sampling theorem, a Butterworth second-order low-pass filter is used to filter high-frequency noise above the cutoff frequency by setting the cutoff frequency as shown in the following relationship (2) according to the differences between master and slave cells, while retaining the effective signal related to the health status of the cell.

[0083] (2)

[0084] In the formula: This indicates the filter cutoff frequency (Hz).

[0085] Furthermore, the transfer function is shown in the following relation (3):

[0086] (3)

[0087] In the formula: This represents the transfer function, which describes the relationship between the input and output signals during anti-aliasing filtering. Represents a complex frequency variable, used to characterize the frequency characteristics of a signal; This indicates the cutoff angular frequency of the filter, expressed in radians per second.

[0088] In one optional embodiment, the signal preprocessing method represents a means of further optimizing and extracting features from the signal after anti-aliasing filtering. It can be performed in combination with the differentiating characteristics of master and slave cell signals and may include wavelet denoising, moving average filtering, etc.

[0089] In one optional embodiment, based on the acquired master and slave cell raw signals, frequency aliasing distortion is first eliminated by anti-aliasing filtering, and then the raw sensor signals are processed by a signal preprocessing method to transform them into standardized features that are strongly correlated with the cell health status and adapted to neural network input, thereby constructing the corresponding target master cell feature matrix.

[0090] Step S205: Based on the target main cell feature matrix and multi-dimensional weighted adjacency matrix, the health status prediction result of the main cell of the container energy storage system is obtained through spatiotemporal graph neural network prediction.

[0091] In one optional embodiment, Spatio-Temporal Graph Neural Networks (ST-GNN) represent a neural network model specifically designed for processing spatio-temporal graph data. It combines the ideas of Graph Neural Networks (GNNs) and spatio-temporal modeling, enabling it to simultaneously capture both spatial dependencies (relationships between nodes) and temporal dynamics (changes over time) in the data. This embodiment presents a core prediction model that integrates physical mechanism constraints and spatio-temporal feature learning for predicting the state of health (SOH) of battery cells in containerized energy storage systems. Essentially, it models the physical relationships between battery cells using a graph structure and captures aging dynamics using spatio-temporal modules, accurately mapping the physical characteristics of the master cell to the slave cell. This model may include a physical feature embedding layer, a thermally constrained 3D convolutional layer, a current balance graph attention layer, a feature fusion layer, and an output layer.

[0092] In one optional embodiment, by combining the target main cell feature matrix and the multidimensional weighted adjacency matrix, the spatiotemporal graph neural network learns the mapping relationship between the main cell features and the slave cell health status under the constraint of electrochemical mechanism.

[0093] Specifically, the aging mechanism information is first injected through the physical feature embedding layer, then the spatial physical correlation is captured through the thermal constraint and current balance modules, and finally, the health status prediction results of the battery cells are output through the feature fusion and output layer. This not only ensures the prediction accuracy (error ≤ 2.3%), but also has physical interpretability, solving the problems of traditional pure data models being detached from electrochemical mechanisms and having poor prediction stability.

[0094] The health status prediction method for containerized energy storage systems provided in this embodiment encodes the spatial and electrical coupling relationship of the cell clusters in the containerized energy storage system and determines a multi-dimensional weighted adjacency matrix through three weighted terms: thermal conduction, electrical impedance, and current distribution. This overcomes the limitation of traditional all-sensor schemes that treat cells as independent units. Furthermore, by combining the multi-dimensional weighted adjacency matrix and using an improved non-dominated sorting genetic algorithm for multi-objective optimization to determine the master-slave cell group, the spatial bias of traditional sampling monitoring and the cost increase of all-sensor schemes are solved. Furthermore, by acquiring raw signals according to the master-slave cell group, differentiated data acquisition is achieved, reducing unnecessary sensor deployment, lowering hardware costs and data transmission pressure, while ensuring the integrity of core status data. Furthermore, by combining anti-aliasing filtering methods and signal preprocessing methods to establish the target master cell feature matrix, signal noise and dimensional influences are eliminated, enabling the extraction of core features strongly correlated with the cell health status. This provides high-quality input for neural network prediction and helps improve prediction accuracy. Furthermore, by incorporating a spatiotemporal graph neural network constrained by electrochemical mechanisms for prediction, the shortcomings of purely data-driven models that are detached from electrochemical mechanisms are overcome. This enables high-precision predictions on resource-constrained BMS embedded devices and solves the problem of excessive prediction errors under complex operating conditions. Therefore, by implementing this invention, the stability and accuracy of prediction errors in large-scale energy storage systems are improved, costs are reduced, and a balance between accurate prediction and cost reduction is achieved.

[0095] In some optional implementations, step S201 above includes:

[0096] Step S2011: Obtain the absolute position coordinates of the center point of each cell in the containerized energy storage system and establish a distance matrix.

[0097] In one optional embodiment, by acquiring the three-dimensional spatial coordinates of all battery cells inside the container through precise positioning technology and calculating the Euclidean distance between any two battery cells, a distance matrix that quantifies the spatial positional relationship of the battery cells can be constructed.

[0098] In one optional embodiment, a laser positioning system is used to scan the entire container energy storage module, establish a three-dimensional coordinate system (X, Y, Z axes), and record the absolute position coordinates of the center point of each cell.

[0099] Furthermore, calculate any two battery cells and The Euclidean distance between them is calculated, and a distance matrix is ​​formed. The following relation (4) is shown:

[0100] (4)

[0101] In the formula: Indicates the total number of battery cells; Indicates battery cell and The Euclidean distance between them is shown in the following equation (5):

[0102] (5)

[0103] In the formula: and They represent battery cells respectively. and The absolute position coordinates of the center point.

[0104] Step S2012: Obtain multiple distribution ratios of current in multiple branches under initial static conditions in the parallel branch connection topology diagram of the container energy storage system.

[0105] In an optional embodiment, the parallel branch connection topology diagram is used to reflect the circuit connection between cells in the containerized energy storage system.

[0106] In an optional embodiment, the allocation ratio represents the percentage difference between the current in any two parallel branches of the battery cells and the average current in all branches of the system under the initial static operating conditions of the packaged energy storage system. This ratio is used to quantify the current balance characteristics of the electrical topology between the battery cells and to reflect the consistency of the branch current distribution.

[0107] In one optional embodiment, based on the circuit connection topology of the container energy storage system, the current value of each parallel branch is measured under initial static conditions, and the current distribution ratio of any two battery cells in the branch can be calculated, thereby quantifying the current balance characteristics of the electrical topology.

[0108] In one optional embodiment, a parallel branch connection topology map is obtained from the battery management system (BMS), and the parallel branch to which each cell belongs and the connection relationship between the branches are determined.

[0109] Furthermore, the system is placed in an initial static condition, i.e., no charging / discharging load or a fixed light load state, and the real-time current value of each parallel branch is measured using branch current sensors. Then, the average current is calculated based on the current values ​​of all branches, as shown in the following equation (6):

[0110] (6)

[0111] In the formula: Indicates all branches The average current; Indicates the first Real-time current values ​​of each branch.

[0112] Furthermore, for any two battery cells and Based on the current of its respective branch, the current distribution ratio of each branch is calculated as shown in the following formula (7):

[0113] (7)

[0114] In the formula: Indicates the allocation ratio; and They represent battery cells respectively. and The current value of the branch circuit.

[0115] Step S2013: Apply a preset AC excitation signal to the current of each branch in the parallel branch connection topology diagram, and obtain multiple AC impedance phase differences between different adjacent cells.

[0116] In an optional embodiment, the preset AC excitation signal is a specific frequency AC signal applied to measure the electrical coupling characteristics between battery cells, used to excite the AC impedance response of the battery cells.

[0117] In an optional embodiment, the AC impedance phase difference represents the absolute value of the phase difference between the voltage response and the current excitation of adjacent cells after a preset AC excitation signal is applied. It is used to characterize the coupling strength of the electrical characteristics between cells and reflect the correlation of the electrical parameters of the cells.

[0118] In one optional embodiment, a preset AC excitation signal is applied to each parallel branch by a dedicated excitation device, the voltage response of adjacent cells is collected synchronously, and the phase difference between voltage and current is calculated, thereby obtaining the AC impedance phase difference that characterizes the electrical coupling strength between cells.

[0119] In one optional embodiment, a preset AC excitation signal is applied to each parallel branch of the container energy storage system through a dedicated excitation source. The signal type is a 10kHz sinusoidal AC current, and the amplitude is set to 5% C-rate to ensure that the signal is stable and conforms to the electrical characteristics of the battery cell.

[0120] Furthermore, a high-precision phase meter is used to synchronously acquire the voltage response signal of each cell under AC excitation. Then, the voltage response signal of adjacent cells is analyzed. and The phase relationship between the voltage response and the current excitation is analyzed, and the phase difference between the two is calculated and its absolute value is taken, thus finally obtaining the corresponding AC impedance phase difference. .

[0121] Step S2014: Construct a multidimensional weighted adjacency matrix based on the distance matrix, multiple allocation ratios, and multiple AC impedance phase differences.

[0122] In an optional embodiment, the spatial thermal conduction relationship between battery cells (i.e., distance matrix), electrical topology current balancing characteristics (i.e., distribution ratio), and electrical coupling strength (i.e., AC impedance phase difference) are integrated. A multidimensional weighted adjacency matrix is ​​constructed by weighted fusion to quantify the spatial-electrical coupling relationship of the battery cell cluster. The corresponding multidimensional weighted adjacency matrix can then be constructed, as shown in the following equation (8):

[0123] (8)

[0124] Among them, the weighting coefficient , , It is the optimal ratio obtained through orthogonal experiments, used to ensure the balance between heat conduction, electrical impedance and current distribution.

[0125] Furthermore, this multidimensional weighted adjacency matrix possesses the following characteristics, including symmetry ( ), normalization (diagonal elements) ), decay (as As the thermal conductivity term increases, its exponential decrease occurs.

[0126] In some optional implementations, step S202 above includes:

[0127] Step S2021: Based on the multidimensional weighted adjacency matrix and the historical temperature dataset of the container energy storage system, determine the multi-objective optimization function.

[0128] In an alternative embodiment, the multi-objective optimization function is used to characterize the minimization of thermal characterization error and the minimization of topological similarity error, as shown in the following relationship (9):

[0129] (9)

[0130] In the formula: Indicates a subset of main battery cells; express The corresponding main cell adjacency matrix; The root mean square error between the temperature rise curve of the main cell group and the whole set is shown in the following relationship (10); Represents the adjacency matrix of the main battery cells and the universal matrix The Frobenius norm difference is shown in the following relation (11); and This represents the weighting coefficient.

[0131] (10)

[0132] (11)

[0133] In the formula: Indicates thermal characterization error; Indicates topological similarity error; Indicates the main cell subset At any moment The average temperature; Indicates the entire system at time [time]. The average temperature; Indicates the number of time points; and These represent the adjacency matrix of the main battery cells. and the universal matrix matrix elements, and They represent the first Line and number List.

[0134] Furthermore, the constraints are shown in the following relation (12):

[0135] (12)

[0136] In the formula: This indicates the total number of battery cells.

[0137] Step S2022: The improved non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function and determine the master cell group and slave cell group of the container energy storage system.

[0138] In one optional embodiment, an improved non-dominated sorting genetic algorithm is used to perform evolutionary solutions on the multi-objective optimization function. Through processes such as population initialization, evolutionary operation, and constraint verification, a subset of master cells that meet the constraints and have the optimal optimization objective is selected from the Pareto front solution set, and the remaining cells constitute the slave cell group.

[0139] In one alternative embodiment, the selection of the main battery cell is a multi-objective optimization problem. The objective is to maximize the information coverage of a representative subset while satisfying a quantity constraint (not exceeding 5% of the total number of cells). Based on the constructed physical topology, an improved NSGA-II algorithm is used to select the optimal subset of main battery cells. Specifically, this includes:

[0140] (1) Algorithm initialization: Use binary encoding to indicate whether each cell is selected as the master cell. The chromosome length is the total number of cells. (gene locus) Indicates selection. (Indicates not selected); initialize the population size to 100, randomly generate the initial population, and ensure that each individual meets the main cell number constraint, i.e., the number of selected cells. .

[0141] (2) Perform bi-objective fitness calculation: For each individual in the population, i.e. a subset of main cells The evaluation includes two fitness functions: thermal characterization error. and topological similarity error As shown in equations (10) and (11) above.

[0142] (3) Execution of evolutionary operations:

[0143] (a) Selection: Each time, 5 individuals are randomly selected, and the one with the best fitness is selected to enter the next generation;

[0144] (b) Crossover: Simulated binary crossover is used, with a crossover probability of 0.9 and a distribution index of 20;

[0145] (c) Mutation: Each gene position is flipped with a probability of 0.02 (0 becomes 1 or 1 becomes 0). If the number constraint is violated after mutation, it is readjusted.

[0146] (4) Constraint handling:

[0147] (a) Spatial coverage constraint: Ensure that each 3×3×3 cell cube contains at least one main cell, and perform bit flip dynamic repair on individuals that violate the constraint;

[0148] (b) Hard constraint on the number of main cells: .

[0149] (5) Optimal solution output: Stop iteration after the termination condition is met (generation number ≥ 100), extract the Pareto front solution set, and select the solution that satisfies the condition. and The solution is used as the main battery cell assembly The remaining cells form a cell bank; if there is no solution that simultaneously satisfies both conditions, the weighting coefficients are adjusted according to actual needs. and And solve it again.

[0150] In some optional implementations, step S204 above includes:

[0151] Step S2041: Based on preset differential filtering parameters, anti-aliasing filtering is performed on the main cell group signal and the slave cell group signal in the original signal to obtain the initial main cell group signal and the initial slave cell group signal.

[0152] In an optional embodiment, the preset differential filtering parameters are a set of key anti-aliasing filtering parameters specifically set to adapt to the sampling frequency and application differences of the master and slave cell signals. The purpose is to configure different cutoff frequencies for the Butterworth second-order low-pass filter. The specific parameters can be referred to in step S204.

[0153] In one optional embodiment, differentiated filtering parameters are set according to the sampling frequency and application differences of the master and slave cell signals, and anti-aliasing filtering is performed on the two types of original signals respectively to eliminate frequency aliasing distortion and retain the effective signals related to the health status of the cells, namely the initial master cell group signal and the initial slave cell group signal.

[0154] In an optional embodiment, the voltage, current, and temperature signals of the main battery cell group in the original signal are processed according to... The parameters are used to perform anti-aliasing filtering to obtain the initial main cell group signal; the voltage signal of the slave cell group in the original signal is then processed according to... The parameters are used to perform anti-aliasing filtering to obtain the initial signal from the battery cell group.

[0155] Step S2042: Based on the initial main cell group signal and the initial slave cell group signal, the initial main cell feature matrix is ​​established after processing by the signal preprocessing method.

[0156] In one optional embodiment, the initial master and slave cell group signals after anti-aliasing filtering are purified by a differentiated denoising algorithm, and core feature parameters that are strongly correlated with the health status of the cells are extracted. Finally, the corresponding initial master cell feature matrix is ​​constructed, realizing the transformation of the original signal into effective features.

[0157] Specifically, step S2042 includes:

[0158] Step a1: Perform wavelet denoising processing on the initial main cell group signal to obtain the target main cell group signal.

[0159] In one optional embodiment, by employing a wavelet denoising algorithm to perform high-precision denoising on the initial main cell group signal, it is possible to eliminate random noise while preserving key aging characteristics such as ohmic internal resistance increment and temperature rise rate to the maximum extent.

[0160] In one alternative embodiment, the denoising parameters are first selected: the Daubechies wavelet basis is used, and a 5-level decomposition hierarchy is set to determine the threshold processing rules for the high-frequency detail coefficients.

[0161] Then, the initial main cell group signal is decomposed into low-frequency approximation coefficients, i.e. effective features, and high-frequency detail coefficients, i.e. noise. Thresholding is then performed on the high-frequency detail coefficients to filter out noise components.

[0162] Finally, the processed low-frequency approximation coefficients and the threshold-processed high-frequency detail coefficients are reconstructed to obtain the target main cell group signal, as shown in the following relationship (13):

[0163] (13)

[0164] In the formula: Indicates the target main battery cell group signal; Indicates the number of thresholds after threshold processing Layer detail factor; Indicates the first Wavelet basis functions corresponding to the layer; This represents the total number of wavelet decomposition layers.

[0165] Step a2: Perform a moving average filtering on the initial slave cell group signal to obtain the target slave cell group signal.

[0166] In one alternative embodiment, by employing a lightweight moving average filtering algorithm, the random fluctuations of the initial cell group voltage signal are smoothed, which can control the computational complexity while ensuring signal stability, thereby adapting to the needs of cell deviation detection.

[0167] In an alternative embodiment, the sliding window length is first determined. Then, for the initial slave cell group voltage signal, the average voltage within the window is calculated according to the following relationship (14) to obtain the target slave cell group signal:

[0168] (14)

[0169] In the formula: This represents the smoothed voltage value, which is the target signal from the battery cell assembly. Indicates the first The original voltage sample value of the cell at any given time is the original data before filtering (initial cell group signal).

[0170] Step a3: Based on the target main cell group signal and the target slave cell group signal, feature extraction is performed to obtain multiple feature parameters.

[0171] In one optional embodiment, key physical features that characterize the aging state of the cells are extracted from the target main cell group signal, and auxiliary verification information is obtained from the target main cell group signal, thereby forming a set of feature parameters that are strongly correlated with the state of health (SOH), i.e., multiple feature parameters.

[0172] In one optional embodiment, several feature parameters are shown in Table 2 below:

[0173] Table 2. List of Core Feature Extractions

[0174]

[0175] in, Indicates open-circuit voltage; express Operating voltage at any given time; Indicates the load current; express The temperature of a moment.

[0176] Step a4: Establish the initial main cell feature matrix based on multiple feature parameters.

[0177] In one optional embodiment, the extracted core feature parameters of the main battery cell are organized in a structured manner, and an initial main battery cell feature matrix is ​​constructed.

[0178] For example, based on the number of main battery cells ( The number of rows is denoted by ), and the number of columns is denoted by 3 core features. An initial main cell feature matrix of M×3 dimensions is constructed, as shown in the following relation (15):

[0179] (15)

[0180] In the formula: This represents the initial main cell characteristic matrix.

[0181] Step S2043: Normalize the initial main cell feature matrix to obtain the target main cell feature matrix.

[0182] In an optional embodiment, the various feature parameters in the initial main cell feature matrix are subjected to Min-Max normalization, as shown in the following relationship (16):

[0183] (16)

[0184] In the formula: This represents the feature parameter values ​​after normalization. This represents the minimum value of the characteristic parameter; This represents the maximum value of the characteristic parameter.

[0185] Furthermore, normalization can eliminate the dimensional differences between different features, keep the feature scale consistent, accelerate the convergence speed of the subsequent neural network, and avoid prediction bias caused by inconsistent feature scales.

[0186] In some optional implementations, step S205 above includes:

[0187] Step S2051: Input the target main cell feature matrix into the physical feature embedding layer of the spatiotemporal graph neural network for processing to obtain the physical embedding feature matrix.

[0188] In an alternative embodiment, the physical feature embedding layer The input layer of the spatiotemporal graph neural network (ST-GNN) is used to convert the low-dimensional physical features of the main battery cell into high-dimensional feature vectors that can be processed by the neural network. At the same time, it embeds electrochemical mechanism information (such as EIS spectral features) to inject physical constraints into the subsequent spatiotemporal feature learning and avoid the black box defects of pure data-driven learning.

[0189] In an optional embodiment, the standardized target main cell feature matrix is ​​input into the physical feature embedding layer. Through electrochemical feature encoding and linear transformation, the low-dimensional physical features are transformed into high-dimensional embedding vectors. At the same time, electrochemical mechanism constraints are injected to obtain the corresponding physical embedding feature matrix, as shown in the following relationship (17):

[0190] (17)

[0191] In the formula: Represents the physical embedding feature matrix; Represents the electrochemical characteristic encoder function; Represents the physical feature weight matrix; This indicates a deviation in physical characteristics.

[0192] Step S2052: Input the physical embedding feature matrix and the preset heat dissipation system parameter set into the thermal constraint 3D convolutional layer of the spatiotemporal graph neural network for processing to obtain the spatial feature matrix that integrates the thermal distribution correlation of the main battery cell.

[0193] In an alternative embodiment, thermally constrained 3D convolutional layers As the spatial feature extraction layer of ST-GNN, the convolution kernel is designed based on the three-dimensional heat dissipation path of the container energy storage system. By using weight modulation to force the model to learn spatial heat conduction correlations that conform to thermodynamic laws, it can accurately capture the temperature gradient transfer effect between cells and enhance the characterization ability of heat distribution-related features.

[0194] In an optional embodiment, the physical embedded feature matrix is ​​combined with preset heat dissipation system parameters, and the thermal correlation features of the main battery cell in three-dimensional space are aggregated through thermally constrained 3D convolution operation, and the temperature field distribution law is simulated, so as to finally obtain the feature matrix that fuses the spatial thermal distribution correlation, i.e., the spatial feature matrix.

[0195] For example, the convolution kernel weights are modulated according to the heat dissipation path to enhance feature extraction along the airflow direction, as shown in the following relationship (18):

[0196] (18)

[0197] In the formula: This represents the weight matrix of a thermally constrained 3D convolutional layer. Indicates 3D convolution kernel in The weight value of the location; This represents the vector from the center of the convolution kernel to the adjacent cell; This represents the thermal conduction attenuation function.

[0198] Furthermore, convolution operations are performed according to the following relation (19), and the spatial feature matrix is ​​calculated:

[0199] (19)

[0200] In the formula: Represents the spatial characteristic matrix.

[0201] Step S2053: Input the spatial feature moments and multidimensional weighted adjacency matrix into the current balance graph attention layer of the spatiotemporal graph neural network for processing to obtain the electrical feature matrix that integrates spatial thermal features and electrical balance constraints.

[0202] In an alternative embodiment, the current balance map attention layer This represents the electrical feature fusion layer of ST-GNN. By introducing a graph attention mechanism and current deviation penalty constraint, and dynamically adjusting the attention weights, it captures the impact of uneven current distribution in parallel branches on cell aging, suppresses interference in the electrical imbalance region, and improves the learning accuracy of electrical topology association features.

[0203] In an optional embodiment, the spatial feature matrix, the multidimensional weighted adjacency matrix, and the branch current data are input into the attention layer of the current balance graph. Through current deviation penalty and attention mechanism, the electrical topology association features are aggregated to obtain an electrical feature matrix that integrates spatial thermal features and electrical balance constraints.

[0204] In one alternative embodiment, a current deviation penalty mechanism is first constructed, which may include physical constraints for current deviation penalty. and current balance correction factor The relationships are shown in equations (20) and (21) respectively:

[0205] (20)

[0206] (twenty one)

[0207] In the formula: This represents the penalty intensity coefficient.

[0208] Further, graph attention calculation is performed:

[0209] (1) Input feature transformation: ,in It is a battery cell. The feature vector (refers to the initial features of the battery cell, i.e., each row of the main battery cell feature matrix). It is a learnable linear transformation matrix.

[0210] (2) Attention energy function (introducing a current physical constraint term into the standard GAT energy function) The following relation (22) is shown:

[0211] (twenty two)

[0212] In the formula: Indicates battery cell and Attention energy value; A learnable vector representing the attention mechanism; Indicates feature similarity; This indicates a penalty for current deviation.

[0213] (3) The normalized attention coefficient is shown in the following relationship (23):

[0214] (twenty three)

[0215] In the formula: Indicates battery cell A collection of associated battery cells; This represents an exponential function.

[0216] (4) Feature aggregation output, as shown in the following relation (24):

[0217] (twenty four)

[0218] In the formula: Indicates battery cell The output vector after feature aggregation is the feature values ​​of the current balance graph attention layer, which are aggregated after processing by the current balance graph attention layer. Each line; This represents the ELU activation function.

[0219] (5) Multi-attention extension integration, as shown in the following relation (25):

[0220] (25)

[0221] In the formula: Represents the electrical characteristic matrix; This represents the attention layer operation function of the current balance diagram; This represents branch current data, which can include the current of each branch and the average current.

[0222] Step S2054: Input the physical embedding feature matrix, spatial feature matrix and electrical feature matrix into the feature fusion layer of the spatiotemporal graph neural network to obtain the fused feature matrix.

[0223] In an optional embodiment, the feature fusion layer This is used to concatenate physical embedding features, spatial thermal features, and electrical topology features by dimension, extract key cross-dimensional correlation information through nonlinear transformation, and thus achieve deep fusion of multi-dimensional features.

[0224] In an optional embodiment, the physical embedding feature matrix, spatial feature matrix, and electrical feature matrix are concatenated column-wise, and the feature dimensions are compressed through nonlinear transformation to achieve deep integration of multi-dimensional features, thereby finally obtaining a fused feature matrix. The following relation (26) is shown:

[0225] (26)

[0226] in, Indicates feature concatenation operation; This represents the fusion weight matrix.

[0227] Step S2055: Input the fused feature moments into the output layer of the spatiotemporal graph neural network, and process them using the state transition operator to obtain the cell health status prediction result of the container energy storage system.

[0228] In an optional embodiment, the output layer is used to establish a mapping relationship between the fusion features of the main cell and the SOH of the slave cell through a state transition operator, and to output a normalized SOH prediction value in combination with a Sigmoid activation function.

[0229] In an optional embodiment, the state transition operator represents a matrix of size M×K (M is the number of master cells and K is the number of slave cells), which physically represents the influence coefficient of the master cell characteristic changes on the aging state of slave cells, and is used to encode the aging coupling strength between cells.

[0230] In one optional embodiment, the fused feature matrix is ​​input to the output layer, a master-slave cell feature mapping is established through a state transition operator, and the SOH prediction results of all slave cells are output by combining deviation correction and activation functions.

[0231] For example, for each cell Calculate the predicted SOH value The following relation (27) is shown:

[0232] (27)

[0233] in, This represents the Sigmoid activation function; Indicates output deviation; Indicates the main battery cell assembly; Indicates a collection of battery cells; (Dimension M×K) represents the state transition operator, as shown in the following relation (28):

[0234] (28)

[0235] in, Indicates the main battery cell Characteristic changes from battery cell Aging state The influence coefficient describes the aging coupling strength between cells (e.g., the accelerated aging effect of heat transfer caused by uneven heat dissipation).

[0236] Furthermore, by traversing all the slave cells, the health status prediction matrix of the slave cells (size K×1) is obtained, as shown in the following relation (29):

[0237] (29)

[0238] in, .

[0239] In some alternative implementations, it also includes:

[0240] Step b1: Calculate multiple predicted cell voltage values ​​based on the predicted cell health status results.

[0241] Step b2: Determine multiple voltage prediction deviation rates based on multiple predicted values ​​of cell voltage and multiple measured values ​​of cell voltage.

[0242] Step b3: Determine whether to trigger correction based on multiple voltage prediction deviation rates.

[0243] Step b4: When calibration is triggered, measure the key aging parameters of the main cell by AC impedance spectroscopy and determine the impedance fitting residual.

[0244] Step b5: Update the state transition operator based on multiple voltage prediction deviation rates and impedance fitting residuals.

[0245] In one alternative embodiment, based on the predicted SOH from the cell, the SOH is converted to SOC using a simplified battery model that includes internal resistance, and then mapped to obtain the predicted value of the cell voltage.

[0246] For example, a simplified battery model (including internal resistance parameters) is used to convert the SOH value of each cell into a corresponding SOC value. Then, based on the SOC-voltage characteristic curve of the lithium-ion battery, the SOC value is mapped to a predicted voltage value.

[0247] Furthermore, by comparing the predicted and measured values ​​of the cell voltage, the voltage deviation rate of each cell is calculated, and the difference between the predicted value and the actual state is quantified, as shown in the following relationship (30):

[0248] (30)

[0249] In the formula: Indicates the voltage prediction deviation rate; Indicates battery cell The predicted voltage value; Indicates battery cell The measured voltage value.

[0250] Furthermore, based on the voltage prediction deviation rate, and combined with the triggering condition shown in the following relationship (31), it is determined whether the state transition operator update needs to be initiated, thereby realizing dynamic triggering of the correction:

[0251] (31)

[0252] Furthermore, when the triggering condition is met:

[0253] i) Suspend the operation of the main battery cell (≤300ms);

[0254] ii) Apply a sweep frequency stimulus: , Current;

[0255] iii) Collect the voltage response, as shown in the following equation (32):

[0256] (32)

[0257] In the formula: Represents frequency The cell's complex impedance is below; Represents frequency AC voltage response under the following conditions; Represents frequency AC current excitation; Represents the phase factor; This represents the frequency-dependent phase difference function.

[0258] iv) Analyze EIS aging parameters, including ohmic impedance, charge transfer impedance, and Weber diffusion impedance.

[0259] v) Calculate the residual loss function, as shown in the following equation (33):

[0260] (33)

[0261] In the formula: This represents the impedance fitting residual; Represents frequency The theoretical impedance of the model is as follows; Represents frequency The measured complex impedance is as follows; Indicates the number of sweep points; This indicates a frequency sweep.

[0262] Furthermore, based on the voltage prediction deviation rate and impedance fitting residual, incremental updates of the state transition operator are achieved through gradient optimization, while boundary constraints are applied to ensure physical rationality, dynamically correcting the master-slave cell state mapping relationship, specifically including:

[0263] (1) Gradient calculation, as shown in the following relation (34):

[0264] (34)

[0265] In the formula: Representing the residual loss to the state transition operator The gradient; This indicates the operator increment (approaching 0).

[0266] (2) Incremental update, as shown in the following relations (35) and (36):

[0267] (35)

[0268] (36)

[0269] In the formula: Indicates the learning rate; This represents the Sigmoid bias sensitivity weighting function; This represents the updated state transition operator.

[0270] (3) Boundary protection mechanism, as shown in the following equation (37):

[0271] (37)

[0272] In the formula: The state transition operator represents the first... Line number The elements of the column.

[0273] Among them, the lower limit of 0.2 is the minimum aging correlation between cells (even the farthest cells still have a basic correlation), and the upper limit of 1.5 is the maximum aging transfer coefficient (extreme thermal coupling conditions of cells that are close to each other).

[0274] In one example, a method for predicting the health status of a containerized energy storage system based on master-slave cell monitoring and spatiotemporal graph neural networks is provided, including:

[0275] Step 1: System initialization and main battery cell selection, such as Figure 3 As shown.

[0276] 1.1 Physical Topology Modeling:

[0277] This step aims to accurately quantify the physical relationships between cells in a containerized energy storage system, providing a scientific basis for subsequent main cell selection. The specific implementation includes the following key technical aspects:

[0278] 1) Accurate acquisition of spatial coordinates.

[0279] A laser positioning system is used to scan the entire containerized energy storage module, establishing a three-dimensional coordinate system (X, Y, Z axes). The absolute position coordinates of the center point of each cell are recorded, and the coordinates of any two cells are calculated. and The Euclidean distance between them forms a distance matrix. As shown in the above relations (4) and (5).

[0280] 2) Precise analysis of electrical topology.

[0281] The parallel branch connection topology was obtained from the battery management system (BMS), and the current distribution ratio of each branch was measured under the initial static conditions, as shown in the above relationship (6). A 10 kHz AC excitation signal was applied through a dedicated excitation source, and the phase difference of AC impedance between adjacent cells was measured using a high-precision phase meter. .

[0282] 3) Precise configuration of thermodynamic parameters.

[0283] Input key parameters of the cooling system (usually integrated inside the container system), including the airflow direction vector. (Typical Z-axis direction), heat dissipation coefficient (Optimized values ​​for lithium iron phosphate batteries). Acquire historical thermal imaging data and establish a temperature field distribution model.

[0284] 4) Construction of multidimensional adjacency matrix.

[0285] By combining the three types of physical association factors, a weighted adjacency matrix is ​​constructed, as shown in the above relation (8).

[0286] This matrix possesses the following characteristics, including symmetry ( ), normalization (diagonal elements) ), decay (as As the thermal conductivity term increases, its exponential decrease occurs.

[0287] 1.2 Multi-objective optimization solution:

[0288] The selection of main battery cells is a multi-objective optimization problem. Its objective is to maximize the information coverage of a representative subset while satisfying quantity constraints (not exceeding 5% of the total number of cells). Based on the constructed physical topology, an improved NSGA-II algorithm is used to select the optimal subset of main battery cells.

[0289] 1) Evolutionary algorithm initialization:

[0290] Binary encoding is used to indicate whether each cell is selected as the master cell. The chromosome length is the total number of cells. (gene locus) Indicates selection. (Indicates not selected); initialize the population size to 100, randomly generate the initial population, and ensure that each individual meets the main cell number constraint, i.e., the number of selected cells. .

[0291] 2) Bi-objective fitness calculation (for each individual in the population, i.e., a subset of master cells) Evaluation (including two fitness functions):

[0292] The thermal characterization error (calculated as the root mean square error between the temperature curve of the main cell group and the temperature curve of the entire dataset. Temperature data are from historical operating datasets) is shown in the above relation (10).

[0293] Topological similarity (compute subsets) adjacency matrix Adjacency matrix of the entire set The difference), as shown in the above relation (11).

[0294] 3) Execution of evolutionary operations:

[0295] Selection (5 individuals are randomly selected each time, and the one with the best fitness is selected to enter the next generation); Crossover (simulated binary crossover is used, with a crossover probability of 0.9 and a distribution index of 20); Mutation (each gene position is flipped with a probability of 0.02 (0 becomes 1 or 1 becomes 0). If the number constraint is violated after mutation, it is readjusted).

[0296] 4) Constraint handling:

[0297] Spatial coverage constraint (each 3×3×3 cell cube must contain at least one main cell), dynamic repair (bit flip correction for chromosomes that violate the constraint), and hard constraint on the number of main cells. .

[0298] 5) Output of the optimal solution:

[0299] After the termination condition is met (generation number ≥ 100), extract the Pareto front solution set and select the optimal solution based on the following criteria: Output main battery cell assembly And its three-dimensional distribution coordinates. If there is no solution that satisfies both conditions simultaneously, the weighting coefficients are adjusted according to actual needs. and Solve it again.

[0300] Step 2: System initialization and main battery cell selection.

[0301] This step enables hierarchical data acquisition and signal processing between master and slave battery cells, providing high-quality input features for the core prediction model. The specific implementation includes the following key steps:

[0302] 2.1 Sensor hierarchical configuration:

[0303] Different hardware configurations are adopted based on the different roles of the master and slave cells. Taking Table 1 as an example, the master cell uses high-density full parameter acquisition (voltage / current / temperature); the slave cell only needs basic voltage monitoring (no current / temperature sensors).

[0304] 2.2 Signal acquisition and preprocessing process, such as Figure 4 As shown.

[0305] 1) Anti-aliasing filter (general for master-slave cells): Filter type: Butterworth second-order low-pass filter; filter cutoff frequency as shown in the above relationship (2); transfer function as shown in the above relationship (3).

[0306] 2) Differential sampling:

[0307] Main cell: Employs Σ-Δ ADC to achieve 24-bit high-resolution sampling, satisfying the sampling theorem. ;

[0308] From the cell: 12-bit SAR ADC sampling at a rate of 0.1 Hz.

[0309] 3) Dedicated noise reduction processing:

[0310] Main cell: Wavelet denoising algorithm (Daubechies wavelet basis, 5-level decomposition), as shown in the above relation (13);

[0311] From the cell: moving average filter, as shown in the above relationship (14).

[0312] 4) Core feature extraction, as shown in Table 2 above.

[0313] 2.3 Feature Fusion and Standardization (organizing the extracted physical features into a matrix input):

[0314] 1) Main cell characteristic matrix As shown in the above relation (15).

[0315] 2) Standardization: Min-Max normalization (eliminating the influence of feature dimensions and accelerating model convergence), as shown in the above relation (16).

[0316] Step 3: Spatiotemporal graph neural network prediction.

[0317] This step involves constructing a spatiotemporal graph neural network (ST-GNN) based on a physical-data fusion framework to achieve a precise mapping from main cell characteristics to cell health status. The core technological innovation lies in integrating electrochemical mechanism constraints into a deep learning architecture.

[0318] 3.1 Network architecture overview, such as Figure 5 As shown.

[0319] 1) Input data: Main cell feature matrix and adjacency matrix (size M×M).

[0320] 2) Output data: from the cell health status matrix As shown in the above relation (29).

[0321] 3.2 Implementation of key network layers:

[0322] 1) Physical Feature Embedding Layer Its function is to convert electrochemical parameters into features that can be processed by a neural network, as shown in the above relation (17).

[0323] 2) Thermally constrained 3D convolutional layers Its function is to model the spatial thermal conduction effect of the battery, as shown in the above relationships (18) and (19).

[0324] 3) Current balance diagram attention layer This is a key component of the Spatiotemporal Graphical Neural Network (ST-GNN), specifically designed to capture the impact of uneven current distribution in parallel branches within a battery system and embed this physical constraint into the neural network's learning process. In containerized energy storage systems, the current distribution in parallel branches directly affects the spatial distribution characteristics of cell aging, such as... Figure 6 As shown.

[0325] i) Current deviation penalty mechanism, as shown in the above relationships (20) and (21).

[0326] ii) Graph attention calculation: including input feature transformation, attention energy function, normalized attention coefficient, and feature aggregation output, refer to the relevant description in step S2053 above.

[0327] iii) The implementation of multi-attention head extension is shown in the above relation (25).

[0328] 4) Feature Fusion Layer Its function is to integrate physical and spatiotemporal characteristics, as shown in the above relation (26).

[0329] 5) Output layer Its function is to predict the SOH value of the cell, as shown in the above relationship (27).

[0330] Step 4: Dynamic correction of the migration operator.

[0331] This step establishes a closed-loop control system for master-slave cell state transition. By real-time correction of the transition operator, the long-term stability of the prediction model is ensured, thus solving the interruption calibration problem of traditional solutions. Figure 7 and Figure 8 The figures shown are the system architecture diagram and the operation flow sequence diagram, respectively.

[0332] For specific procedures, please refer to the relevant descriptions in steps b1 to b5 above.

[0333] The health status prediction method for containerized energy storage systems based on master-slave cell monitoring and spatiotemporal graph neural networks provided in this example has the following beneficial effects:

[0334] 1. Hardware costs and complexity are significantly reduced.

[0335] This example innovatively employs a physical topology mapping mechanism to construct a master-slave monitoring architecture. By integrating correlation analysis of spatial heat conduction, impedance coupling, and current distribution, it achieves for the first time the coverage of over 96% of system status information using a representative subset of less than 5% of the total number of cells. Since only a few master cells require the deployment of high-precision sensors (acquiring all parameters including current, voltage, and temperature), the sensor requirements of traditional full-monitoring solutions are reduced by over 95%. In a real-world case with a 3200-cell scale, the number of sensors decreased from 9600 to 320, the hardware cost percentage dropped from 35% to 13.6%, and the communication cable length was shortened by over 80%, completely resolving the contradiction between the economics and engineering feasibility of sensor deployment in large-scale energy storage systems.

[0336] 2. Spatial prediction accuracy has been significantly improved.

[0337] This example eliminates spatial bias accumulation through a triple physical constraint mechanism: First, the thermally constrained convolutional layer in the spatiotemporal graph neural network (ST-GNN) strengthens the weights by 40% along the heat dissipation channel direction, accurately simulating the temperature gradient conduction law; second, the current balance graph attention layer actively attenuates the associated weights by 40% when it detects a branch current deviation exceeding 5%, suppressing the impact of electrical imbalance on prediction; finally, the three-dimensional adjacency matrix completely maps the coupling relationship between the cell's spatial location, electrical characteristics, and thermodynamic parameters. This enables the system to stabilize the spatial prediction error below 1.8% after 500 charge-discharge cycles, a 75% reduction compared to the 7.2% error of the traditional sampling scheme, and improves the temperature field fitting accuracy by 74%, fundamentally solving the prediction distortion problem caused by uneven cell aging.

[0338] 3. Breakthrough improvement in the real-time performance of edge computing.

[0339] To address the computational limitations of embedded devices in Battery Management Systems (BMS), this example develops a two-stage lightweight technology: At the model architecture level, electrochemical aging parameters (ohmic impedance increment, SEI film capacitance) are compressed into low-dimensional features through a physical feature embedding layer, avoiding the neural network from repeatedly learning fundamental physical laws and reducing computational load by 89%; at the dynamic correction level, incremental updates using transfer operators replace global model retraining, requiring only adjustments to less than one-thousandth of the parameters to complete prediction calibration. This is the first time that the engineering requirements for online real-time prediction in large-scale energy storage systems have been met.

[0340] 4. Significantly enhanced adaptability to extreme working conditions.

[0341] The core advantage of this example lies in its deep integration of electrochemical mechanisms into the prediction framework: when a voltage deviation exceeding the threshold is detected, the change in ohmic impedance and charge transfer impedance inside the cell is directly measured through the AC impedance spectroscopy (EIS) recalibration module, which improves the prediction accuracy by 77% compared to pure data models; at the same time, by setting a safety boundary for the migration operator parameters, the aging correlation strength between cells is limited to 0.2 to 1.5 times the physical possibility, effectively preventing model collapse caused by local cell failure or voltage acquisition failure, and the prediction error remains below 3.5% under extreme conditions such as sudden current changes and local short circuits.

[0342] 5. Fundamental guarantee of stability during long-term service.

[0343] The migration operator dynamic correction system, built upon the closed-loop feedback control concept, innovatively transforms voltage deviation into a physical calibration signal: when the deviation between the measured voltage of the battery cell and the predicted value exceeds 5%, it automatically triggers impedance spectrum scanning and analyzes changes in the aging mechanism. Through gradient optimization algorithms, it directionally adjusts the migration operator parameters, enabling the prediction model to continuously track the aging trajectory of the battery cell. This system maintains a prediction error below 3% throughout 1000 charge-discharge cycles, improving stability by more than 80% compared to the uncorrected model, and each correction requires only 300 milliseconds without interrupting system operation. This mechanism fundamentally solves the industry-wide problem of inaccurate prediction models throughout the entire lifecycle of energy storage systems.

[0344] In summary, this example demonstrates a breakthrough by employing three innovative methods: physical topology mapping for selecting the main cell cluster, embedding electrochemical constraints into the ST-GNN network, and closed-loop correction using migration operators. This achieves a 95% reduction in sensors while maintaining stable prediction errors within 3% for large-scale energy storage systems. This solution successfully resolves three core contradictions: hardware cost versus prediction accuracy, algorithm complexity versus edge computing, and short-term accuracy versus long-term stability. It provides the industry with a new paradigm for cell health management that combines economic efficiency and reliability.

[0345] This embodiment also provides a health status prediction device for a containerized energy storage system. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0346] This embodiment provides a health status prediction device for a containerized energy storage system, such as... Figure 9 As shown, the device includes:

[0347] Modeling module 301 is used to perform physical topology modeling of the container energy storage system and obtain a multi-dimensional weighted adjacency matrix. The multi-dimensional weighted adjacency matrix is ​​used to encode the spatial and electrical coupling relationship of the cell cluster in the container energy storage system through three types of weight terms: thermal conduction, electrical impedance, and current distribution.

[0348] The optimization module 302 is used to perform multi-objective optimization based on the multi-dimensional weighted adjacency matrix and the historical temperature dataset of the container energy storage system, and to determine the master cell group and slave cell group of the container energy storage system.

[0349] The acquisition module 303 is used to acquire the raw signals of the container energy storage system based on the main cell group and the slave cell group.

[0350] The processing module 304 is used to establish the target main cell feature matrix based on the original signal, after processing by anti-aliasing filtering and signal preprocessing methods.

[0351] The prediction module 305 is used to predict the health status of the battery cells of the container energy storage system based on the feature matrix of the target main battery cell and the multidimensional weighted adjacency matrix through a spatiotemporal graph neural network. The spatiotemporal graph neural network incorporates electrochemical mechanism constraints.

[0352] The containerized energy storage system health status prediction device provided in this embodiment of the invention can execute the containerized energy storage system health status prediction method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method. Further functional descriptions of the above modules are the same as in the corresponding embodiments described above, and will not be repeated here.

[0353] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0354] The following is a detailed reference. Figure 10 This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 401, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 402 or a program loaded from memory 408 into random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the electronic device. The processor 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0355] Typically, the following devices can be connected to I / O interface 405: input devices 406 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 407 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 408 including, for example, magnetic tapes, hard disks, etc.; and communication devices 409. Communication device 409 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 10 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0356] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 409, or installed from a memory 408, or installed from a ROM 402. When the computer program is executed by the processor 401, it performs the functions defined in the containerized energy storage system health status prediction method of the embodiments of the present invention.

[0357] Figure 10 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0358] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the containerized energy storage system health status prediction method shown in the above embodiments is implemented.

[0359] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0360] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for predicting the health status of a containerized energy storage system, characterized in that, The method includes: A physical topology model of the container energy storage system is performed to obtain a multidimensional weighted adjacency matrix. The multidimensional weighted adjacency matrix is ​​used to encode the spatial and electrical coupling relationship of the cell cluster in the container energy storage system through three types of weight terms: thermal conduction, electrical impedance, and current distribution. Based on the multidimensional weighted adjacency matrix and the historical temperature dataset of the container energy storage system, an improved non-dominated sorting genetic algorithm is used for multi-objective optimization to determine the master cell group and slave cell group of the container energy storage system. The raw signals of the container energy storage system are obtained based on the main battery cell group and the slave battery cell group. Based on the original signal, the target main cell feature matrix is ​​established after processing by anti-aliasing filtering and signal preprocessing methods. Based on the target main cell feature matrix and the multidimensional weighted adjacency matrix, the health status prediction result of the main cell of the container energy storage system is obtained through spatiotemporal graph neural network prediction. The spatiotemporal graph neural network incorporates electrochemical mechanism constraints. Specifically, based on the target main cell feature matrix and the multidimensional weighted adjacency matrix, the spatiotemporal graph neural network predicts the cell health status of the containerized energy storage system, including: The target main cell feature matrix is ​​input into the physical feature embedding layer of the spatiotemporal graph neural network for processing to obtain a physical embedding feature matrix. The physical feature embedding layer represents the input layer of the spatiotemporal graph neural network and is used to convert the low-dimensional physical features of the main cell into a high-dimensional feature vector that can be processed by the neural network, while embedding electrochemical mechanism information. The physical embedding feature matrix and the preset heat dissipation system parameter set are input into the thermally constrained 3D convolutional layer of the spatiotemporal graph neural network for processing, to obtain a spatial feature matrix that integrates the thermal distribution correlation of the main battery cell. The thermally constrained 3D convolutional layer represents the spatial feature extraction layer of the spatiotemporal graph neural network. By designing the convolutional kernel and weight modulation based on the three-dimensional heat dissipation path of the container energy storage system, the model is forced to learn spatial heat conduction correlation that conforms to the thermodynamic laws, accurately capturing the temperature gradient transfer effect between battery cells and enhancing the characterization ability of heat distribution-related features. The spatial feature moments and the multidimensional weighted adjacency matrix are input into the current balance graph attention layer of the spatiotemporal graph neural network for processing, to obtain an electrical feature matrix that fuses spatial thermal features and electrical balance constraints. The current balance graph attention layer represents the electrical feature fusion layer of the spatiotemporal graph neural network. By introducing a graph attention mechanism and current deviation penalty constraints, and dynamically adjusting the attention weights, the influence of uneven current distribution in parallel branches on cell aging is captured, interference in the electrical imbalance region is suppressed, and the learning accuracy of electrical topology correlation features is improved. The physical embedding feature matrix, the spatial feature matrix, and the electrical feature matrix are input into the feature fusion layer of the spatiotemporal graph neural network to obtain a fused feature matrix. The feature fusion layer is used to concatenate the physical embedding features, spatial thermal features, and electrical topology features according to dimensions, and extract cross-dimensional key correlation information through nonlinear transformation. The fused feature moments are input into the output layer of the spatiotemporal graph neural network and processed using the state transition operator to obtain the predicted health status of the slave cells in the container energy storage system. The output layer is used to establish the mapping relationship between the fused features of the master cells and the SOH of the slave cells through the state transition operator, and outputs the normalized SOH prediction value in combination with the Sigmoid activation function.

2. The method according to claim 1, characterized in that, Physical topology modeling of the containerized energy storage system yields a multidimensional weighted adjacency matrix, including: Obtain the absolute position coordinates of the center point of each cell in the containerized energy storage system and establish a distance matrix; The current distribution ratios of multiple branches under initial static conditions are obtained in the parallel branch connection topology diagram of the container energy storage system. The parallel branch connection topology diagram is used to reflect the circuit connection between cells in the container energy storage system. A preset AC excitation signal is applied to the current of each branch in the parallel branch connection topology diagram, and multiple AC impedance phase differences between different adjacent cells are obtained. The multidimensional weighted adjacency matrix is ​​constructed based on the distance matrix, the multiple allocation ratios, and the multiple AC impedance phase differences.

3. The method according to claim 1, characterized in that, Based on the multidimensional weighted adjacency matrix and the historical temperature dataset of the containerized energy storage system, an improved non-dominated sorting genetic algorithm is used for multi-objective optimization to determine the master and slave cell groups of the containerized energy storage system, including: Based on the multidimensional weighted adjacency matrix and the historical temperature dataset of the container energy storage system, a multi-objective optimization function is determined. The multi-objective optimization function is used to characterize the minimization of thermal characterization error and the minimization of topological similarity error. The improved non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function and determine the master cell group and the slave cell group of the container energy storage system.

4. The method according to claim 1, characterized in that, Based on the original signal, after processing with anti-aliasing filtering and signal preprocessing methods, a target main cell feature matrix is ​​established, including: Based on preset differential filtering parameters, anti-aliasing filtering is performed on the main cell group signal and the slave cell group signal in the original signal to obtain the initial main cell group signal and the initial slave cell group signal. Based on the initial main cell group signal and the initial slave cell group signal, the initial main cell feature matrix is ​​established after processing by the signal preprocessing method. The initial main cell feature matrix is ​​normalized to obtain the target main cell feature matrix.

5. The method according to claim 4, characterized in that, Based on the initial main cell group signal and the initial slave cell group signal, after processing by the signal preprocessing method, an initial main cell feature matrix is ​​established, including: The initial main cell group signal is subjected to wavelet denoising processing to obtain the target main cell group signal; The initial slave cell group signal is subjected to moving average filtering to obtain the target slave cell group signal; Based on the target main cell group signal and the target secondary cell group signal, feature extraction is performed to obtain multiple feature parameters; The initial main cell feature matrix is ​​established based on the multiple feature parameters.

6. The method according to claim 1, characterized in that, The method further includes: Based on the predicted cell health status, multiple predicted cell voltage values ​​are calculated. Based on the multiple predicted values ​​of cell voltage and the multiple measured values ​​of cell voltage, multiple voltage prediction deviation rates are determined; Whether to trigger correction is determined based on the multiple voltage prediction deviation rates; When calibration is triggered, key aging parameters of the main cell are measured by AC impedance spectroscopy, and the impedance fitting residual is determined. The state transition operator is updated based on the plurality of voltage prediction deviation rates and the impedance fitting residuals.

7. A health status prediction device for a containerized energy storage system, characterized in that, The device includes: The modeling module is used to perform physical topology modeling of the container energy storage system and obtain a multi-dimensional weighted adjacency matrix. The multi-dimensional weighted adjacency matrix is ​​used to encode the spatial and electrical coupling relationship of the cell cluster in the container energy storage system through three types of weight terms: thermal conduction, electrical impedance, and current distribution. An optimization module is used to perform multi-objective optimization based on the multi-dimensional weighted adjacency matrix and the historical temperature dataset of the container energy storage system, using an improved non-dominated sorting genetic algorithm, and to determine the master cell group and slave cell group of the container energy storage system. The acquisition module is used to acquire the raw signals of the container energy storage system based on the main cell group and the slave cell group; The processing module is used to establish the target main cell feature matrix based on the original signal through anti-aliasing filtering and signal preprocessing methods; The prediction module is used to obtain the prediction result of the cell health status of the container energy storage system based on the target main cell feature matrix and the multidimensional weighted adjacency matrix through a spatiotemporal graph neural network. The spatiotemporal graph neural network incorporates electrochemical mechanism constraints. Specifically, the prediction module is used for: The target main cell feature matrix is ​​input into the physical feature embedding layer of the spatiotemporal graph neural network for processing to obtain a physical embedding feature matrix. The physical feature embedding layer represents the input layer of the spatiotemporal graph neural network and is used to convert the low-dimensional physical features of the main cell into a high-dimensional feature vector that can be processed by the neural network, while embedding electrochemical mechanism information. The physical embedding feature matrix and the preset heat dissipation system parameter set are input into the thermally constrained 3D convolutional layer of the spatiotemporal graph neural network for processing, to obtain a spatial feature matrix that integrates the thermal distribution correlation of the main battery cell. The thermally constrained 3D convolutional layer represents the spatial feature extraction layer of the spatiotemporal graph neural network. By designing the convolutional kernel and weight modulation based on the three-dimensional heat dissipation path of the container energy storage system, the model is forced to learn spatial heat conduction correlation that conforms to the thermodynamic laws, accurately capturing the temperature gradient transfer effect between battery cells and enhancing the characterization ability of heat distribution-related features. The spatial feature moments and the multidimensional weighted adjacency matrix are input into the current balance graph attention layer of the spatiotemporal graph neural network for processing, to obtain an electrical feature matrix that fuses spatial thermal features and electrical balance constraints. The current balance graph attention layer represents the electrical feature fusion layer of the spatiotemporal graph neural network. By introducing a graph attention mechanism and current deviation penalty constraints, and dynamically adjusting the attention weights, the influence of uneven current distribution in parallel branches on cell aging is captured, interference in the electrical imbalance region is suppressed, and the learning accuracy of electrical topology correlation features is improved. The physical embedding feature matrix, the spatial feature matrix, and the electrical feature matrix are input into the feature fusion layer of the spatiotemporal graph neural network to obtain a fused feature matrix. The feature fusion layer is used to concatenate the physical embedding features, spatial thermal features, and electrical topology features according to dimensions, and extract cross-dimensional key correlation information through nonlinear transformation. The fused feature moments are input into the output layer of the spatiotemporal graph neural network and processed using the state transition operator to obtain the predicted health status of the slave cells in the container energy storage system. The output layer is used to establish the mapping relationship between the fused features of the master cells and the SOH of the slave cells through the state transition operator, and outputs the normalized SOH prediction value in combination with the Sigmoid activation function.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the health status prediction method for the containerized energy storage system according to any one of claims 1 to 6.

9. A computer program product, characterized in that, It includes computer instructions for causing a computer to execute the health status prediction method for a containerized energy storage system as described in any one of claims 1 to 6.

Images