A method and system for quickly handling thermal runaway lithium batteries in energy storage power stations

Abstract

The present invention belongs to the field of thermal runaway detection, and provides a rapid processing method and system for thermal runaway lithium batteries in energy storage power stations, including: collecting state parameters of each lithium battery in the energy storage power station at a preset frequency; constructing a first network diagram according to the physical distribution; constructing a second network diagram according to the electrical distribution; constructing a first neural network; performing a first convolution calculation according to the collected lithium battery state parameter sequence and the first network diagram to obtain a first convolution result; performing a second convolution calculation according to the collected lithium battery state parameter sequence and the second network diagram to obtain a second convolution result; a feature fusion layer performs weighted fusion of the first convolution result and the second convolution result to obtain hidden features; performing an isolation forest analysis on the hidden features to obtain an abnormal score of each node, and isolating lithium batteries with high abnormal scores. The present invention can accurately and timely perform thermal runaway detection.

Description

Technical Field

[0001] The present invention belongs to the field of thermal runaway detection, and in particular relates to a method and system for quickly processing a thermal runaway lithium battery in an energy storage power station. Background Art

[0002] Energy storage power stations are facilities used to store electrical energy. Their purpose is to balance the power system by storing electrical energy when power supply does not match demand. Energy storage power stations can store excess electrical energy when power demand is low and release it when power demand is high, thereby improving the efficiency and stability of the power system.

[0003] In energy storage power stations, lithium batteries are a common energy storage technology. However, lithium batteries may encounter a serious safety issue during use, called thermal runaway. Thermal runaway means that when the temperature of a lithium battery rises, the internal chemical reaction intensifies, further raising the temperature, forming a vicious cycle, and may eventually cause the battery to catch fire or explode.

[0004] Thermal runaway can be controlled to a certain extent by real-time monitoring of the battery status. However, in energy storage power stations, there are many lithium batteries and they are densely arranged. The battery status affects each other, and the battery working status and parameter changes show a complex nonlinear relationship. For example, when the temperature of one of two adjacent batteries rises, it will affect the temperature rise of the adjacent battery; when the lithium battery is rapidly charged and discharged, the temperature rise is a normal phenomenon; when the lithium battery is not rapidly charged and discharged, its temperature rise is a dangerous signal; and different temperatures for current, voltage, and duration are also different; due to the mutual influence between batteries, the influence between different currents and durations, and the nonlinear relationship between various parameters, it is difficult for the existing technology to have a threshold with high sensitivity and small error to monitor the temperature, voltage, current, etc. of lithium batteries, resulting in the inability to quickly handle thermal runaway lithium batteries. Summary of the invention

[0005] In order to solve the problems in the prior art, the present invention provides a method for quickly processing a thermal runaway lithium battery in an energy storage power station, comprising the following steps: collecting state parameters of each lithium battery in the energy storage power station at a preset frequency, wherein the state parameters include at least temperature, current, and voltage; obtaining the physical distribution of the lithium batteries in the energy storage power station, and constructing a first network diagram according to the physical distribution, wherein each lithium battery in the first network diagram is a node, and when the lithium batteries are adjacent, the corresponding nodes are connected; obtaining the electrical distribution of the lithium batteries in the energy storage power station, and constructing a second network diagram according to the electrical distribution, wherein each lithium battery in the second network diagram is a node, and when the lithium batteries are connected in series, the corresponding nodes are connected; constructing a first neural network, wherein the first neural network is constructed according to the first neural network. The first neural network is a multi-layer graph neural network, which includes at least two layers; the first layer performs a first convolution calculation based on the collected lithium battery state parameter sequence and the first network graph to obtain a first convolution result; the second layer performs a second convolution calculation based on the collected lithium battery state parameter sequence and the second network graph to obtain a second convolution result; the first neural network also includes a feature fusion layer, which performs weighted fusion of the first convolution result and the second convolution result to obtain hidden features; an isolation forest analysis is performed on the hidden features to obtain an abnormality score for each node, and the lithium batteries corresponding to the nodes whose abnormality scores are higher than a preset percentage of the average abnormality score are isolated.

[0006] Furthermore, an adjacency matrix is ​​used to represent the first network graph. In the first network graph, if battery i and battery j are adjacent, the element in the i-th row and j-th column of the matrix element is 1, otherwise it is 0; an adjacency matrix is ​​used to represent the second network graph. In the second network graph, if battery i and battery j are connected in series, the element in the i-th row and j-th column of the matrix element is 1, otherwise it is 0.

[0007] Further, performing a first convolution calculation according to the collected lithium battery state parameter sequence and the first network diagram to obtain a first convolution result includes:

[0008] The convolution calculation formula of the first layer is expressed as:

[0009]

[0010] in

[0011] H1: The output feature matrix of the first layer, with a shape of n×D1, where n is the number of nodes and D1 is the output feature dimension;

[0012] The normalized adjacency matrix of the first network graph, representing the physically distributed adjacency relationship, has a shape of n×n and is calculated as Where A1 is the original adjacency matrix and D1 is the degree matrix;

[0013] X1: The input feature matrix of the first layer, with a shape of t×n×F, where t is the number of time steps, n is the number of nodes, and F is the input feature dimension;

[0014] W1: The weight matrix of the first layer, which is a learnable parameter;

[0015] σ: nonlinear activation function, used to introduce nonlinearity to enhance the expressiveness of the model.

[0016] Furthermore, performing a second convolution calculation according to the collected lithium battery state parameter sequence and the second network graph to obtain a second convolution result includes:

[0017] The convolution calculation formula of the second layer is expressed as:

[0018]

[0019] in

[0020] H2: The output feature matrix of the second layer, with a shape of n×D2, where n is the number of nodes and D2 is the output feature dimension;

[0021] The normalized adjacency matrix of the second network graph, representing the physically distributed adjacency relationship, has a shape of n×n and is calculated as Where A2 is the original adjacency matrix and D2 is the degree matrix;

[0022] X2: The input feature matrix of the second layer, with a shape of t×n×F, where t is the number of time steps, n is the number of nodes, and F is the input feature dimension;

[0023] W2: weight matrix of the second layer, which is a learnable parameter;

[0024] σ: nonlinear activation function, used to introduce nonlinearity to enhance the expressiveness of the model.

[0025] Furthermore, when the first neural network is trained:

[0026] Use reconstruction loss:

[0027]

[0028] in is the hidden feature X of the i-th node (i) is the input feature of the i-th node, and N is the number of network layers;

[0029] The gradient is calculated through the back-propagation algorithm, and the model parameters are updated according to the gradient. The updated parameters include:

[0030] The weight matrix W1 of the first layer;

[0031] The weight matrix W2 of the second layer;

[0032] Weight coefficients α and β of the feature fusion layer.

[0033] Another aspect of the present invention provides a rapid processing system for thermal runaway lithium batteries in an energy storage power station, comprising the following modules:

[0034] A collection module, used to collect the state parameters of each lithium battery in the energy storage power station at a preset frequency, wherein the state parameters include at least: temperature, current, and voltage;

[0035] A mapping module, used to obtain the physical distribution of lithium batteries in the energy storage power station, and construct a first network diagram according to the physical distribution, wherein each lithium battery in the first network diagram is a node, and when the lithium batteries are adjacent, the corresponding nodes are connected; obtain the electrical distribution of lithium batteries in the energy storage power station, and construct a second network diagram according to the electrical distribution, wherein each lithium battery in the second network diagram is a node, and when the lithium batteries are connected in series, the corresponding nodes are connected;

[0036] A computing module is used to construct a first neural network, wherein the first neural network is a multi-layer graph neural network, and the multi-layer graph neural network includes at least two layers; the first layer performs a first convolution calculation based on the collected lithium battery state parameter sequence and the first network graph to obtain a first convolution result; the second layer performs a second convolution calculation based on the collected lithium battery state parameter sequence and the second network graph to obtain a second convolution result; the first neural network also includes a feature fusion layer, and the feature fusion layer performs weighted fusion on the first convolution result and the second convolution result to obtain a hidden feature;

[0037] The early warning module is used to perform isolation forest analysis on the hidden features to obtain an abnormality score for each node, and isolate the lithium batteries corresponding to the nodes whose abnormality scores are higher than a preset percentage of the average abnormality score.

[0038] The present invention can produce the following beneficial effects through the above technical solution:

[0039] Graph neural networks can simultaneously process the physical and electrical distribution information of lithium batteries, extracting the complex features of nodes (batteries) through multi-layer graph convolutional networks. The adjacency matrix is ​​used to integrate the information of each node and its neighboring nodes, which is very effective in capturing the mutual influence between batteries (heat conduction and current conduction). Through training, the weight matrix can be adaptively adjusted to extract the most useful features for subsequent anomaly detection.

[0040] Isolation Forest does not require pre-labeled abnormal data. It determines abnormal points by the length of isolated paths of samples, which is very suitable for unlabeled data environments. It can process large-scale data and is suitable for real-time monitoring of lithium battery status. At the same time, Isolation Forest is very sensitive to abnormal points in the data and can effectively identify potential thermal runaway phenomena. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0043] Below, the invention is preferably described in conjunction with the accompanying drawings and specific implementation methods.

[0044] This embodiment solves the above problem through the following steps:

[0045] In one embodiment, reference Figure 1 The present invention provides a rapid treatment method for thermal runaway lithium batteries in energy storage power stations, comprising the following steps:

[0046] In order to monitor the operating status of the battery in real time, detect and deal with potential problems early, and ensure the safe and efficient operation of the energy storage system, it is necessary to collect the parameters of the lithium batteries in the energy storage power station. The status parameters of each lithium battery in the energy storage power station are collected at a preset frequency, and the status parameters include at least: temperature, current, and voltage.

[0047] Temperature: The temperature sensor can measure the operating temperature of the battery. Too high or too low a temperature will affect the performance and life of the battery. Current: The current sensor is used to measure the charge and discharge current of the battery. Monitoring the current helps to understand the charge and discharge status and load conditions of the battery. Voltage: The voltage sensor is used to measure the terminal voltage of the battery. Voltage is an important indicator for judging the charge and discharge status of the battery.

[0048] In practical applications, embedded sensors and data acquisition systems can be used to transmit data to a central monitoring system for processing and storage via wireless or wired means, thus enabling remote monitoring and management of battery packs.

[0049] Among the collected parameters, the temperature is affected by the distance between the batteries. For this purpose, a distance network is constructed to obtain the physical distribution of the lithium batteries in the energy storage power station. A first network diagram is constructed based on the physical distribution. Each lithium battery in the first network diagram is a node. When the lithium batteries are adjacent, the corresponding nodes are connected.

[0050] Specifically, determine the actual location of each lithium battery in the energy storage power station. This information can be obtained through battery layout diagrams or installation records.

[0051] The location coordinates of each battery are labeled (eg, using a two-dimensional or three-dimensional coordinate system).

[0052] Each lithium battery is represented as a node. Based on the physical distribution information, the connection between adjacent batteries is determined. Adjacent batteries refer to batteries that are close in space. A distance threshold can be set to determine whether the batteries are adjacent. Connect the nodes corresponding to adjacent batteries to form edges in the network graph.

[0053] The adjacency matrix in graph theory is used to represent the network graph. In the adjacency matrix, if battery i and battery j are adjacent, the element in the i-th row and j-th column of the matrix element is 1, otherwise it is 0.

[0054] The first network diagram is constructed by the above method. The first network diagram reflects the neighboring relationship between the batteries and can determine the temperature influence between the batteries.

[0055] For example, it is assumed that the connection relationship of the physical distribution is as follows:

[0056] Battery 0 and Battery 1 are adjacent

[0057] Battery 1 and Battery 2 are adjacent

[0058] Battery 2 and Battery 3 are adjacent

[0059] Battery 3 and Battery 4 are adjacent

[0060] Battery 4 and Battery 5 are adjacent

[0061] The corresponding first network diagram can be expressed as

[0062]

[0063] Among the collected parameters, current has the greatest impact on battery temperature, and batteries with the same current can provide a mutual reference for the increase in temperature, so a network of series-connected batteries is constructed. The electrical distribution of lithium batteries in the energy storage power station is obtained, and a second network diagram is constructed based on the electrical distribution. Each lithium battery in the second network diagram is a node. When lithium batteries are connected in series, the corresponding nodes are connected.

[0064] Specifically, determine how each lithium battery is connected in the electrical system, especially which batteries are connected in series. Represent each lithium battery as a node.

[0065] Based on the electrical distribution information, determine the connections between the series-connected batteries and add corresponding edges in the network graph.

[0066] Similar to the first network, the adjacency matrix in graph theory is used to represent the series graph. For each battery group connected in series, the adjacency matrix is ​​updated so that the matrix elements of adjacent batteries are 1, indicating that there is a connection between them.

[0067] For example, it is assumed that the series relationship of the electrical distribution is as follows:

[0068] Battery 0, 1, 2 connected in series

[0069] Battery 3, 4, 5 connected in series

[0070] Then the second network graph can be expressed as

[0071]

[0072] A first neural network is constructed, wherein the first neural network is a multi-layer graph neural network, and the multi-layer graph neural network includes at least two layers.

[0073] Graph neural network is a neural network model that can process graph structured data. It gradually extracts feature information of nodes and their neighborhoods through multiple graph convolution layers, thereby gradually extracting higher-level features. Common graph neural networks include GCN (Graph Convolutional Network), GAT (Graph Attention Network), etc.

[0074] The graph convolution layer captures the structural information of nodes and their neighborhoods through the adjacency matrix. Through multiple layers of convolution and activation functions, it can capture the nonlinear characteristics of graph structure data.

[0075] In the aforementioned steps, in order to discover the thermal interference relationship caused by spatial proximity, the spatial characteristics of each battery and its adjacent batteries are extracted based on the physical adjacency matrix.

[0076] In the aforementioned steps, in order to analyze the current distribution and voltage variation law, the electrical connection characteristics of each battery and its series-connected batteries are extracted based on the electrical adjacency matrix.

[0077] By representing the physical layer and the electrical layer as an adjacency matrix, the physical and electrical distribution characteristics of the battery can be comprehensively analyzed through graph neural networks for status monitoring and anomaly detection.

[0078] The first layer performs convolution calculation according to the collected lithium battery state parameter sequence to obtain a first convolution result.

[0079] The convolution calculation formula of the first layer can be expressed as:

[0080]

[0081] in

[0082] H1: Output feature matrix of the first layer. The shape is n×D1, where n is the number of nodes (lithium batteries) and D1 is the output feature dimension.

[0083] The normalized adjacency matrix of the first layer. It represents the adjacency relationship of the physical distribution and has a shape of n×n. It is calculated as Where A1 is the original adjacency matrix and D1 is the degree matrix.

[0084] X1: Input feature matrix of the first layer. At time step t, the shape is t×n×F, where t is the number of time steps (how many consecutive data), n is the number of nodes (lithium batteries), and F is the input feature dimension (in this embodiment, at least three dimensions: temperature, current, and voltage);

[0085] W1: The weight matrix of the first layer. It is a learnable parameter.

[0086] σ: nonlinear activation function, preferably ReLU. It is used to introduce nonlinearity to enhance the expressiveness of the model.

[0087] The second layer performs convolution calculation according to the collected lithium battery state parameter sequence to obtain a second convolution result;

[0088] The convolution calculation formula of the second layer can be expressed as:

[0089]

[0090] in

[0091] H2: Output feature matrix of the second layer. The shape is n×D2, where n is the number of nodes (lithium batteries) and D2 is the output feature dimension.

[0092] The normalized adjacency matrix of the second layer. It represents the adjacency relationship of the physical distribution and has a shape of n×n. It is calculated as Where A2 is the original adjacency matrix and D2 is the degree matrix.

[0093] X2: Input feature matrix of the second layer. At time step t, the shape is t×n×F, where t is the number of time steps (how many consecutive data), n is the number of nodes (lithium batteries), and F is the input feature dimension (in this embodiment, at least three dimensions: temperature, current, and voltage);

[0094] W2: The weight matrix of the second layer. It is a learnable parameter.

[0095] Different graph structures (physical distribution and electrical distribution) reflect different relationships between lithium batteries:

[0096] Physical distribution: indicates the spatial proximity of lithium batteries and can reflect information such as temperature and physical interference.

[0097] Electrical distribution: It indicates the series relationship of lithium batteries in the circuit and can reflect electrical characteristics such as current and voltage.

[0098] The graph neural network extracts the features of nodes and their neighborhoods through the graph convolution layer. Different graph structures capture different types of features. Weighted fusion of these features can improve the expressiveness and robustness of the features, thereby improving the performance of the model. By combining the information of the two graph structures, the state characteristics of the lithium battery can be captured more comprehensively. To this end, the first neural network also includes a feature fusion layer, which performs weighted fusion of the first convolution result and the second convolution result to obtain hidden features.

[0099] H hidden =αH1+βH2

[0100] H hidden :The hidden layer features obtained

[0101] α, β: weight coefficients, which are trainable parameters.

[0102] The feature fusion layer adjusts the contribution of different features by linearly combining different convolution results and using weight coefficients α and β. This linear combination can be regarded as a weighted average, which enables the model to balance the importance of different graph structure features.

[0103] The weight coefficients α and β are learnable parameters that can be optimized through back propagation, so that the model can automatically adjust the ratio of feature fusion and find the optimal feature combination during training.

[0104] The information provided by different graph structures may be complementary. Fusing this information can reduce the limitations of a single graph structure information and improve the model's ability to represent complex graph structure data.

[0105] In this embodiment, the multi-layer graph neural network is used to generate hidden features, so it is a self-supervised learning model. During training, reconstruction loss (such as mean square error loss, MSE) is used, that is, the output features of the model should reconstruct the input features as much as possible:

[0106]

[0107] in is the hidden feature X of the i-th node (i) is the input feature of the i-th node.

[0108] The gradient is calculated through the back-propagation algorithm, and the model parameters are updated according to the gradient. The updated parameters include:

[0109] The weight matrix W1 of the first layer

[0110] The weight matrix W2 of the second layer

[0111] Weight coefficients α and β of the feature fusion layer.

[0112] In the aforementioned steps, the battery features are extracted through a multi-layer graph neural network. Since time series data is used in the extraction process, when multiple cycles of data are used, such as collecting data once every minute and processing one hour of data each time, each battery has 60 sets of data, so the dimension of the features will be very high. In addition, there is currently no data set with clear labels for thermal runaway. In many cases, the abnormal state of lithium batteries is difficult to pre-define (such as through thresholds) or label. Therefore, this embodiment also requires an unsupervised method to separate the abnormal data. Therefore, an isolation forest analysis is performed on the hidden features to obtain the abnormal score of each node, and the lithium batteries corresponding to the nodes whose abnormal scores are higher than the preset percentage of the average abnormal score are isolated.

[0113] Using hidden features H hidden Train an Isolation Forest model. Isolation Forest isolates samples by building random trees and calculating the anomaly score for each sample.

[0114] After training the isolation forest model, calculate the anomaly score of each node, calculate the average anomaly score of all nodes, set a preset percentage threshold (e.g. 20%), and determine the threshold to identify nodes with anomaly scores higher than the threshold as abnormal nodes.

[0115] Thermal runaway of lithium batteries refers to overheating or chemical reactions inside the battery that cause the temperature to rise sharply, which may cause fire or explosion. Through the above method, lithium batteries that may be in thermal runaway can be detected and isolated in time to prevent catastrophic consequences.

[0116] On the other hand, the present invention also provides a rapid processing system for thermal runaway lithium batteries in an energy storage power station, comprising:

[0117] A collection module, used to collect the state parameters of each lithium battery in the energy storage power station at a preset frequency, wherein the state parameters include at least: temperature, current, and voltage;

[0118] A mapping module, used to obtain the physical distribution of lithium batteries in the energy storage power station, and construct a first network diagram according to the physical distribution, wherein each lithium battery in the first network diagram is a node, and when the lithium batteries are adjacent, the corresponding nodes are connected; obtain the electrical distribution of lithium batteries in the energy storage power station, and construct a second network diagram according to the electrical distribution, wherein each lithium battery in the second network diagram is a node, and when the lithium batteries are connected in series, the corresponding nodes are connected;

[0119] A computing module is used to construct a first neural network, wherein the first neural network is a multi-layer graph neural network, and the multi-layer graph neural network includes at least two layers; the first layer performs a first convolution calculation based on the collected lithium battery state parameter sequence and the first network graph to obtain a first convolution result; the second layer performs a second convolution calculation based on the collected lithium battery state parameter sequence and the second network graph to obtain a second convolution result; the first neural network also includes a feature fusion layer, and the feature fusion layer performs weighted fusion on the first convolution result and the second convolution result to obtain a hidden feature;

[0120] The early warning module is used to perform isolation forest analysis on the hidden features to obtain an abnormality score for each node, and isolate the lithium batteries corresponding to the nodes whose abnormality scores are higher than a preset percentage of the average abnormality score.

[0121] Furthermore, the specific implementation methods of the above-mentioned rapid processing system for thermal runaway lithium batteries in an energy storage power station are the same as the rapid processing method for thermal runaway lithium batteries in an energy storage power station, and all further technical solutions in the rapid processing method for thermal runaway lithium batteries in an energy storage power station are completely introduced into the rapid processing system for thermal runaway lithium batteries in an energy storage power station.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

[0123] The present invention does not particularly specify the structure of some modules, which shall be subject to the contents recorded in the prior art. The prior art mentioned in the above background technology section and the specific embodiment section of the present invention can be used as part of the present invention to understand the meaning of some technical features or parameters. The scope of protection of the present invention shall be subject to the contents actually recorded in the claims.

Claims

1. A rapid treatment method for thermal runaway lithium batteries in energy storage power stations, characterized in that The method comprises the following steps: Collecting the state parameters of each lithium battery in the energy storage power station at a preset frequency, wherein the state parameters include at least: temperature, current, and voltage; Obtaining the physical distribution of lithium batteries in the energy storage power station, and constructing a first network graph according to the physical distribution, wherein each lithium battery in the first network graph is a node, and when the lithium batteries are adjacent, the corresponding nodes are connected; Obtaining electrical distribution of lithium batteries in the energy storage power station, and constructing a second network diagram according to the electrical distribution, wherein each lithium battery in the second network diagram is a node, and when the lithium batteries are connected in series, the corresponding nodes are connected; Constructing a first neural network, wherein the first neural network is a multi-layer graph neural network, and the multi-layer graph neural network includes at least two layers; The first layer performs a first convolution calculation according to the collected lithium battery state parameter sequence and the first network graph to obtain a first convolution result; The second layer performs a second convolution calculation according to the collected lithium battery state parameter sequence and the second network graph to obtain a second convolution result; The first neural network further includes a feature fusion layer, wherein the feature fusion layer performs weighted fusion on the first convolution result and the second convolution result to obtain a hidden feature; An isolation forest analysis is performed on the hidden features to obtain an abnormality score for each node, and lithium batteries corresponding to nodes whose abnormality scores are higher than a preset percentage of the average abnormality score are isolated.

2. The rapid treatment method for thermal runaway lithium batteries in energy storage power stations according to claim 1, characterized in that: The first network graph is represented by using an adjacency matrix. In the first network graph, if battery i and battery j are adjacent, the element in the i-th row and j-th column of the matrix element is 1, otherwise it is 0; The second network graph is represented by an adjacency matrix. In the second network graph, if battery i and battery j are connected in series, the element in the i-th row and j-th column of the matrix element is 1, otherwise it is 0.

3. The rapid treatment method for thermal runaway lithium batteries in energy storage power stations according to claim 2, characterized in that The performing a first convolution calculation according to the collected lithium battery state parameter sequence and the first network diagram to obtain a first convolution result includes: The convolution calculation formula of the first layer is expressed as: in H1: The output feature matrix of the first layer, with a shape of n×D1, where n is the number of nodes and D1 is the output feature dimension; The normalized adjacency matrix of the first network graph, representing the physically distributed adjacency relationship, has a shape of n×n and is calculated as Where A1 is the original adjacency matrix and D1 is the degree matrix; X1: The input feature matrix of the first layer, with a shape of t×n×F, where t is the number of time steps, n is the number of nodes, and F is the input feature dimension; W1: The weight matrix of the first layer, which is a learnable parameter; σ: nonlinear activation function, used to introduce nonlinearity to enhance the expressiveness of the model.

4. The rapid treatment method for thermal runaway lithium batteries in energy storage power stations according to claim 2, characterized in that The performing a second convolution calculation according to the collected lithium battery state parameter sequence and the second network diagram to obtain a second convolution result includes: The convolution calculation formula of the second layer is expressed as: in H2: The output feature matrix of the second layer, with a shape of n×D2, where n is the number of nodes and D2 is the output feature dimension; The normalized adjacency matrix of the second network graph, representing the physically distributed adjacency relationship, has a shape of n×n and is calculated as Where A2 is the original adjacency matrix and D2 is the degree matrix; X2: The input feature matrix of the second layer, with a shape of t×n×F, where t is the number of time steps, n is the number of nodes, and F is the input feature dimension; W2: weight matrix of the second layer, which is a learnable parameter; σ: nonlinear activation function, used to introduce nonlinearity to enhance the expressiveness of the model.

5. The rapid treatment method for thermal runaway lithium batteries in energy storage power stations according to claim 1, characterized in that When the first neural network is trained: Use reconstruction loss: in is the hidden feature X of the i-th node (i) is the input feature of the i-th node, and N is the number of network layers; The gradient is calculated through the back-propagation algorithm, and the model parameters are updated according to the gradient. The updated parameters include: The weight matrix W1 of the first layer; The weight matrix W2 of the second layer; Weight coefficients α and β of the feature fusion layer.

6. A rapid treatment system for thermal runaway lithium batteries in energy storage power stations, characterized in that The system includes the following modules: A collection module, used to collect the state parameters of each lithium battery in the energy storage power station at a preset frequency, wherein the state parameters include at least: temperature, current, and voltage; A mapping module, used to obtain the physical distribution of lithium batteries in the energy storage power station, and construct a first network diagram according to the physical distribution, wherein each lithium battery in the first network diagram is a node, and when the lithium batteries are adjacent, the corresponding nodes are connected; obtain the electrical distribution of lithium batteries in the energy storage power station, and construct a second network diagram according to the electrical distribution, wherein each lithium battery in the second network diagram is a node, and when the lithium batteries are connected in series, the corresponding nodes are connected; A computing module is used to construct a first neural network, wherein the first neural network is a multi-layer graph neural network, and the multi-layer graph neural network includes at least two layers; the first layer performs a first convolution calculation based on the collected lithium battery state parameter sequence and the first network graph to obtain a first convolution result; the second layer performs a second convolution calculation based on the collected lithium battery state parameter sequence and the second network graph to obtain a second convolution result; the first neural network also includes a feature fusion layer, and the feature fusion layer performs weighted fusion on the first convolution result and the second convolution result to obtain a hidden feature; The early warning module is used to perform an isolation forest analysis on the hidden features to obtain an abnormality score for each node, and to isolate the lithium batteries corresponding to the nodes whose abnormality scores are higher than a preset percentage of the average abnormality score.

7. The rapid treatment system for thermal runaway lithium batteries in energy storage power stations according to claim 6, characterized in that: The first network graph is represented by using an adjacency matrix. In the first network graph, if battery i and battery j are adjacent, the element in the i-th row and j-th column of the matrix element is 1, otherwise it is 0; The second network graph is represented by an adjacency matrix. In the second network graph, if battery i and battery j are connected in series, the element in the i-th row and j-th column of the matrix element is 1, otherwise it is 0.

8. The rapid treatment system for thermal runaway lithium batteries in energy storage power stations according to claim 7, characterized in that The performing a first convolution calculation according to the collected lithium battery state parameter sequence and the first network diagram to obtain a first convolution result includes: The convolution calculation formula of the first layer is expressed as: in H1: The output feature matrix of the first layer, with a shape of n×D1, where n is the number of nodes and D1 is the output feature dimension; The normalized adjacency matrix of the first network graph, representing the physically distributed adjacency relationship, has a shape of n×n and is calculated as Where A1 is the original adjacency matrix and D1 is the degree matrix; X1: The input feature matrix of the first layer, with a shape of t×n×F, where t is the number of time steps, n is the number of nodes, and F is the input feature dimension; W1: The weight matrix of the first layer, which is a learnable parameter; σ: nonlinear activation function, used to introduce nonlinearity to enhance the expressiveness of the model.

9. The rapid treatment system for thermal runaway lithium batteries in energy storage power stations according to claim 7, characterized in that The performing a second convolution calculation according to the collected lithium battery state parameter sequence and the second network diagram to obtain a second convolution result includes: The convolution calculation formula of the second layer is expressed as: in H2: The output feature matrix of the second layer, with a shape of n×D2, where n is the number of nodes and D2 is the output feature dimension; The normalized adjacency matrix of the second network graph, representing the physically distributed adjacency relationship, has a shape of n×n and is calculated as Where A2 is the original adjacency matrix and D2 is the degree matrix; X2: The input feature matrix of the second layer, with a shape of t×n×F, where t is the number of time steps, n is the number of nodes, and F is the input feature dimension; W2: weight matrix of the second layer, which is a learnable parameter; σ: nonlinear activation function, used to introduce nonlinearity to enhance the expressiveness of the model.

10. The rapid treatment system for thermal runaway lithium batteries in energy storage power stations according to claim 6, characterized in that When the first neural network is trained: Use reconstruction loss: in is the hidden feature X of the i-th node (i) is the input feature of the i-th node, and N is the number of network layers; The gradient is calculated through the back-propagation algorithm, and the model parameters are updated according to the gradient. The updated parameters include: The weight matrix W1 of the first layer; The weight matrix W2 of the second layer; Weight coefficients α and β of the feature fusion layer.

Images