CN-121543462-B - Container energy storage system health state prediction method, device and program product

CN121543462BCN 121543462 BCN121543462 BCN 121543462BCN-121543462-B

Abstract

The invention relates to the technical field of energy storage systems and discloses a method, a device and a program product for predicting the health state of a container energy storage system, wherein the method comprises the steps of carrying out physical topology modeling on the container energy storage system to obtain a multidimensional weight adjacency matrix; the method comprises the steps of carrying out multi-objective optimization by utilizing an improved non-dominant sorting genetic algorithm based on a multi-dimensional weight adjacency matrix and a historical temperature dataset of a container energy storage system, determining a master cell group and a slave cell group of the container energy storage system, acquiring original signals of the container energy storage system according to the master cell group and the slave cell group, establishing a target master cell feature matrix by processing the original signals through an anti-aliasing filtering method and a signal preprocessing method, and obtaining a slave cell health state prediction result of the container energy storage system by means of time-space diagram neural network prediction based on the target master cell feature matrix and the multi-dimensional weight adjacency matrix.

Inventors

LI YUXUAN
DING KE
ZHANG HAIZHEN
XIE YURONG
DENG RUIFENG
MOU MIN
YANG HAOJIE
LIU LILI
WANG YUANCHEN
Jiao Junhao

Assignees

华电电力科学研究院有限公司

Dates

Publication Date: 20260508
Application Date: 20260120

Claims (9)

1. A method for predicting the health of a container energy storage system, the method comprising: Performing physical topology modeling on a container energy storage system to obtain a multi-dimensional weight adjacency matrix, wherein the multi-dimensional weight adjacency matrix is used for encoding the space and electric coupling relation of a battery cell cluster in the container energy storage system through three weight items of heat conduction, electric impedance and current distribution; Performing multi-objective optimization by utilizing an improved non-dominant ranking genetic algorithm based on the multi-dimensional weight adjacency matrix and the historical temperature dataset of the container energy storage system, and determining a master cell group and a slave cell group of the container energy storage system; acquiring an original signal of the container energy storage system according to the master cell group and the slave cell group; based on the original signal, processing by an anti-aliasing filtering method and a signal preprocessing method, and establishing a target main cell characteristic matrix; based on the target main cell characteristic matrix and the multidimensional weight adjacent matrix, obtaining a prediction result of the health state of the secondary cell of the container energy storage system through space-time diagram neural network prediction, wherein the space-time diagram neural network is fused with electrochemical mechanism constraint; based on the target main cell feature matrix and the multidimensional weight adjacent matrix, the prediction result of the health state of the secondary cell of the container energy storage system is obtained through space-time diagram neural network prediction, and the method comprises the following steps: Inputting the target main cell feature matrix into a physical feature embedding layer of the space-time diagram neural network for processing to obtain a physical embedding feature matrix, wherein the physical feature embedding layer represents an input layer of the space-time diagram neural network and is used for converting low-dimensional physical features of the main cell into high-dimensional feature vectors which can be processed by the neural network and embedding electrochemical mechanism information; Inputting the physical embedded feature matrix and a preset heat dissipation system parameter set into a thermal constraint 3D convolution layer of the space-time diagram neural network for processing to obtain a spatial feature matrix fused with the thermal distribution association of the main battery cells, wherein the thermal constraint 3D convolution layer represents a spatial feature extraction layer of the space-time diagram neural network, and the thermal gradient transfer effect among the battery cells is accurately captured by learning spatial heat conduction association conforming to thermodynamic rules based on a three-dimensional heat dissipation path design convolution kernel and a weight modulation forced model of a container energy storage system so as to strengthen the characterization capability of relevant features of the thermal distribution; Inputting the space feature moment and the multidimensional weight adjacent matrix into a current balance diagram attention layer of the space-time diagram neural network for processing to obtain an electrical feature matrix fusing space thermal features and electrical balance constraints, wherein the current balance diagram attention layer represents an electrical feature fusion layer of the space-time diagram neural network, and capturing the influence of current distribution non-uniformity of parallel branches on cell aging, inhibiting interference of an electrical unbalanced region and improving learning accuracy of electrical topology association features by introducing a diagram attention mechanism and current deviation punishment constraint and dynamically adjusting attention weight; Inputting the physical embedded feature matrix, the spatial feature matrix and the electrical feature matrix into a feature fusion layer of the space-time diagram neural network to be processed to obtain a fusion feature matrix, wherein the feature fusion layer is used for splicing physical embedded features, spatial thermal features and electrical topological features according to dimensions and extracting cross-dimension key associated information through nonlinear transformation; and inputting the fusion characteristic moment into an output layer of the space-time diagram neural network, and processing by using a state migration operator to obtain the health state prediction result of the secondary battery cell of the container energy storage system, wherein the output layer is used for establishing a mapping relation between a main battery cell fusion characteristic and the SOH of the secondary battery cell through the state migration operator, and outputting a normalized SOH prediction value by combining a Sigmoid activation function.
2. The method of claim 1, wherein modeling the physical topology of the container energy storage system to obtain the multi-dimensional weight adjacency matrix comprises: Acquiring absolute position coordinates of a central point of each electric core in the container energy storage system, and establishing a distance matrix; Obtaining a plurality of distribution ratios of a plurality of branch currents under an initial static working condition in a parallel branch connection topological graph of the container energy storage system, wherein the parallel branch connection topological graph is used for reflecting the circuit connection condition between electric cores in the container energy storage system; Applying a preset alternating current excitation signal to each branch current in the parallel branch connection topological graph, and acquiring a plurality of alternating current impedance phase differences between different adjacent battery cells; And constructing the multi-dimensional weight adjacency matrix according to the distance matrix, the distribution ratios and the alternating current impedance phase differences.
3. The method of claim 1, wherein performing multi-objective optimization using a modified non-dominant ranking genetic algorithm based on the multi-dimensional weight adjacency matrix and the historical temperature dataset of the container energy storage system and determining a master cell set and a slave cell set of the container energy storage system comprises: determining a multi-objective optimization function based on the multi-dimensional weight adjacency matrix and the historical temperature dataset of the container energy storage system, the multi-objective optimization function being used to characterize minimization of thermal characterization errors and minimization of topological similarity errors; And solving the multi-objective optimization function by utilizing the improved non-dominant ranking genetic algorithm, and determining the master cell group and the slave cell group of the container energy storage system.
4. The method of claim 1, wherein establishing a target primary cell feature matrix based on the original signal via anti-aliasing filtering method and signal preprocessing method processing comprises: Based on preset differential filtering parameters, anti-aliasing filtering is respectively carried out on a master cell group signal and a slave cell group signal in the original signal to obtain an initial master cell group signal and an initial slave cell group signal; Based on the initial master cell group signal and the initial slave cell group signal, processing the initial master cell group signal by the signal preprocessing method, and establishing an initial master cell characteristic matrix; and carrying out normalization processing on the initial main cell characteristic matrix to obtain the target main cell characteristic matrix.
5. The method of claim 4, wherein establishing an initial master cell feature matrix based on the initial master cell group signal and the initial slave cell group signal via the signal preprocessing method comprises: Performing wavelet denoising processing on the initial main cell group signal to obtain a target main cell group signal; Performing moving average filtering processing on the initial slave cell group signal to obtain a target slave cell group signal; Performing feature extraction based on the target main cell group signal and the target slave cell group signal to obtain a plurality of feature parameters; and establishing the initial main cell characteristic matrix according to the characteristic parameters.
6. The method according to claim 1, wherein the method further comprises: Calculating a plurality of secondary battery cell voltage predicted values according to the secondary battery cell health state predicted result; Determining a plurality of voltage prediction deviation rates according to the plurality of slave cell voltage predicted values and the plurality of slave cell voltage actual measurement values; judging whether to trigger correction according to the voltage prediction deviation rates; When triggering correction, measuring main cell aging key parameters through alternating current impedance spectrum, and determining impedance fitting residual errors; And updating the state migration operator according to the plurality of voltage prediction deviation rates and the impedance fitting residual errors.
7. A container energy storage system health state prediction device, the device comprising: The modeling module is used for performing physical topology modeling on the container energy storage system to obtain a multi-dimensional weight adjacency matrix, wherein the multi-dimensional weight adjacency matrix is used for encoding the space and electric coupling relation of the battery cell clusters in the container energy storage system through three weight items of heat conduction, electric impedance and current distribution; the optimizing module is used for carrying out multi-objective optimization by utilizing an improved non-dominant sorting genetic algorithm based on the multi-dimensional weight adjacency matrix and the historical temperature dataset of the container energy storage system, and determining a master cell group and a slave cell group of the container energy storage system; the acquisition module is used for acquiring an original signal of the container energy storage system according to the master cell group and the slave cell group; The processing module is used for processing the original signals through an anti-aliasing filtering method and a signal preprocessing method to establish a target main cell characteristic matrix; the prediction module is used for obtaining a prediction result of the health state of the secondary battery core of the container energy storage system through space-time diagram neural network prediction based on the target main battery core feature matrix and the multidimensional weight adjacent matrix, and the space-time diagram neural network is fused with electrochemical mechanism constraint; the prediction module is specifically configured to: Inputting the target main cell feature matrix into a physical feature embedding layer of the space-time diagram neural network for processing to obtain a physical embedding feature matrix, wherein the physical feature embedding layer represents an input layer of the space-time diagram neural network and is used for converting low-dimensional physical features of the main cell into high-dimensional feature vectors which can be processed by the neural network and embedding electrochemical mechanism information; Inputting the physical embedded feature matrix and a preset heat dissipation system parameter set into a thermal constraint 3D convolution layer of the space-time diagram neural network for processing to obtain a spatial feature matrix fused with the thermal distribution association of the main battery cells, wherein the thermal constraint 3D convolution layer represents a spatial feature extraction layer of the space-time diagram neural network, and the thermal gradient transfer effect among the battery cells is accurately captured by learning spatial heat conduction association conforming to thermodynamic rules based on a three-dimensional heat dissipation path design convolution kernel and a weight modulation forced model of a container energy storage system so as to strengthen the characterization capability of relevant features of the thermal distribution; Inputting the space feature moment and the multidimensional weight adjacent matrix into a current balance diagram attention layer of the space-time diagram neural network for processing to obtain an electrical feature matrix fusing space thermal features and electrical balance constraints, wherein the current balance diagram attention layer represents an electrical feature fusion layer of the space-time diagram neural network, and capturing the influence of current distribution non-uniformity of parallel branches on cell aging, inhibiting interference of an electrical unbalanced region and improving learning accuracy of electrical topology association features by introducing a diagram attention mechanism and current deviation punishment constraint and dynamically adjusting attention weight; Inputting the physical embedded feature matrix, the spatial feature matrix and the electrical feature matrix into a feature fusion layer of the space-time diagram neural network to be processed to obtain a fusion feature matrix, wherein the feature fusion layer is used for splicing physical embedded features, spatial thermal features and electrical topological features according to dimensions and extracting cross-dimension key associated information through nonlinear transformation; and inputting the fusion characteristic moment into an output layer of the space-time diagram neural network, and processing by using a state migration operator to obtain the health state prediction result of the secondary battery cell of the container energy storage system, wherein the output layer is used for establishing a mapping relation between a main battery cell fusion characteristic and the SOH of the secondary battery cell through the state migration operator, and outputting a normalized SOH prediction value by combining a Sigmoid activation function.
8. A computer-readable storage medium having stored thereon computer instructions for causing a computer to perform the method of predicting the health of a container energy storage system according to any one of claims 1 to 6.
9. A computer program product comprising computer instructions for causing a computer to perform the method of predicting the health of a container energy storage system as claimed in any one of claims 1 to 6.

Description

Container energy storage system health state prediction method, device and program product Technical Field The invention relates to the technical field of energy storage systems, in particular to a method, a device and a program product for predicting the health state of a container energy storage system. Background In the field of battery cell level health status monitoring, a full sensor monitoring scheme is most widely adopted, and the method is used for distributing current, voltage and temperature sensors for each battery cell so as to realize SOH prediction. However, this approach results in an exponential increase in hardware cost, namely more than 9600 sensors are required to be installed in a 3200 cell system, accounting for more than 35% of the total cost. More seriously, this approach creates ten thousand data channels, resulting in a battery management system communication delay of over 200 milliseconds, far beyond real-time control requirements. The root cause of these drawbacks is that this approach fails to take into account the physical relevance of the cell clusters-in fact, in container energy storage systems, there is significant spatial thermally conductive coupling between the cells (temperature gradient less than 8 ℃ per m) and current distribution uniformity of the electrical topology (parallel branch current deviation less than 3%), but these relevance characteristics are completely ignored. Another common solution is a sampling monitoring scheme by randomly monitoring about 10% of the cell sensors. However, this approach presents serious spatial bias problems, resulting in a prediction error exceeding 7% due to sampling of areas not covered by thermal runaway (e.g., corner cells aged 15-20% faster than center cells). Meanwhile, the method can ignore the current interaction among parallel branches, so that the SOC estimation offset is larger than 8%, and errors can accumulate with time, namely, after 200 charge and discharge cycles, the prediction error is expanded to be 3.2 times of the initial value. The method is based on the adoption of a static independent sampling model, and the problems of propagation effect of cell aging in space-time dimension and state migration mechanism of sensor data to non-monitoring points cannot be solved. The third category of technology is a pure data driven scheme that uses deep learning networks such as LSTM to implement SOH predictions, but relies on full-scale cell data input. This can lead to data dimension disasters-when the number of cores exceeds 500, model parameters can explode to 10 billion orders of magnitude, making deployment difficult in engineering practice. More specifically, the method has extremely high requirements on hardware, can run only by a GPU server, exceeds the bearing capacity of embedded equipment of a battery management system, has a prediction error of 12% under low-temperature and other complex working conditions, is completely separated from the constraint of electrochemical aging mechanisms (such as overpotential rise, SEI film thickening and other physical processes), and has weak migration capacity, so that a model needs to be retrained every time a new batch of batteries are replaced. These problems are essentially due to the simple view of the predictive problem as a purely mathematical map, neglecting the physical interpretability of the electrochemical aging mechanism and the state-related properties under the three-dimensional spatial constraints of the energy storage system. In summary, the core contradiction faced by the prior art is that the cost is greatly increased due to the fact that mass sensors are required to be deployed to realize accurate prediction, and the prediction precision is obviously sacrificed due to the fact that the cost is reduced. This contradiction has become a key technical bottleneck that restricts the commercialization of large-scale container energy storage systems. Disclosure of Invention The invention provides a method, a device and a program product for predicting the health state of a container energy storage system, which are used for solving the problems of accurate prediction and unbalanced cost reduction existing in the existing battery cell level health state monitoring technology. In a first aspect, the present invention provides a method for predicting a health status of a container energy storage system, the method comprising: The method comprises the steps of carrying out physical topology modeling on a container energy storage system to obtain a multi-dimensional weight adjacency matrix, carrying out heat conduction, electric impedance and current distribution on the multi-dimensional weight adjacency matrix to encode the space and electric coupling relation of a battery core cluster in the container energy storage system, carrying out multi-objective optimization by utilizing an improved non-dominant sorting genetic algorithm based on a historical temperature data