Thermal runaway risk prediction method and apparatus, device, and storage medium

Abstract

The present application relates to the technical field of batteries, and discloses a thermal runaway risk prediction method and apparatus, a device, and a storage medium. The method comprises: processing, by at least two neural network layers in a target prediction model, thermal runaway risk parameters layer by layer, wherein the target prediction model is obtained by pre-training on the basis of state vectors of a plurality of time steps, so that the memory capability of the model for past state sequences can be enhanced, and thus the model can better learn the dynamic characteristics of an energy storage battery system, and captures a complex temporal association relationship among multiple variables, thereby improving the accuracy of thermal runaway risk prediction and reducing the safety risk of the energy storage battery system.

Description

Thermal runaway risk prediction method, device, equipment and storage medium

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on February 5, 2024, with application number 202410167552.5 and invention name “Thermal runaway risk prediction method, device, equipment and storage medium”, the entire contents of which are incorporated by reference into the application. Technical Field

[0002] The present application relates to the field of battery technology, and in particular to a method, device, equipment and storage medium for predicting thermal runaway risk. Background Art

[0003] At present, high-voltage energy storage battery systems are widely used in energy storage, grid peak-valley balance and other fields. However, during long-term operation, high-voltage energy storage battery systems may face safety hazards such as overcharging, over-discharging, and overheating. If not handled in a timely manner, serious accidents such as fire and explosion may occur.

[0004] However, the relevant energy storage battery thermal runaway risk prediction methods are difficult to capture the complex correlation between multiple variables in the energy storage battery system, resulting in low prediction accuracy and high safety risks.

[0005] Application Contents

[0006] In view of the above problems, the present application provides a thermal runaway risk prediction method, device, equipment and storage medium, aiming to solve the problem that the relevant energy storage battery thermal runaway risk prediction method is difficult to capture the complex correlation between multiple variables in the energy storage battery system, resulting in low prediction accuracy and high safety risks.

[0007] In a first aspect, the present application provides a method for predicting thermal runaway risk, the method comprising:

[0008] Obtain thermal runaway risk parameters for each battery pack in the energy storage battery system;

[0009] Processing the thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model to obtain a thermal runaway probability for each battery pack, wherein the target prediction model is pre-trained based on state vectors of multiple time steps, and the state vectors are used to enhance the model's memory of past state sequences of the thermal runaway risk parameters;

[0010] The total probability of thermal runaway risk of the energy storage battery system is predicted based on the thermal runaway probability of each battery pack.

[0011] In the technical solution of the embodiment of the present application, thermal runaway risk parameters are processed layer by layer through at least two neural network layers in the target prediction model, and the target prediction model is pre-trained based on state vectors of multiple time steps in advance, thereby enhancing the model's memory ability of past state sequences, thereby enabling the model to better learn the dynamic characteristics of the energy storage battery system, capture the complex temporal correlation between multiple variables, improve the accuracy of thermal runaway risk prediction, and reduce the safety risk of the energy storage battery system.

[0012] In some embodiments, before obtaining the thermal runaway risk parameters of each battery pack of the energy storage battery system, the method further includes:

[0013] The thermal runaway risk training data of the energy storage battery system is divided into a state sequence of multiple time steps, and the state sequence of each time step is converted into a vector form to obtain the state vector of each time step;

[0014] Assigning representations of child nodes of the tree according to the state vectors of each time step, and recursively combining the representations of the child nodes into a representation of the parent node starting from the root node of the tree until a representation of the entire tree is obtained;

[0015] The representation of the root node of the tree is input into an initial prediction model to train at least two neural network layers in the initial prediction model to obtain a target prediction model.

[0016] In the technical solution of the embodiment of the present application, the variables of the battery system are formed into a state vector, and the model input is adjusted through a recursive network to train the model, thereby further enhancing the model's ability to remember past state sequences, thereby enabling the model to better learn the dynamic characteristics of the energy storage battery system and capture the complex temporal correlation between multiple variables.

[0017] In some embodiments, inputting the representation of the root node of the tree into an initial prediction model to train at least two neural network layers in the initial prediction model to obtain a target prediction model includes:

[0018] Inputting the representation of the root node of the tree into an initial prediction model as a model to perform forward pre-training on at least two neural network layers in the initial prediction model;

[0019] After the neural network layer completes the forward pre-training, the neural network layer is backward fine-tuned through the back propagation network in the initial prediction model to obtain the target prediction model.

[0020] In the technical solution of the embodiment of the present application, by performing forward pre-training and backward fine-tuning on at least two neural network layers in the initial prediction model, the model training effect can be improved and the prediction ability of the target prediction model can be enhanced.

[0021] In some embodiments, inputting the representation of the root node of the tree as a model into an initial prediction model to perform forward pre-training on at least two neural network layers in the initial prediction model includes:

[0022] Inputting the representation of the root node of the tree as input data into a neural network layer in an initial prediction model, and reconstructing the input data back into original data through the neural network layer to obtain an output result;

[0023] Calculating a reconstruction error loss function of the neural network layer according to the input data and the output result, and performing forward pre-training on the neural network layer according to the reconstruction error loss function;

[0024] After the neural network layer completes forward pre-training, the hidden unit state of the neural network layer is input into the neural network layer in the initial prediction model as input data of the next neural network layer, and the step of reconstructing the input data back to the original data through the neural network layer is returned until all neural network layers in the initial prediction model complete forward pre-training.

[0025] In the technical solution of the embodiment of the present application, the forward pre-training is improved by introducing the reconstruction error loss function in the model training process, thereby solving the problem that the model may find it difficult to capture the characteristics of abnormal data due to the small amount of abnormal data, and thus improving the model prediction ability.

[0026] In some embodiments, after the neural network layer completes forward pre-training, fine-tuning the neural network layer backward through the back propagation network in the initial prediction model to obtain a target prediction model includes:

[0027] After the neural network layer completes forward pre-training, labeling normal data and abnormal data in the thermal runaway risk training data;

[0028] Respectively calculating a reconstruction error of the normal data and a reconstruction error of the abnormal data, and calculating a loss function according to the reconstruction error of the normal data and the reconstruction error of the abnormal data;

[0029] The target prediction model is obtained by fine-tuning the neural network layer backward through the back propagation network in the initial prediction model based on the loss function.

[0030] In the technical solution of the embodiment of the present application, the forward pre-training is improved by introducing the reconstruction error loss function in the model training process, and the data is annotated for anomalies in the backward fine-tuning of the model, thereby solving the problem that the model may find it difficult to capture the characteristics of abnormal data due to the small amount of abnormal data, and further improving the model prediction ability.

[0031] In some embodiments, respectively calculating the reconstruction error of the normal data and the reconstruction error of the abnormal data, and calculating the loss function according to the reconstruction error of the normal data and the reconstruction error of the abnormal data, comprises:

[0032] Respectively calculating the reconstruction error of the normal data and the reconstruction error of the abnormal data, and obtaining the normal data weight and the abnormal data weight, wherein the abnormal data weight is greater than a preset threshold;

[0033] A loss function is calculated based on the normal data weight, the reconstruction error of the normal data, the abnormal data weight, and the reconstruction error of the abnormal data.

[0034] In the technical solution of the embodiment of the present application, by anomaly marking the data and assigning higher weights to abnormal data samples through the weighted form of the loss function, the abnormal data characteristics can be better captured, thereby improving the model prediction ability.

[0035] In some embodiments, the neural network layer includes a continuous restricted Boltzmann machine, and reconstructing the input data back to original data through the neural network layer to obtain an output result includes:

[0036] The input data is reconstructed back to the original data through the Gaussian distribution activation function of the continuous restricted Boltzmann machine to obtain an output result, wherein the Gaussian distribution activation function includes a sigmoid function after retaining the sigmoid transfer function, removing the data discretization process, and adding a Gaussian random noise variable.

[0037] In the technical solution of the embodiment of the present application, the input data is reconstructed back to the original data through the Gaussian distribution activation function of the continuous restricted Boltzmann machine to obtain the output result, thereby ensuring that there are continuous state values ​​between the visible layer neurons and the hidden layer neurons when extracting features, thereby realizing continuous state transition.

[0038] In some embodiments, dividing the thermal runaway risk training data of the energy storage battery system into a state sequence of multiple time steps, and converting the state sequence of each time step into a vector form to obtain a state vector of each time step includes:

[0039] Obtain at least two thermal runaway risk parameter samples for each battery pack of the energy storage battery system, and preprocess the thermal runaway risk parameter samples;

[0040] Obtain a training data set from the preprocessed thermal runaway risk parameter samples;

[0041] The training data of the training data set are divided into state sequences of multiple time steps, and the state sequence of each time step is converted into a vector form to obtain a state vector of each time step.

[0042] In the technical solution of the embodiment of the present application, by preprocessing the thermal runaway risk parameter samples according to the parameter characteristics of the thermal runaway risk parameter samples, the difficulty of subsequent data processing can be reduced and the efficiency of thermal runaway risk prediction can be improved.

[0043] In some embodiments, after inputting the representation of the root node of the tree into an initial prediction model to train at least two neural network layers in the initial prediction model to obtain a target prediction model, the method further includes:

[0044] Obtaining a test sample set from the pre-processed thermal runaway risk parameter sample;

[0045] Testing the target prediction model based on the test data in the test sample set to obtain a test result;

[0046] The parameters of the target prediction model are adjusted according to the test results.

[0047] In the technical solution of the embodiment of the present application, the thermal runaway risk parameter samples are first divided into a training data set and a test data set, and then the initial prediction model is trained with the training data set to obtain a target prediction model. The target prediction model is then tested based on the test data set, and the parameters of the target prediction model are adjusted according to the test results. In this way, the test and adjustment can be performed after the initial prediction model is trained, thereby further improving the prediction ability of the target prediction model.

[0048] In some embodiments, processing the thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model to obtain the thermal runaway probability of each battery pack includes:

[0049] Use the target prediction model to predict the number of batteries at risk of thermal runaway in each battery pack of the energy storage battery system;

[0050] The thermal runaway probability of each battery pack is calculated according to the number of batteries with thermal runaway risks in each battery pack.

[0051] In the technical solution of the embodiment of the present application, the number of batteries at risk of thermal runaway in each battery pack of the energy storage battery system is first predicted by a target prediction model, and then the thermal runaway probability of each battery pack in the energy storage battery system is calculated based on the number of batteries at risk of thermal runaway in each battery pack. In this way, the thermal runaway probability of each battery pack can be calculated from the perspective of each battery in each battery pack, and then the thermal runaway probability of each battery pack can be calculated from a finer dimension, further improving the accuracy of thermal runaway risk prediction.

[0052] In some embodiments, after predicting the total probability of thermal runaway risk of the energy storage battery system based on the thermal runaway probability of each battery pack, the method further includes:

[0053] Determining a thermal runaway risk level of the energy storage battery system according to a total probability of thermal runaway risk of the energy storage battery system;

[0054] Thermal runaway warning information is generated according to the thermal runaway risk level, and a thermal runaway warning is performed based on the thermal runaway warning information.

[0055] In the technical solution of the embodiment of the present application, the thermal runaway risk level of the energy storage battery system is first determined based on the total probability of thermal runaway risk of the energy storage battery system, and then thermal runaway warning information is generated based on the thermal runaway risk level. Thermal runaway warning is performed based on the thermal runaway warning information, thereby enabling timely warning of potential safety hazards, effectively reducing the risk of accidents, and improving the safety and reliability of the energy storage battery system.

[0056] In a second aspect, the present application provides a thermal runaway risk prediction device, comprising:

[0057] A parameter acquisition module is used to obtain thermal runaway risk parameters of each battery pack in the energy storage battery system;

[0058] a model processing module, configured to process the thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model to obtain a thermal runaway probability for each battery pack, wherein the target prediction model is pre-trained based on state vectors for multiple time steps, and the state vectors are used to enhance the model's memory of past state sequences of the thermal runaway risk parameters;

[0059] The risk prediction module is used to predict the total probability of thermal runaway risk of the energy storage battery system according to the thermal runaway probability of each battery pack.

[0060] In a third aspect, the present application provides a thermal runaway risk prediction device, which includes: a memory, a processor, and a thermal runaway risk prediction program stored in the memory and executable on the processor, wherein the thermal runaway risk prediction program implements the thermal runaway risk prediction method described above when executed by the processor.

[0061] In a fourth aspect, the present application provides a storage medium storing a thermal runaway risk prediction program, which implements the thermal runaway risk prediction method described above when executed by a processor.

[0062] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] FIG1 is a flow chart of an embodiment of a method for predicting thermal runaway risk according to the present invention;

[0064] FIG2 is a flow chart of another embodiment of the thermal runaway risk prediction method of the present application;

[0065] FIG3 is a schematic diagram of model training in some embodiments of the thermal runaway risk prediction method of the present application;

[0066] FIG4 is a flow chart of another embodiment of the thermal runaway risk prediction method of the present application;

[0067] FIG5 is a flow chart of another embodiment of the thermal runaway risk prediction method of the present application;

[0068] FIG6 is a flow chart of another embodiment of the thermal runaway risk prediction method of the present application;

[0069] FIG7 is a flow chart of another embodiment of the thermal runaway risk prediction method of the present application;

[0070] FIG8 is a flow chart of another embodiment of the thermal runaway risk prediction method of the present application;

[0071] FIG9 is a specific flow chart of some embodiments of the thermal runaway risk prediction method of the present application;

[0072] FIG10 is a structural block diagram of an embodiment of a thermal runaway risk prediction device of the present application;

[0073] FIG11 is a schematic structural diagram of a thermal runaway risk prediction device for a hardware operating environment according to an embodiment of the present application. DETAILED DESCRIPTION

[0074] The following embodiments of the technical solution of the present application will be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present application and are therefore only examples and are not intended to limit the scope of protection of the present application.

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned figure descriptions are intended to cover non-exclusive inclusions.

[0076] In the description of the embodiments of this application, the technical terms "first" and "second" are used only to distinguish different objects and should not be understood to indicate or imply relative importance or implicitly specify the quantity, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, the meaning of "plurality" is more than two, unless otherwise clearly and specifically defined.

[0077] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0078] In the description of the embodiments of this application, the term "and / or" is simply a description of the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent the following three situations: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0079] In the description of the embodiments of the present application, the term "multiple" refers to more than two (including two). Similarly, "multiple groups" refers to more than two groups (including two groups), and "multiple pieces" refers to more than two pieces (including two pieces).

[0080] In the description of the embodiments of the present application, the technical terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the embodiments of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the embodiments of the present application.

[0081] In the description of the embodiments of the present application, unless otherwise clearly specified and limited, technical terms such as "installed", "connected", "connected", and "fixed" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the internal connection of two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the embodiments of the present application can be understood according to specific circumstances. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0082] Currently, market developments indicate that batteries are increasingly being used. They are not only used in energy storage systems such as hydropower, thermal, wind, and solar power plants, but are also widely used in electric vehicles like electric bicycles, electric motorcycles, and electric cars, as well as in military equipment and aerospace. As battery applications continue to expand, market demand is also growing.

[0083] High-voltage energy storage battery systems are widely used in energy storage, grid peak-valley balancing, and other fields. However, during long-term operation, these systems may face safety hazards such as overcharging, over-discharging, and overheating. If not promptly addressed, these systems can cause serious accidents such as fires and explosions. Existing methods for predicting thermal runaway risk in energy storage batteries struggle to capture the complex relationships between the various variables in these systems, resulting in low prediction accuracy and significant safety risks.

[0084] For example, a related method for predicting the risk of thermal runaway of energy storage batteries is to convert the battery energy storage system into a two-node electrothermal model and obtain the historical data of the ambient temperature and the historical data of the surface temperature in the two-node electrothermal model; predict the first surface temperature at the current moment based on the above parameters; establish a multi-scale long short-term memory network (LSTM) network, input the historical surface temperature, historical ambient temperature and historical electrothermal parameters of the battery into the multi-scale LSTM network, and predict the second surface temperature at the current moment; perform weighted summation of the first surface temperature and the second surface temperature to calculate the coupled surface temperature at the current moment; establish a classification network with cross entropy as the loss function, input the coupled surface temperature into the classification network to determine whether the coupled surface temperature reaches or exceeds the preset threshold; when the coupled surface temperature reaches or exceeds the preset threshold, the battery energy storage system is warned, which can quickly and accurately warn the battery energy storage system.

[0085] However, related energy storage battery thermal runaway risk prediction methods have the following problems: 1. The battery energy storage system is modeled as a two-node electrothermal model, which may not fully capture the complex thermal coupling characteristics of the actual battery system, resulting in prediction errors; 2. The construction and training of multi-scale LSTM networks may require a large amount of computing resources and data, and the network design and parameter adjustment may be relatively complex.

[0086] Specifically, for the thermal runaway risk prediction problem in high-voltage energy storage battery systems, existing methods struggle to capture the complex relationships between multiple variables in the system through mathematical modeling. Furthermore, because fault data accounts for a relatively small proportion of battery data, common supervised methods suffer from significant errors in judging small sample sizes. The LSTM algorithm is not particularly effective at feature learning and dimensionality reduction, making it unable to fully utilize the information in the data for modeling and prediction. Furthermore, LSTM models are typically single-layer recurrent neural networks, which may not effectively capture high-order relationships in the data.

[0087] Therefore, in order to solve the problem that the relevant energy storage battery thermal runaway risk prediction methods are difficult to capture the complex correlation between multiple variables in the energy storage battery system, resulting in low prediction accuracy and high safety risks, the thermal runaway risk parameters can be processed layer by layer through at least two neural network layers in the target prediction model, and the target prediction model is pre-trained based on the state vectors of multiple time steps in advance, so as to enhance the model's memory ability of past state sequences, thereby enabling the model to better learn the dynamic characteristics of the energy storage battery system, capture the complex temporal correlation between multiple variables, improve the accuracy of thermal runaway risk prediction, and reduce the safety risk of the energy storage battery system.

[0088] In a first aspect, an embodiment of the present application provides a method for predicting thermal runaway risk. Referring to FIG. 1 , FIG. 1 is a flow chart of an embodiment of a method for predicting thermal runaway risk.

[0089] In this embodiment, the thermal runaway risk prediction method includes the following steps:

[0090] Step S10: Obtain thermal runaway risk parameters of each battery pack of the energy storage battery system.

[0091] The execution subject of this embodiment can be a thermal runaway risk prediction device with data processing, network communication and program running functions, such as a computer, a host computer, etc., or other electronic devices that can achieve the same or similar functions. This embodiment does not impose any restrictions on this.

[0092] The thermal runaway risk parameters of each battery pack of the energy storage battery system include, but are not limited to, parameters that can reflect the thermal runaway risk of the battery, such as voltage, current, temperature, and CO concentration of each battery pack of the energy storage battery system. This embodiment does not limit this.

[0093] Of course, in order to further improve the accuracy of thermal runaway risk prediction, in this embodiment, the thermal runaway probability of each battery pack can also be calculated from the perspective of the batteries in each battery pack, and then the thermal runaway probability of each battery pack can be calculated from a finer dimension, that is, the thermal runaway risk parameters of each battery pack of the energy storage battery system include but are not limited to the voltage, current, temperature and CO concentration of the single battery of each battery pack of the energy storage battery system, and other parameters that can reflect the thermal runaway risk of the battery. This embodiment does not limit this.

[0094] Obtaining the thermal runaway risk parameters of each battery pack of the energy storage battery system can be achieved by monitoring the thermal runaway risk parameters of each battery pack of the energy storage battery system in real time through sensors pre-installed on each battery pack of the energy storage battery system. For example, a gas sensor pre-installed on each battery pack of the energy storage battery system can be used to monitor the CO concentration generated by battery overcharging, over-discharging, etc. in real time.

[0095] Step S20: Processing the thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model to obtain the thermal runaway probability of each battery pack, wherein the target prediction model is pre-trained based on state vectors of multiple time steps, and the state vectors are used to enhance the model's memory ability of the past state sequence of the thermal runaway risk parameters.

[0096] The target prediction model represents the thermal runaway risk parameters layer by layer through at least two stacked neural network layers. Each neural network layer can capture the deep thermal runaway correlation between thermal runaway risk parameters of different depths. The target prediction model includes at least two stacked neural network layers. The input of the bottom neural network layer is the thermal runaway risk parameter. The feature of the output of the bottom neural network layer is the thermal runaway correlation between the thermal runaway risk parameters. The input of the upper layer of the bottom neural network layer is the feature of the output of the bottom neural network layer. The feature of the output of the upper layer of the bottom neural network layer is the deeper thermal runaway correlation between the thermal runaway risk parameters. And so on, until the topmost neural network layer outputs the thermal runaway probability of each battery pack according to the deep thermal runaway correlation between the data inputs, wherein the deep thermal runaway correlation between the thermal runaway risk parameters includes but is not limited to the nonlinear relationship, dynamic relationship and time series relationship between the thermal runaway risk parameters.

[0097] In a specific implementation, the target prediction model includes but is not limited to Convolutional Neural Networks (CNN), Deep Belief Network (DBN) model, Continuous Deep Belief Network (CDBN) model, etc. Among them, the CNN model includes at least two stacked hidden layers, the DBN model includes at least two stacked Restricted Boltzmann Machines (RBM), and the CDBN model includes at least two stacked Continuous Restricted Boltzmann Machines (CRBM).

[0098] In order to more accurately capture the complex correlations between multiple variables in the energy storage battery system, in this embodiment, the target prediction model includes a CDBN model. The CDBN model can more accurately capture the complex correlations between multiple variables in the energy storage battery system, mainly due to its deep structure and feature learning capabilities:

[0099] 1. Deep Structure: The CDBN model consists of at least two stacked continuous restricted Boltzmann machines (CRBMs). This deep structure enables the CDBN model to learn different levels of abstract representation. In energy storage battery systems, nonlinear and high-order relationships may exist between different variables. Through multiple layers of CRBMs, the CDBN can gradually learn these complex relationships, thereby better capturing and expressing information in the energy storage battery system.

[0100] 2. Feature Learning Capabilities: CRBM, as a fundamental component of the CDBN model, possesses powerful feature learning capabilities. CRBM can automatically learn important features from input data through unsupervised learning. In energy storage battery systems, factors such as the thermal runaway risk parameters of each battery pack may affect the thermal runaway probability of each pack, but these effects are not easily observed or understood. Leveraging CRBM's feature learning capabilities, CDBN can extract useful features from raw data and use them as input for next-level learning and prediction, thereby better capturing the complex relationships within the energy storage battery system.

[0101] In addition, RBM generally processes binary data such as 0 or 1, but this embodiment needs to process continuous real-valued numbers such as voltage, current, temperature, and CO concentration. Therefore, CRBM is introduced in this embodiment. By processing the thermal runaway risk parameters through at least two stacked CRBMs in the CDBN model, a deep thermal runaway correlation between the thermal runaway risk parameters is obtained.

[0102] In summary, the CDBN model, through its deep structure and feature learning capabilities, can more accurately capture the complex correlations between multiple variables in the energy storage battery system and improve the accuracy of thermal runaway probability prediction.

[0103] In the specific implementation, X in As the input data of the CDBN model, Y out As the output data of the CDBN model, X in =[V,I,T,CO],Y out =P i , where P i is the thermal runaway probability of the mth battery pack.

[0104] In addition, in order to enhance the target prediction model's ability to remember past state sequences, so that the model can better learn the dynamic characteristics of the energy storage battery system and capture the complex temporal correlation between multiple variables, in this embodiment, the target prediction model is pre-trained based on the state vectors of multiple time steps. The specific steps may be to divide the thermal runaway risk training data of the energy storage battery system into a state sequence of multiple time steps, and convert the state sequence of each time step into a vector form to obtain the state vector of each time step. The state vector of each time step is input into the initial prediction model to train at least two neural network layers in the initial prediction model to obtain the target prediction model.

[0105] For ease of understanding, the following examples are given, but they do not limit the embodiments of the present application. In one example, the thermal runaway risk training data of the energy storage battery system (such as voltage (V), current (I), temperature (T), CO concentration) is divided into a state sequence of multiple time steps, and the state sequence of multiple time steps is expressed in vector form to obtain the state vector of each time step. For each time step t, the state vector S t =[V t ,T t ,I t ,CO t ], contains multiple state information, among which V t represents the voltage at time step t, T t represents the temperature at time step t, I t represents the current at time step t, CO t represents the CO concentration at time step t.

[0106] Step S30: predicting the total probability of thermal runaway risk of the energy storage battery system according to the thermal runaway probability of each battery pack.

[0107] Predicting the total probability of thermal runaway risk of the energy storage battery system based on the thermal runaway probability of each battery pack can be done by taking the highest thermal runaway probability of each battery pack as the total probability of thermal runaway risk of the energy storage battery system, so as to promptly discover battery packs with thermal runaway risks; or it can be done by sorting the thermal runaway probabilities of each battery pack from large to small, and taking the average of a preset number of thermal runaway probabilities at the top of the sort as the total probability of thermal runaway risk of the energy storage battery system, so as to reduce the impact of individual abnormal data and improve the accuracy of the total probability of thermal runaway risk, wherein the preset number can be pre-set, for example, set to 3.

[0108] Of course, in order to fully consider the different urgency levels of different thermal runaway probabilities and further improve the accuracy of the total thermal runaway risk probability, in this embodiment, the weight corresponding to the thermal runaway probability of each battery pack can also be calculated, and the total thermal runaway risk probability of the energy storage battery system can be calculated based on the thermal runaway probability and weight of each battery pack. This embodiment does not limit this.

[0109] For ease of understanding, the following examples are provided, but are not intended to limit the embodiments of the present application. In one example, the weights corresponding to the thermal runaway probabilities of the respective battery packs may be calculated using a probability formula, where the probability formula is as follows:

[0110] Where k i is the weight corresponding to the thermal runaway probability of the mth battery pack, P i is the thermal runaway probability of the mth battery pack, which is graded according to four risk levels. When the thermal runaway risk of a battery pack is higher, its weight is larger.

[0111] Calculating the total probability of thermal runaway risk of the energy storage battery system according to the thermal runaway probability and weight of each battery pack can be performed by using a total probability calculation formula according to the thermal runaway probability and weight of each battery pack. The total probability calculation formula is as follows:

[0112] P 总 =∑k i ·P i / ∑k i

[0113] Where, P 总 is the total probability of thermal runaway risk of the energy storage battery system, k i is the weight corresponding to the thermal runaway probability of the mth battery pack, P iis the thermal runaway probability of the mth battery pack.

[0114] This embodiment processes thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model, and the target prediction model is pre-trained based on state vectors of multiple time steps in advance, thereby enhancing the model's memory ability of past state sequences, thereby enabling the model to better learn the dynamic characteristics of the energy storage battery system, capture the complex temporal correlations between multiple variables, improve the accuracy of thermal runaway risk prediction, and reduce the safety risks of the energy storage battery system.

[0115] In some embodiments, referring to FIG. 2 , in this embodiment, before step S10, the method further includes:

[0116] Step S01: dividing the thermal runaway risk training data of the energy storage battery system into a state sequence of multiple time steps, and converting the state sequence of each time step into a vector form to obtain a state vector of each time step.

[0117] In order to further enhance the model's ability to remember past state sequences, in this embodiment, the variables of the battery system are formed into a state vector, and the model input is adjusted through a recursive network to train the model.

[0118] Step S02: Assigning values ​​to the representations of the child nodes of the tree according to the state vectors of each time step, and recursively combining the representations of the child nodes into the representation of the parent node starting from the root node of the tree until the representation of the entire tree is obtained.

[0119] Adjusting the model input through the recursive network can specifically be to regard the state vector of each time step as a leaf node of a tree structure. The input of each node in the tree is composed of the representations of its child nodes. Starting from the root node of the tree, the representations of the child nodes are recursively combined into the representation of the parent node until the representation of the entire tree is obtained. The representation of the root node of the tree is used as the input of the CDBN model to train at least two stacked CRBMs in the CDBN model.

[0120] Step S03: Inputting the representation of the root node of the tree into the initial prediction model to train at least two neural network layers in the initial prediction model to obtain a target prediction model.

[0121] For ease of understanding, reference is made to Figure 3 for illustration, but this does not limit the embodiments of the present application. Figure 3 is a schematic diagram of model training for some embodiments of the thermal runaway risk prediction method of the present application. In the figure, the thermal runaway risk training data of the energy storage battery system (such as voltage (V), current (I), temperature (T), CO concentration) are divided into state sequences of multiple time steps, and the state sequences of multiple time steps are expressed in vector form to obtain the state vectors of each time step. For each time step t, the state vector S t =[V t ,T t ,I t ,CO t ], contains multiple state information, among which V t represents the voltage at time step t, T t represents the temperature at time step t, I t represents the current at time step t, CO t Represents the CO concentration at time step t. The state vector of each time step is regarded as a leaf node of a tree structure. The input of each node in the tree is composed of the representations of its child nodes. Starting from the root node of the tree, the representations of the child nodes are recursively combined into the representation of the parent node until the representation of the entire tree is obtained. The representation of the root node of the tree is used as the input of the CDBN model to train at least two stacked CRBMs in the CDBN model.

[0122] This embodiment forms the variables of the battery system into a state vector and adjusts the model input through a recursive network to train the model, thereby further enhancing the model's ability to remember past state sequences, thereby enabling the model to better learn the dynamic characteristics of the energy storage battery system and capture the complex temporal correlations between multiple variables.

[0123] In some embodiments, referring to FIG. 4 , in this embodiment, step S03 includes:

[0124] Step S031: Inputting the representation of the root node of the tree as a model into an initial prediction model to perform forward pre-training on at least two neural network layers in the initial prediction model.

[0125] In this embodiment, by performing forward pre-training and backward fine-tuning on at least two neural network layers in the initial prediction model, the model training effect can be improved and the prediction ability of the target prediction model can be enhanced.

[0126] Step S032: After the forward pre-training of the neural network layer is completed, the neural network layer is backward fine-tuned through the back propagation network in the initial prediction model to obtain a target prediction model.

[0127] Since RBN has a poor ability to extract abstract feature vectors, in this embodiment, a CDBN model is constructed including at least two stacked CRBM and back propagation (BP) networks, and Gaussian distribution is used to process input feature vectors for thermal runaway risk prediction.

[0128] For ease of understanding, reference is made to Figure 3 for illustration, but this does not limit the present application. Figure 3 is a schematic diagram of model training for some embodiments of the thermal runaway risk prediction method of the present application. In the figure, the CDBN model is composed of a series of superimposed CRBMs and a layer of BP network, wherein the input of the bottom CRBM is the training data Xin, and the input of the next layer of CRBM is the features of the output of the previous layer of CRBM. Each two adjacent layers of the network are used as a CRBM. After layer-by-layer training, the parameters of the entire network are initialized. The BP network is added to the top layer of the CDBN, and the training data set is used for back propagation and gradient descent algorithm to fine-tune the parameters of the entire network.

[0129] In order to fit the training dataset, the network parameters need to be further adjusted. Therefore, the training process of the CDBN model can be divided into two processes: forward pre-training and backward fine-tuning.

[0130] Forward pre-training can be to take the training data as input from the bottom layer and train the first layer of CRBM. Each CRBM will extract the feature representation of the corresponding level and use its output as the input of the next layer of CRBM. The above process is repeated, training layer by layer until the desired depth is reached.

[0131] Backward fine-tuning requires a backpropagation algorithm. After unsupervised training, except for the top-level CRBM, the remaining layers are divided into upward weights and downward weights. Based on the gradient descent theory, the downward weights between layers are modified to generate the bottom-level weights. In addition, backward fine-tuning can modify the upward weights between layers, that is, iteratively optimize the weights and offsets in the network to achieve the minimum mean square error, complete the fine-tuning of the entire model to obtain the final prediction results, and at the same time improve the learning ability of the network.

[0132] The CRBM output of the last layer is used as the input of the top-level BP network for regression prediction. The model is adjusted using gradient descent. The weights are updated with the goal of minimizing the error between the model's predicted output and the actual output. The learning rate, step size, and number of iterations are set. The network is optimized using the backpropagation algorithm, and the loss rate and accuracy are calculated.

[0133] For ease of understanding, the following examples are provided, but are not intended to limit the present application. In one example, the training process of the CDBN model can be divided into two steps: forward pre-training and backward fine-tuning:

[0134] (1) Forward pre-training includes the following steps:

[0135] RBMs typically process binary data such as 0 or 1, but this embodiment needs to process continuous real-valued numbers such as voltage, current, temperature, and CO concentration. These data have a time dimension and multiple feature dimensions, requiring appropriate model design. Therefore, CRBMs are introduced. The CD algorithm is used to initialize the states of the visible layer neurons and train the first CRBM. The weights and offsets are used to retain the state of the previous CRBM hidden layer units as the next input vector. After training, the second CRBM is stacked on top of the first CRBM. The above steps are repeated multiple times until all CRBM layers are trained. Since the training set data is labeled, when training the top-level CRBM, the neurons with the classification labels also need to be trained.

[0136] A. Define the class of the hidden layer of the Multilayer Perceptron (MLP).

[0137] The input layer and hidden layer are fully connected, and the output of the hidden layer is the number of neurons. If the input is an n-dimensional vector and the hidden layer has m neurons, there are a total of n*m weights, which can be represented as an n-row, m-column matrix, where each column corresponds to the connection weight of each neuron in the hidden layer. The training model here uses a multidimensional vector as input, with each row serving as the input to the MLP. The activation function can be set to the sigmoid function. A softmax regression function is then defined, which multiplies the output of the previous layer by the weights and adds the bias to calculate the probability of each sample being classified as one of the categories. Combining the two categories above results in a three-layer MLP.

[0138] B. Create a CRBM class.

[0139] The sigmoid transfer function is retained, the subsequent data discretization process is removed, and a Gaussian random noise variable with a mean of 0 and a variance of 1 is added to achieve continuous state transition. When extracting features, the visible layer and hidden layer neurons have continuous state values.

[0140] In CRBM, the neuron state of the visible layer is: ν i =σ(Σ j w ij h j +a i +λ·N i (0,1)

[0141] The hidden layer neuron state is: h j =σ(Σ i ν i w ij +b j +λ·N j (0,1)

[0142] Where λ is a constant, and N(0,1) represents a Gaussian random noise variable with a mean of 0 and a variance of 1. The i-th visible layer neuron is represented by v i , the jth hidden layer neuron is represented by h j , w ij represents the weight between the i-th visible layer neuron and the j-th hidden layer neuron, a i is the bias of the i-th visible layer neuron, b j is the bias of the jth hidden layer neuron. Since σ(x) is a continuous function with a range of (0,1), ν i and h j It converges to the interval (0,1) and is continuous within the interval.

[0143] Based on the collected data samples, CRBM uses a specific algorithm to repeatedly adjust the network structure parameters, thereby fitting the samples input to the system. The maximum likelihood learning method is used to maximize the log-likelihood value as the objective function to determine the parameters in the network.

[0144] Similarly, CRBM training focuses on solving the problem of approximate optimization. This approach draws on the principles of the simulated annealing algorithm to maximize the joint probability distribution of the visible and hidden layers. The contrastive divergence algorithm, a commonly used fast learning algorithm for RBMs, is introduced. This algorithm uses sample data to directly initialize the states of visible layer neurons, replacing random initialization. This allows the partial derivatives of weights and biases to be calculated, thereby completing the weight update. The update formula is as follows: Δa i =ε(<ν i > data -<ν i > rec ) Δb j =ε( <h j > data - <h j > rec )

[0145] Here, ε represents the learning efficiency, and the parameters with subscripts “data” and “rec” represent the input data and the reconstructed data, respectively.

[0146] The CRBM is trained by setting the number of layers, units per layer, and iterations. The number of units in the visible and hidden layers is defined, and a complete forward and backward propagation process and objective function are constructed. The CRBM is trained using the CD-k algorithm. The Gibbs sampling step size is set, and the activation probabilities of the hidden layers are derived from the visible layers. The hidden layer states are randomly generated and used to reconstruct the visible layer states. This is repeated k times, and the updated values ​​of each parameter are calculated with the cross-entropy of the original input. Starting from the input layer, the weights and biases of each CRBM are trained layer by layer to generate the upper layer data.

[0147] (2) Backward fine-tuning includes the following steps:

[0148] Supervised training requires a backpropagation algorithm. After unsupervised training, except for the top-level CRBM, the remaining layers are divided into upward weights and downward weights. Based on the gradient descent theory, the downward weights between layers are modified to generate the bottom-level weights. In addition, reverse fine-tuning can modify the upward weights between layers, that is, iteratively optimize the weights and offsets in the network to achieve the minimum mean square error, complete the fine-tuning of the entire model to obtain the final prediction results, and at the same time improve the network's learning ability.

[0149] The CRBM output of the last layer is used as the input of the top-level BP network for regression prediction. The model is adjusted using gradient descent. The weights are updated with the goal of minimizing the error between the model's predicted output and the actual output. The learning rate, step size, and number of iterations are set. The network is optimized using the backpropagation algorithm, and the loss rate and accuracy are calculated.

[0150] 3) Establish an evaluation system

[0151] The prediction model is evaluated using the RMSE function and the coefficient of determination (R-Square, R2). RMSE, or root mean square error, is the square root of the ratio of the square of the deviation between the predicted value and the actual value to the number of observations. The formula is as follows:

[0152] The coefficient of determination (R-Square, R2) is a statistical indicator used to evaluate the model. Its value is between (0,1]. The closer it is to 1, the better the prediction effect of the model.

[0153] y out represents the predicted value, Indicates the actual value, y mean Represents the mean.

[0154] This embodiment performs forward pre-training and backward fine-tuning on at least two neural network layers in the initial prediction model, thereby improving the model training effect and enhancing the prediction capability of the target prediction model.

[0155] In some embodiments, referring to FIG. 5 , in this embodiment, step S031 includes:

[0156] Step S0311: Input the representation of the root node of the tree as input data into the neural network layer in the initial prediction model, and reconstruct the input data back to original data through the neural network layer to obtain an output result.

[0157] In the battery system, due to the small amount of abnormal data, it may be difficult for the model to capture the characteristics of the abnormal data. Therefore, in order to solve this problem, in this embodiment, a reconstruction error loss function is introduced in the model training process to improve the forward pre-training.

[0158] Step S0312: Calculate the reconstruction error loss function of the neural network layer according to the input data and the output result, and perform forward pre-training on the neural network layer according to the reconstruction error loss function.

[0159] For ease of understanding, reference is made to FIG3 for illustration, but this does not limit the present application. FIG3 is a schematic diagram of model training for some embodiments of the thermal runaway risk prediction method of the present application. In the figure, in each forward pre-training stage, the reconstruction error for normal data is small, while the reconstruction error loss function for abnormal data is large. By introducing the reconstruction error as a loss function, the model will be forced to learn to distinguish between normal data and abnormal data, thereby improving the ability to identify abnormal data. The calculation formula of the reconstruction error is as follows: Reconstruction error = ||XX′|| 2

[0160] Where X is the input data, and X′ is the result of reconstructing the input data back to the original data through the model.

[0161] Step S0313: After the neural network layer completes forward pre-training, the hidden unit state of the neural network layer is input into the neural network layer in the initial prediction model as the input data of the next neural network layer, and the step of reconstructing the input data back to the original data through the neural network layer is returned until all neural network layers in the initial prediction model complete forward pre-training.

[0162] This embodiment improves forward pre-training by introducing a reconstruction error loss function during the model training process, thereby solving the problem that the model may have difficulty capturing the characteristics of abnormal data due to the small amount of abnormal data, thereby improving the model's prediction ability.

[0163] In one embodiment, the neural network layer includes: a continuous restricted Boltzmann machine, and reconstructing the input data back to the original data through the neural network layer to obtain the output result includes: reconstructing the input data back to the original data through the Gaussian distribution activation function of the continuous restricted Boltzmann machine to obtain the output result, and the Gaussian distribution activation function includes retaining the sigmoid transfer function, removing the data discretization process, and adding a Gaussian random noise variable. The sigmoid function.

[0164] In order to ensure that there are continuous state values ​​between the visible layer neurons and the hidden layer neurons when extracting features and realize continuous state transition, in this embodiment, the Gaussian distribution activation function of the continuous restricted Boltzmann machine includes retaining the sigmoid transfer function, removing the data discretization process, and adding the sigmoid function after adding the Gaussian random noise variable. The Gaussian distribution activation function of the continuous restricted Boltzmann machine is used to reconstruct the input data back to the original data to obtain the output result.

[0165] This embodiment reconstructs the input data back to the original data through the Gaussian distribution activation function of the continuous restricted Boltzmann machine to obtain the output result, thereby ensuring that there are continuous state values ​​between the visible layer neurons and the hidden layer neurons when extracting features, thereby achieving continuous state transition.

[0166] In some embodiments, referring to FIG. 5 , in this embodiment, step S032 includes:

[0167] Step S0321: After the forward pre-training of the neural network layer is completed, normal data and abnormal data in the thermal runaway risk training data are labeled.

[0168] In the battery system, due to the small amount of abnormal data, it may be difficult for the model to capture the characteristics of the abnormal data. Therefore, in order to solve this problem, in this embodiment, a reconstruction error loss function is introduced in the model training process to improve the forward pre-training, and the data is annotated for abnormalities in the backward fine-tuning of the model.

[0169] When obtaining thermal runaway risk training data, a binary label can be defined to indicate whether the temperature exceeds a threshold. This binary label can be used to label normal data and abnormal data in the thermal runaway risk training data. The thermal runaway risk training data corresponding to the temperature exceeding the threshold is abnormal data, and vice versa.

[0170] Step S0322: Calculate the reconstruction error of the normal data and the reconstruction error of the abnormal data respectively, and calculate the loss function according to the reconstruction error of the normal data and the reconstruction error of the abnormal data.

[0171] For ease of understanding, reference is made to Figure 3 for illustration, but this does not limit the present application. Figure 3 is a schematic diagram of model training for some embodiments of the thermal runaway risk prediction method of the present application. In the figure, normal data and abnormal data in the thermal runaway risk training data are marked, and the reconstruction error of the normal data and the reconstruction error of the abnormal data are calculated respectively. The loss function is calculated based on the reconstruction error of the normal data and the reconstruction error of the abnormal data. The specific formula is as follows: Weighted LOSS =Normal LOSS +AnormalLOSS .

[0172] In the formula, Weighted LOSS is the loss function, Normal LOSS is the reconstruction error of normal data, Anormal LOSS is the reconstruction error of abnormal data.

[0173] Step S0323: Based on the loss function, fine-tune the neural network layer through the back propagation network in the initial prediction model to obtain a target prediction model.

[0174] Supervised training requires a backpropagation algorithm. After unsupervised training, except for the top-level CRBM, the remaining layers are divided into upward weights and downward weights. Based on the gradient descent theory, the downward weights between layers are modified to generate the bottom-level weights. In addition, reverse fine-tuning can modify the upward weights between layers, that is, iteratively optimize the weights and offsets in the network to achieve the minimum mean square error, complete the fine-tuning of the entire model to obtain the final prediction results, and at the same time improve the network's learning ability.

[0175] The CRBM output of the last layer is used as the input of the top-level BP network for regression prediction. The model is adjusted using gradient descent. The weights are updated with the goal of minimizing the error between the model's predicted output and the actual output. The learning rate, step size, and number of iterations are set. The network is optimized using the backpropagation algorithm, and the loss rate and accuracy are calculated.

[0176] This embodiment improves forward pre-training by introducing a reconstruction error loss function during the model training process, and anomalies are marked on the data during the backward fine-tuning of the model, thereby solving the problem that the model may find it difficult to capture the characteristics of abnormal data due to the small amount of abnormal data, and further improving the model's prediction ability.

[0177] In some embodiments, the step S0322 includes: respectively calculating the reconstruction error of the normal data and the reconstruction error of the abnormal data, and obtaining the normal data weight and the abnormal data weight, wherein the abnormal data weight is greater than a preset threshold; and calculating the loss function based on the normal data weight, the reconstruction error of the normal data, the abnormal data weight, and the reconstruction error of the abnormal data.

[0178] In order to better capture the characteristics of abnormal data, in this embodiment, the data is annotated and a higher weight is assigned to the abnormal data samples through the weighted form of the loss function. The specific formula is as follows: Weighted LOSS =α·Normal LOSS +β·Anormal LOSS

[0179] Where, NormalLOSS is the loss function of normal data, Anormal LOSS is the loss function for abnormal data, through Anormal LOSS The model learns the characteristics and patterns of abnormal data. The reconstruction error between abnormal data and normal data is large. LOSS The higher the value of Anormal LOSS , the model will adjust the parameters to reduce the model's reconstruction error of abnormal data and improve the ability to identify abnormal samples. The specific formula is as follows: Anormal LOSS =||Input-Reconstructied|| 2

[0180] Where Input is the actual observation data of the battery system, and Reconstructied represents the data reconstructed by the CDBN model decoder.

[0181] This embodiment labels the data as abnormal and assigns higher weights to abnormal data samples through the weighted form of the loss function, so as to better capture the characteristics of abnormal data and further improve the model prediction ability.

[0182] In some embodiments, referring to FIG. 6 , in this embodiment, step S01 includes:

[0183] Step S011: obtaining at least two thermal runaway risk parameter samples of each battery pack of the energy storage battery system, and preprocessing the thermal runaway risk parameter samples.

[0184] In order to eliminate the dimensional influence between different indicators and reduce the difficulty of subsequent data processing, in this embodiment, for a high-voltage energy storage battery system, in order to reduce the difficulty of subsequent data processing, in this embodiment, the data of each dimension in the original data set is preprocessed according to the parameter characteristics of the thermal runaway risk parameters in the original data set.

[0185] The at least two thermal runaway risk parameter samples of each battery pack of the energy storage battery system include but are not limited to sample data such as voltage, current, temperature and CO concentration of each battery pack of the energy storage battery system, which is not limited in this embodiment.

[0186] When obtaining thermal runaway risk parameter samples, a binary label can be defined to indicate whether the temperature exceeds the threshold, so as to facilitate subsequent abnormal data annotation.

[0187] Preprocessing includes, but is not limited to, outlier removal and normalization. The specific processing method can be determined based on the parameter characteristics of the thermal runaway risk parameter sample. For example, in a high-voltage energy storage battery system, to eliminate the dimensionality effects between different indicators and reduce the difficulty of subsequent data processing, in this embodiment, the thermal runaway risk parameter sample is normalized. For example, data such as voltage, current, temperature, and CO concentration are normalized to the interval [0, 1], rendering them dimensionless.

[0188] In the specific implementation, normalization is performed through the Min-Max normalization method, and the normalization formula is:

[0189] Where, Represents the normalized index. represents each sample value, and and Represent the maximum and minimum values ​​of the samples respectively.

[0190] Step S012: Obtain a training data set from the preprocessed thermal runaway risk parameter samples.

[0191] The training data set can be obtained from the pre-processed thermal runaway risk parameter samples in the order of the acquisition time of the thermal runaway risk parameter samples, with the first 80% of the thermal runaway risk parameter samples as the training data set and the remaining part as the test data set.

[0192] Step S013: Divide the training data of the training data set into state sequences of multiple time steps, and convert the state sequence of each time step into a vector form to obtain a state vector of each time step.

[0193] For ease of understanding, the following examples are provided, but are not intended to limit the embodiments of the present application. In one example, the training data (e.g., voltage (V), current (I), temperature (T), and CO concentration) of a training data set are divided into state sequences of multiple time steps, and the state sequences of the multiple time steps are represented in vector form to obtain a state vector for each time step.

[0194] This embodiment pre-processes the thermal runaway risk parameter samples according to their parameter characteristics, thereby reducing the difficulty of subsequent data processing and improving the efficiency of thermal runaway risk prediction.

[0195] In some embodiments, referring to FIG. 6 , in this embodiment, after step S03, the following steps are further included:

[0196] Step S04: Acquire a test sample set from the pre-processed thermal runaway risk parameter samples.

[0197] In this embodiment, the preprocessed thermal runaway risk parameter samples are first divided into a training data set and a test data set, and then the initial prediction model is trained with the training data set to obtain a target prediction model. The target prediction model is then tested based on the test data set, and the parameters of the target prediction model are adjusted according to the test results. In this way, the test and adjustment can be performed after the initial prediction model is trained, thereby further improving the prediction ability of the target prediction model.

[0198] From the pre-processed thermal runaway risk parameter samples, the first 80% of the thermal runaway risk parameter samples can be used as a training data set and the remaining part can be used as a test data set in the order of the acquisition time of the thermal runaway risk parameter samples.

[0199] Step S05: testing the target prediction model based on the test data in the test sample set to obtain a test result.

[0200] The target prediction model is tested based on the test data in the test sample set. The test results can be obtained by using corresponding evaluation indicators (such as accuracy, precision, recall, etc.) to evaluate the performance of the target prediction model on the test data set according to the requirements of the specific task, and obtaining the test results of the target prediction model.

[0201] Step S06: Adjust the parameters of the target prediction model according to the test results.

[0202] It is difficult to ensure that the optimal parameters can be selected within a certain range during model training. For deep learning algorithms, parameter selection is crucial. Therefore, in order to further improve the predictive ability of the target prediction model, the parameters of the target prediction model can be adjusted according to the test results. The hyperparameters of the target prediction model can be adjusted according to the test results using intelligent optimization algorithms (such as genetic algorithms, simulated annealing, fruit fly optimization, particle swarm optimization, gray wolf optimization algorithm, etc.) to make the network structure of the target prediction model tend to be optimal.

[0203] This embodiment first divides the thermal runaway risk parameter samples into a training data set and a test data set, then trains the initial prediction model with the training data set to obtain a target prediction model, then tests the target prediction model based on the test data set, and then adjusts the parameters of the target prediction model according to the test results, so that the test adjustment can be performed after the initial prediction model is trained, thereby further improving the prediction ability of the target prediction model.

[0204] In some embodiments, referring to FIG. 7 , in this embodiment, step S20 includes:

[0205] Step S201: predicting the number of batteries with thermal runaway risks in each battery pack of the energy storage battery system using a target prediction model.

[0206] In this embodiment, the target prediction model is first used to predict the number of batteries at risk of thermal runaway in each battery pack of the energy storage battery system. Then, the thermal runaway probability of each battery pack in the energy storage battery system is calculated based on the number of batteries at risk of thermal runaway in each battery pack. This allows the thermal runaway probability of each battery pack to be calculated from the perspective of each battery in the battery pack, and further allows the thermal runaway probability of each battery pack to be calculated from a finer dimension, further improving the accuracy of the thermal runaway risk prediction.

[0207] Step S202: Calculating the thermal runaway probability of each battery pack according to the number of batteries with thermal runaway risk in each battery pack.

[0208] For ease of understanding, the following examples are given, but they do not limit the embodiments of the present application. In one example, it is assumed that the energy storage battery system has m battery packs, each battery pack contains n single cells, and the probability of thermal runaway is P i (i=1…m), The number of batteries at risk of thermal runaway:

[0209] Where Y out is the output data of the target prediction model, P i is the thermal runaway probability of the mth battery pack. The energy storage battery system has m battery packs, and each battery pack contains n single cells. is the number of batteries at risk of thermal runaway.

[0210] This embodiment first predicts the number of batteries at risk of thermal runaway in each battery pack of the energy storage battery system through a target prediction model, and then calculates the thermal runaway probability of each battery pack in the energy storage battery system based on the number of batteries at risk of thermal runaway in each battery pack. This allows the thermal runaway probability of each battery pack to be calculated from the perspective of each battery in the battery pack, and further allows the thermal runaway probability of each battery pack to be calculated from a finer dimension, further improving the accuracy of thermal runaway risk prediction.

[0211] In some embodiments, referring to FIG. 8 , in this embodiment, after step S30, the method further includes:

[0212] Step S40: determining a thermal runaway risk level of the energy storage battery system according to the total probability of thermal runaway risk of the energy storage battery system.

[0213] In order to provide timely warnings of potential safety hazards, effectively reduce the risk of accidents, and improve the safety and reliability of the energy storage battery system, in this embodiment, the thermal runaway risk level of the energy storage battery system is determined based on the total probability of thermal runaway risk of the energy storage battery system, and thermal runaway warning information is generated based on the thermal runaway risk level. Thermal runaway warning is then performed based on the thermal runaway warning information.

[0214] In specific implementations, when the total probability of thermal runaway risk is ≥70%, it can be considered that there is a high thermal runaway risk and immediate human intervention may be required. The thermal runaway risk level is level 2. When the total probability of thermal runaway risk is 50% ≤ and <70%, it can be considered that there is a thermal runaway risk and the risk needs to be investigated. The thermal runaway risk level is level 1. When the total probability of thermal runaway risk is <50%, it can be considered that there is no thermal runaway risk. The thermal runaway risk level is level 1. The specific formula is as follows:

[0215] Where Ptotal is the total probability of thermal runaway risk of the energy storage battery system.

[0216] Step S50: generating thermal runaway warning information according to the thermal runaway risk level, and performing a thermal runaway warning based on the thermal runaway warning information.

[0217] Different thermal runaway risk levels have different degrees of criticality, so different thermal runaway warning messages need to be generated based on the different thermal runaway risk levels. For example, when the thermal runaway risk level is level 2, the thermal runaway warning message will be "The energy storage battery system is at a high risk of thermal runaway, requiring immediate human intervention." When the thermal runaway risk level is level 1, the thermal runaway warning message will be "The energy storage battery system is at a risk of thermal runaway, requiring risk investigation." When the thermal runaway risk level is level 1, the thermal runaway warning message may not be generated, or the thermal runaway warning message may be "The energy storage battery system is at no risk of thermal runaway." This embodiment does not impose any restrictions on this.

[0218] Thermal runaway warning based on thermal runaway warning information can be performed by displaying the thermal runaway warning information on the thermal runaway warning device to perform thermal runaway warning; or the thermal runaway warning information can be sent to the terminal device corresponding to the thermal runaway warning device so that maintenance personnel can view it remotely. This embodiment does not limit this.

[0219] This embodiment first determines the thermal runaway risk level of the energy storage battery system based on the total probability of thermal runaway risk of the energy storage battery system, then generates thermal runaway warning information based on the thermal runaway risk level, and performs a thermal runaway warning based on the thermal runaway warning information, thereby providing timely warnings of potential safety hazards, effectively reducing the risk of accidents, and improving the safety and reliability of the energy storage battery system.

[0220] For ease of understanding, reference is made to FIG9 for illustration, but this does not limit the present application. FIG9 is a specific flow chart of some embodiments of the thermal runaway risk prediction method of the present application. In the figure, the thermal runaway risk prediction method mainly includes the following steps: obtaining thermal runaway risk parameter samples of each battery pack of the energy storage battery system, and preprocessing the data of each dimension in the thermal runaway risk parameter sample, obtaining a training sample set and a test sample set, constructing a CDBN model, and training the CDBN model based on the training data set through vector transformation and recursion (the training steps include updating model parameters, cross-validation, calculating errors, and intelligent optimization algorithms in the figure), obtaining a target CDBN model that meets the termination conditions, and inputting the test data set into the target CDBN model. The standard CDBN model tests and adjusts the target CDBN model, and deploys the target CDBN model to obtain at least two thermal runaway risk parameters for each battery pack of the energy storage battery system. The thermal runaway risk parameters are processed layer by layer through at least two stacked CRBMs in the target CDBN model to obtain the thermal runaway probability of each battery pack. The total thermal runaway risk probability of the energy storage battery system is predicted based on the thermal runaway probability of each battery pack. The thermal runaway risk level of the energy storage battery system is determined based on the total thermal runaway risk probability of the energy storage battery system. Thermal runaway warning information is generated based on the thermal runaway risk level, and a thermal runaway warning is performed based on the thermal runaway warning information.

[0221] In a second aspect, referring to FIG. 10 , an embodiment of the present application further provides a thermal runaway risk prediction device, the device comprising:

[0222] A parameter acquisition module 10 is used to obtain thermal runaway risk parameters of each battery pack of the energy storage battery system;

[0223] a model processing module 20 for processing the thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model to obtain a thermal runaway probability for each battery pack, wherein the target prediction model is pre-trained based on state vectors of multiple time steps, and the state vectors are used to enhance the model's ability to memorize past state sequences of the thermal runaway risk parameters;

[0224] The risk prediction module 30 is configured to predict the total probability of thermal runaway risk of the energy storage battery system according to the thermal runaway probability of each battery pack.

[0225] This embodiment processes thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model, and the target prediction model is pre-trained based on state vectors of multiple time steps in advance, thereby enhancing the model's memory ability of past state sequences, thereby enabling the model to better learn the dynamic characteristics of the energy storage battery system, capture the complex temporal correlations between multiple variables, improve the accuracy of thermal runaway risk prediction, and reduce the safety risks of the energy storage battery system.

[0226] Other embodiments or specific implementations of the thermal runaway risk prediction device described in this application can refer to the above-mentioned method embodiments and will not be repeated here.

[0227] In a third aspect, referring to FIG11 , an embodiment of the present application further proposes a thermal runaway risk prediction device, which may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to implement connection and communication between these components. The user interface 1003 may include a display screen. Optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. In this application, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a wireless fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable memory (NVM), such as a disk storage. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0228] Those skilled in the art will understand that the structure shown in FIG11 does not constitute a limitation on the thermal runaway risk prediction device, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0229] As shown in FIG. 11 , the memory 1005 , which is identified as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a thermal runaway risk prediction program.

[0230] In the thermal runaway risk prediction device shown in Figure 11, the network interface 1004 is mainly used to connect to the background server and communicate data with the background server; the user interface 1003 is mainly used to connect to the user device; the thermal runaway risk prediction device calls the thermal runaway risk prediction program stored in the memory 1005 through the processor 1001, and executes the thermal runaway risk prediction method provided in the embodiment of the present application.

[0231] In a fourth aspect, an embodiment of the present application further proposes a storage medium, on which a thermal runaway risk prediction program is stored. When the thermal runaway risk prediction program is executed by a processor, the thermal runaway risk prediction method as described above is implemented.

[0232] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application, and they should all be included in the scope of the claims and specification of the present application. In particular, as long as there is no structural conflict, the various technical features mentioned in the various embodiments can be combined in any way. The present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions that fall within the scope of the claims.

Claims

1. A thermal runaway risk prediction method, wherein: The thermal runaway risk prediction method includes: Obtain thermal runaway risk parameters for each battery pack in the energy storage battery system; Processing the thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model to obtain a thermal runaway probability for each battery pack, wherein the target prediction model is pre-trained based on state vectors of multiple time steps, and the state vectors are used to enhance the model's memory of past state sequences of the thermal runaway risk parameters; The total probability of thermal runaway risk of the energy storage battery system is predicted based on the thermal runaway probability of each battery pack.

2. The thermal runaway risk prediction method according to claim 1, wherein: Before obtaining the thermal runaway risk parameters of each battery pack of the energy storage battery system, the method further includes: The thermal runaway risk training data of the energy storage battery system is divided into a state sequence of multiple time steps, and the state sequence of each time step is converted into a vector form to obtain the state vector of each time step; Assigning representations of child nodes of the tree according to the state vectors of each time step, and recursively combining the representations of the child nodes into a representation of the parent node starting from the root node of the tree until a representation of the entire tree is obtained; The representation of the root node of the tree is input into an initial prediction model to train at least two neural network layers in the initial prediction model to obtain a target prediction model.

3. The thermal runaway risk prediction method according to claim 2, wherein: The step of inputting the representation of the root node of the tree into an initial prediction model to train at least two neural network layers in the initial prediction model to obtain a target prediction model comprises: Inputting the representation of the root node of the tree into an initial prediction model as a model to perform forward pre-training on at least two neural network layers in the initial prediction model; After the neural network layer completes the forward pre-training, the neural network layer is backward fine-tuned through the back propagation network in the initial prediction model to obtain the target prediction model.

4. The thermal runaway risk prediction method according to claim 3, wherein: The step of inputting the representation of the root node of the tree as a model into an initial prediction model to perform forward pre-training on at least two neural network layers in the initial prediction model comprises: Inputting the representation of the root node of the tree as input data into a neural network layer in an initial prediction model, and reconstructing the input data back into original data through the neural network layer to obtain an output result; Calculating a reconstruction error loss function of the neural network layer according to the input data and the output result, and performing forward pre-training on the neural network layer according to the reconstruction error loss function; After the neural network layer completes forward pre-training, the hidden unit state of the neural network layer is input into the neural network layer in the initial prediction model as input data of the next neural network layer, and the step of reconstructing the input data back to the original data through the neural network layer is returned until all neural network layers in the initial prediction model complete forward pre-training.

5. The thermal runaway risk prediction method according to claim 4, wherein: After the forward pre-training of the neural network layer is completed, the neural network layer is fine-tuned backward through the back propagation network in the initial prediction model to obtain a target prediction model, including: After the neural network layer completes forward pre-training, labeling normal data and abnormal data in the thermal runaway risk training data; Respectively calculating a reconstruction error of the normal data and a reconstruction error of the abnormal data, and calculating a loss function according to the reconstruction error of the normal data and the reconstruction error of the abnormal data; The target prediction model is obtained by fine-tuning the neural network layer backward through the back propagation network in the initial prediction model based on the loss function.

6. The thermal runaway risk prediction method according to claim 5, wherein: The respectively calculating the reconstruction error of the normal data and the reconstruction error of the abnormal data, and calculating the loss function according to the reconstruction error of the normal data and the reconstruction error of the abnormal data, comprises: Respectively calculating the reconstruction error of the normal data and the reconstruction error of the abnormal data, and obtaining the normal data weight and the abnormal data weight, wherein the abnormal data weight is greater than a preset threshold; A loss function is calculated based on the normal data weight, the reconstruction error of the normal data, the abnormal data weight, and the reconstruction error of the abnormal data.

7. The thermal runaway risk prediction method according to claim 4, wherein: The neural network layer includes: a continuous restricted Boltzmann machine, and the input data is reconstructed back to original data through the neural network layer to obtain an output result, including: The input data is reconstructed back to the original data through the Gaussian distribution activation function of the continuous restricted Boltzmann machine to obtain an output result, wherein the Gaussian distribution activation function includes a sigmoid function after retaining the sigmoid transfer function, removing the data discretization process, and adding a Gaussian random noise variable.

8. The thermal runaway risk prediction method according to claim 2, wherein: The thermal runaway risk training data of the energy storage battery system is divided into a state sequence of multiple time steps, and the state sequence of each time step is converted into a vector form to obtain the state vector of each time step, including: Obtain at least two thermal runaway risk parameter samples for each battery pack of the energy storage battery system, and preprocess the thermal runaway risk parameter samples; Obtain a training data set from the preprocessed thermal runaway risk parameter samples; The training data of the training data set are divided into state sequences of multiple time steps, and the state sequence of each time step is converted into a vector form to obtain a state vector of each time step.

9. The thermal runaway risk prediction method according to claim 8, wherein: After inputting the representation of the root node of the tree into the initial prediction model to train at least two neural network layers in the initial prediction model to obtain the target prediction model, the method further includes: Obtaining a test sample set from the pre-processed thermal runaway risk parameter sample; Testing the target prediction model based on the test data in the test sample set to obtain a test result; The parameters of the target prediction model are adjusted according to the test results.

10. The thermal runaway risk prediction method according to any one of claims 1 to 9, wherein: The step of processing the thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model to obtain the thermal runaway probability of each battery pack includes: Use the target prediction model to predict the number of batteries at risk of thermal runaway in each battery pack of the energy storage battery system; The thermal runaway probability of each battery pack is calculated according to the number of batteries with thermal runaway risks in each battery pack.

11. The thermal runaway risk prediction method according to any one of claims 1 to 9, wherein: After predicting the total probability of thermal runaway risk of the energy storage battery system according to the thermal runaway probability of each battery pack, the method further includes: Determining a thermal runaway risk level of the energy storage battery system according to a total probability of thermal runaway risk of the energy storage battery system; Thermal runaway warning information is generated according to the thermal runaway risk level, and a thermal runaway warning is performed based on the thermal runaway warning information.

12. A thermal runaway risk prediction device, wherein: The thermal runaway risk prediction device comprises: A parameter acquisition module is used to obtain thermal runaway risk parameters of each battery pack in the energy storage battery system; a model processing module, configured to process the thermal runaway risk parameters layer by layer through at least two neural network layers in a target prediction model to obtain a thermal runaway probability for each battery pack, wherein the target prediction model is pre-trained based on state vectors for multiple time steps, and the state vectors are used to enhance the model's memory of past state sequences of the thermal runaway risk parameters; The risk prediction module is used to predict the total probability of thermal runaway risk of the energy storage battery system according to the thermal runaway probability of each battery pack.

13. A thermal runaway risk prediction device, wherein: The thermal runaway risk prediction device includes: a memory, a processor, and a thermal runaway risk prediction program stored in the memory and executable on the processor. When the thermal runaway risk prediction program is executed by the processor, the thermal runaway risk prediction method according to any one of claims 1 to 11 is implemented.

14. A storage medium, wherein: The storage medium stores a thermal runaway risk prediction program, which, when executed by a processor, implements the thermal runaway risk prediction method according to any one of claims 1 to 11.

Images