Battery pack charge state evaluation method and system, vehicle and storage medium

Abstract

The invention discloses a battery pack charge state evaluation method and system, a vehicle and a storage medium, and belongs to the technical field of batteries, and the battery pack charge state evaluation method comprises the steps: obtaining the current time sequence operation data and theoretical charging charge quantity of a target battery pack under the current charging condition; obtaining a target charge quantity difference evaluation model; the target charge quantity difference evaluation model is obtained by training a bidirectional time sequence neural network model based on an attention mechanism; determining the current charging charge quantity difference of the target battery pack according to the current time sequence operation data and a target charge quantity difference evaluation model; and determining the inconsistency of the charge state of the target battery pack according to the current charging charge quantity difference and the theoretical charging charge quantity. According to the method, the target charge quantity difference evaluation model is obtained through training of the bidirectional time sequence neural network model based on the attention mechanism, so that the charging charge quantity difference is quantitatively evaluated according to the target charge quantity difference evaluation model, and quantitative evaluation of the charge state inconsistency is realized.

Description

Technical Field

[0001] This application relates to the field of battery technology, specifically to battery pack state of charge assessment methods, systems, vehicles, and storage media. Background Technology

[0002] As the proportion of electrification in the passenger vehicle market increases, the safety issues of electric vehicles are gradually emerging and have become a key focus of industry attention. Among the many potential safety problems of electric vehicles, the power battery pack, as a core component, is particularly important for early warning of its failures. The inconsistency of the State of Charge (SOC) between individual cells within the battery pack is one of the key factors affecting the overall performance and safety of the battery pack. Therefore, it is essential to assess the SOC inconsistency of the battery pack.

[0003] However, the current technologies for assessing the SOC inconsistency of battery packs mostly rely on evaluating the static or dynamic voltage differences of individual cells to determine SOC consistency. This makes it difficult to achieve quantitative assessment, resulting in low accuracy. Summary of the Invention

[0004] This invention provides a method, system, vehicle, and storage medium for assessing the state of charge (SOC) of a battery pack, enabling quantitative assessment of SOC inconsistency and improving assessment accuracy.

[0005] Firstly, a method for assessing the state of charge (SOC) of a battery pack is provided, comprising the following steps: Obtain the current time-series operating data and theoretical charge capacity of the target battery pack under the current charging conditions; Obtain the target charge difference assessment model; the target charge difference assessment model is trained based on a bidirectional temporal neural network model with an attention mechanism; The current charge-to-capacity difference of the target battery pack is determined based on the current time-series operating data and the target charge-to-capacity difference assessment model. The inconsistency in determining the state of charge of the target battery pack is based on the difference between the current charging capacity and the theoretical charging capacity.

[0006] In some embodiments, a method for obtaining a target charge difference assessment model includes: Acquire historical time-series operational data of the training battery pack under historical charging conditions; A sample dataset was constructed based on historical time-series operational data; A bidirectional temporal neural network model based on an attention mechanism is trained using a sample dataset to obtain a target charge difference assessment model.

[0007] In some embodiments, the bidirectional temporal neural network model includes: a bidirectional gated recurrent unit encoding layer, an attention mechanism layer, a context vector generation layer, and a fully connected regression layer; the historical time-series running data includes historical running data at multiple time steps; A bidirectional temporal neural network model based on an attention mechanism is trained using a sample dataset to obtain a target charge difference assessment model, including: The target hiding state at each time step is determined based on the bidirectional gated cyclic unit encoding layer, the historical running data of each time step, and the preset time sequence. The training battery pack's current charge-to-capacity difference prediction is output based on the attention mechanism layer, target hidden state, context vector generation layer, and fully connected regression layer. The current training loss of the bidirectional temporal neural network model is determined based on the preset loss function, the predicted value of the current charging load difference, and the actual charge. The model then returns to the step of determining the target hidden state at each time step until the current training loss reaches the preset training loss, thus obtaining the target charge difference evaluation model.

[0008] In some embodiments, the bidirectional gated loop unit encoding layer includes a forward gated loop unit and a backward gated loop unit; the preset time sequence includes a first preset time sequence and a second preset time sequence; The target hiding state at each time step is determined based on the bidirectional gated cyclic unit encoding layer, historical running data at each time step, and a preset time sequence, including: The historical running data of each time step is processed by the forward gated loop unit according to the first preset time order to obtain the first hidden state of each time step along the first preset time order. The historical running data of each time step is processed by the backward gated loop unit according to the second preset time order to obtain the second hidden state of each time step along the second preset time order. Determine the target hidden state for each time step based on the first and second hidden states.

[0009] In some embodiments, the current charge-to-capacity difference prediction of the trained battery pack is output based on the attention mechanism layer, the target hidden state, the context vector generation layer, and the fully connected regression layer, including: The target attention weights for the target hidden state are determined based on the attention mechanism layer and the target hidden state. The context vector for the target dimension is determined based on the context vector generation layer, each target hidden state, and the target attention weights corresponding to each target hidden state. The current charge-to-charge difference is predicted based on the context vector of the fully connected regression layer and the target dimension.

[0010] In some embodiments, the fully connected regression layer includes a first fully connected sublayer, a second fully connected sublayer, and an output layer; Based on the context vector of the fully connected regression layer and the target dimension, the current charging load difference prediction value is output, including: The first fully connected sublayer is used to reduce the dimensionality of the context vector in the target dimension to obtain the context vector in the first dimension. The first-dimensional context vector is reduced in dimensionality by a second fully connected sublayer to obtain a second-dimensional context vector. The context vector of the second dimension is linearly transformed through the output layer to obtain the predicted value of the current charging load difference.

[0011] In some embodiments, the inconsistency of the state of charge of the target battery pack is determined based on the difference in current charging capacity and the theoretical charging capacity. The evaluation method further includes: The inconsistency of the state of charge of the target battery pack is determined by the ratio of the current difference in charge capacity to the theoretical charge capacity.

[0012] Secondly, this application also provides a battery pack state of charge assessment system, comprising: The first acquisition module is used to acquire the current timing operation data of the target battery pack under the current charging condition; The second acquisition module is used to acquire the theoretical charge capacity of the target battery pack under the current time-series operating data in the current charging condition. The third acquisition module is used to acquire the target charge difference assessment model; the target charge difference assessment model is trained based on a bidirectional temporal neural network model with an attention mechanism. The first determining module is used to determine the current charging capacity difference of the target battery pack based on the current time-series operating data and the target charge capacity difference evaluation model. The second determining module is used to determine the inconsistency of the state of charge of the target battery pack based on the difference between the current charging capacity and the theoretical charging capacity.

[0013] Thirdly, this application also provides a vehicle including a battery pack state of charge assessment system as described in the second aspect.

[0014] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to perform the steps of the method described in the first aspect.

[0015] Beneficial Effects: This application provides a method, system, vehicle, and storage medium for assessing the state of charge (SOC) of a battery pack. The method includes: acquiring current time-series operating data and theoretical charge capacity of a target battery pack under the current charging condition; acquiring a target charge capacity difference assessment model; the target charge capacity difference assessment model is trained using a bidirectional temporal neural network model based on an attention mechanism; determining the current charge capacity difference of the target battery pack based on the current time-series operating data and the target charge capacity difference assessment model; and determining the inconsistency of the SOC of the target battery pack based on the current charge capacity difference and the theoretical charge capacity. The battery pack SOC assessment method provided in this application obtains the target charge capacity difference assessment model through training a bidirectional temporal neural network model based on an attention mechanism. It then quantifies the current charge capacity difference of the target battery pack based on the target charge capacity difference assessment model and the current time-series operating data of the target battery pack, thereby achieving a quantitative assessment of the inconsistency of the target battery pack's SOC and improving the accuracy of the assessment. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a battery pack state of charge assessment method provided in the embodiments of this application; Figure 2 This is a schematic diagram of the overall framework of a bidirectional temporal neural network model based on an attention mechanism provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the bidirectional gated cyclic unit coding layer provided in the embodiments of this application; Figure 4 This is a schematic diagram of the overall process of a battery pack inconsistency evaluation method provided in the embodiments of this application; Figure 5 This is a schematic diagram of the principle structure of a battery pack state of charge assessment system provided in the embodiments of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] "A and / or B" includes the following three combinations: A only, B only, and a combination of A and B.

[0021] The use of "applies to" or "configured to" in this application implies open and inclusive language, which does not exclude the applicability to or configuration to devices performing additional tasks or steps. Additionally, the use of "based on" implies openness and inclusivity, because processes, steps, calculations, or other actions "based on" one or more of the stated conditions or values ​​may in practice be based on additional conditions or values ​​beyond those stated.

[0022] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0023] The applicant's research revealed that with the rapid development of the new energy vehicle industry, such as the increasing proportion of electrification, electric vehicle safety has gradually become a focus of industry attention. Among the numerous potential safety issues of electric vehicles, the power battery pack, as the core power source, is crucial for early warning of potential faults. Power batteries exhibit a wide variety of fault types, each of which, over a long period, can pose a significant threat to electric vehicle safety. Among these, the inconsistency in State of Charge (SOC) between individual cells within the battery pack is a key factor affecting the overall performance and safety of the battery pack.

[0024] To meet the range and power requirements of electric vehicles, power battery packs are typically composed of dozens to hundreds of individual cells connected in series and parallel. However, in actual use, due to differences in manufacturing processes, fluctuations in material ratios, and changes in the operating environment, initial inconsistencies in parameters such as capacity, internal resistance, and self-discharge rate are inevitable among individual cells. State of Charge (SOC) inconsistency has a cumulative characteristic; after dozens or even hundreds of charge-discharge cycles, the initial inconsistency gradually accumulates and amplifies, severely impacting the safety and reliability of the battery pack. This initial inconsistency, accumulated over multiple charge-discharge cycles, is further amplified, leading to significant differences in the SOC of individual cells within the battery pack. In actual operation, the usable capacity of the battery pack is limited by the cell with the lowest SOC. Therefore, SOC inconsistency will lead to capacity loss, accelerate overall battery aging, shorten lifespan, exacerbate overcharging and over-discharging of some cells, and even trigger thermal runaway in extreme cases.

[0025] Currently, the industry mainly uses detection methods based on voltage / temperature parameter consistency, which assess SOC consistency by evaluating the static or dynamic voltage differences of individual cells. This type of method for detecting SOC inconsistency typically acquires basic battery parameters from sensors deployed in the battery management system. After preprocessing these parameters, representative features characterizing the battery's external properties are extracted to form the required feature set. Subsequently, unsupervised learning clustering methods, outlier detection, or statistical methods such as Euclidean distance and Pearson correlation coefficient are used to evaluate the correlation between the features of each cell, and further, relevant result evaluation formulas are developed to determine the inconsistency level of the battery pack. The technical limitation of this method lies in the insufficient accuracy and sensitivity of voltage sensor acquisition, especially for lithium iron phosphate batteries. Due to the flat open circuit voltage-state of charge (OCV-SOC) plateau, small SOC differences correspond to extremely small voltage differences.

[0026] In recent years, with the rapid iteration of computer technology, big data technology has gradually been applied to battery management systems. Related technical solutions generally combine feature engineering with traditional machine learning, using extracted statistical features such as voltage and temperature to directly predict battery states such as voltage, SOC, and capacity, thereby further evaluating SOC inconsistencies in the battery pack. However, these methods often focus on short-cycle data windows, relying heavily on insufficient modeling in the long term, and simple neural networks struggle to capture the evolution of inconsistencies across charging cycles.

[0027] For example, the technology for assessing the SOC inconsistency of power battery packs typically involves acquiring relevant basic battery parameters through a data acquisition circuit. Then, methods such as clustering, outlier analysis, or correlation are used to identify abnormal battery cells. Finally, relevant evaluation methods are used to assess the degree of SOC inconsistency of the battery pack. These methods are almost impossible to directly and effectively quantify the degree of SOC inconsistency; they can only demonstrate the correlation between individual battery cells within the pack.

[0028] For example, some SOC inconsistency detection methods, due to the nature of unsupervised learning algorithms or the characteristics of lithium batteries, often suffer from insufficient or underestimated sensitivity in detecting actual inconsistencies, leading to potential risks. Lithium iron phosphate batteries have a very flat OCV-SOC curve plateau region; even relatively significant SOC differences can be further masked by limited voltage sampling accuracy and noise. Furthermore, clustering algorithms, due to inherent limitations, cannot accurately control the clustering accuracy of each data point, resulting in poor adaptability.

[0029] For example, methods based on direct estimation of SOC and capacity are generally implemented through models or machine learning. However, these methods focus on short-cycle data windows, relying heavily on insufficient modeling in the long run. Simple neural networks struggle to capture the inconsistent evolution across charging cycles. The characteristics of the lithium iron phosphate platform region also pose a significant challenge to accurate SOC estimation within the industry. The SOC estimation errors for each individual cell are independent and random, and these errors are further compounded during the calculation of differences. Furthermore, traditional statistical features used in machine learning, such as mean and variance, are difficult to effectively represent battery time-series data. The dynamic time-series patterns of the charging process are over-quantified by statistics, making them insensitive to subtle trends.

[0030] It is evident that assessing the SOC inconsistency among power battery packs is extremely difficult due to the limitations of the accuracy of battery SOC estimation. Directly estimating the SOC of individual battery cells is subject to limitations in the algorithm itself and interference from noise and environmental factors, resulting in initial errors that can lead to errors in quantifying SOC inconsistency. This is particularly true for lithium iron phosphate batteries, where accurate estimation of individual cell SOC is a significant challenge due to voltage plateau effects. Furthermore, other methods that assess basic battery parameters are even less effective at directly quantifying the SOC inconsistency of the entire battery pack.

[0031] Moreover, the relevant technologies lack a true and effective quantification of SOC inconsistency. Under slight SOC differences, the detection accuracy and sensitivity are limited by the voltage plateau effect, and the time series statistical features are difficult to represent the short-term abnormal data characteristics of the battery.

[0032] In view of this, embodiments of this application provide a method, system, vehicle, and storage medium for assessing the state of charge (SOC) of a battery pack. Embodiments of this application train a target SOC difference assessment model using a bidirectional temporal neural network model based on an attention mechanism. The model is then used to quantitatively assess the current charging SOC difference of the target battery pack based on the target SOC difference assessment model and the current temporal operating data of the target battery pack. This achieves a quantitative assessment of the inconsistency of the SOC of the target battery pack, thereby improving the accuracy of the assessment.

[0033] It should be noted that the battery pack provided in this application embodiment is a power battery pack, etc.

[0034] Figure 1 This is a flowchart illustrating a battery pack state-of-charge (SOC) assessment method provided in this embodiment. On one hand, this embodiment provides a battery pack SOC assessment method applicable to vehicle control systems or battery management systems, enabling the quantitative assessment of SOC inconsistencies in battery packs. This method can be executed by a battery pack SOC assessment system, which can be implemented in software and / or hardware and can be configured in the processor or controller of the vehicle control system or battery management system. Please refer to... Figure 1 The method includes the following steps: Step 110: Obtain the current timing operation data and theoretical charge capacity of the target battery pack under the current charging conditions.

[0035] The current charging conditions include fast charging, slow charging, variable current charging, constant current charging, constant voltage charging, etc., as long as it is a charging condition. For example, this application uses slow charging, the most common and longest-lasting condition, as an example for illustration, and will not be elaborated further below.

[0036] The current time-series operational data includes data from multiple dimensions, such as the voltage of each individual cell in the power battery pack. Average cell voltage of power battery pack Total current of power battery pack Real-time input ampere-hours during the current charging process Average temperature of power battery pack .

[0037] The theoretical charge capacity refers to the ideal charge capacity. For example, in this embodiment, considering practical applications, the average charge capacity of the battery pack is taken as the ideal charge capacity.

[0038] Step 120: Obtain the target charge difference assessment model.

[0039] The target charge difference assessment model is trained based on a bidirectional temporal neural network model with an attention mechanism.

[0040] In some embodiments, the method for obtaining a target charge difference assessment model includes: obtaining historical time-series operation data of the training battery pack under historical charging conditions; constructing a sample dataset based on the historical time-series operation data; and training a bidirectional time-series neural network model based on an attention mechanism based on the sample dataset to obtain the target charge difference assessment model.

[0041] The historical charging conditions include fast charging, slow charging, variable current charging, constant current charging, constant voltage charging, etc. Any charging condition is acceptable.

[0042] The training battery pack consists of battery packs from multiple vehicles. For example, the historical time-series operational data of the training battery pack under historical charging conditions includes: collected time-series operational data of all charging conditions across the entire lifecycle of the battery packs of multiple vehicles. Furthermore, the collected historical time-series operational data undergoes data preprocessing. The preprocessing methods include: cleaning and normalizing the data, removing rows containing null values ​​and outlier removal, and then normalizing the data in each dimension column to a standard normal distribution.

[0043] The historical time-series operational data includes data from multiple dimensions, such as the voltage of each individual cell in the power battery pack. Average cell voltage of power battery pack Total current of power battery pack Real-time input ampere-hours during each charging process Average temperature of power battery pack .

[0044] The process of constructing a sample dataset based on historical time-series operational data includes: dividing the time-series operational data for each vehicle's charging cycle into segments, and constructing a sample dataset based on these segments to provide training data for the bidirectional time-series neural network model. Specifically, constructing the sample dataset includes the following steps: Step 1: Utilize the time-series data from each vehicle's charging cycle to construct a multi-time-scale feature system reflecting battery health, based on different time scales. This multi-time-scale feature system specifically includes basic parameter features, derived features, and rolling time-series features.

[0045] Among them, the basic parameter characteristics are directly characterized using data acquired by the acquisition circuit, including the voltage of each cell in the power battery pack. Average cell voltage of power battery pack Total current of power battery pack and the average temperature of the power battery pack The acquisition circuit refers to the hardware sensors and signal conditioning system used to acquire key vehicle operating parameters in real time. Specifically, it includes: a current sensor, a voltage divider sampling network, an NTC temperature probe, a signal conditioning module, and an ADC module.

[0046] Among them, derived features are derived from changes in basic parameter characteristics and are used to describe the state of individual battery cells, including the difference between individual cell voltage and average individual cell voltage. Individual voltage change rate and average single-cell voltage change rate .

[0047] Among them, the rolling time series feature is used to describe the state changes of the battery over time, including the voltage difference of a single cell at five consecutive time steps. (i.e., the voltage of a single cell at time step (t+5) minus the voltage of a single cell at time step (t)) and the voltage difference of the average single cell voltage over the preceding and following 5 time steps. (i.e., the average individual cell voltage at time step (t+5) minus the average individual cell voltage at time step (t), and the rolling average voltage value of the individual cell voltage over the preceding and following 5 time steps.) (i.e., the average value of the individual cell voltage over the five time steps t, t+1, t+2, t+3, t+4, and t+5) and the rolling average voltage value of the individual cell voltage over the five consecutive time steps. (That is, the average value of the average single-cell voltage over the five time steps t, t+1, t+2, t+3, t+4, and t+5). The specific calculation formula is as follows: ; ; ; ; ; ; ; in, This represents the voltage of a single cell at time step (t+5). This represents the voltage of a single cell at time step (t). This represents the average single-cell voltage at time step (t+5); This represents the average single-cell voltage at time step (t).

[0048] Step 2: Based on the constructed multi-timescale feature system, extract all time-series operational data for each vehicle under each charging condition, and form a corresponding sample dataset. For example, randomly shuffle the sample dataset and divide it into a training set and a test set in an 8:2 ratio. The training set is used as input to the bidirectional temporal neural network model, while the test set is used to test the prediction performance of the bidirectional temporal neural network model.

[0049] In some embodiments, after constructing the sample dataset, the standardization process of the sample dataset is also included, namely, using mask sequence packing technology to process variable-length sample data.

[0050] Specifically, data processing of the sample dataset is a crucial step before training the bidirectional temporal neural network model. This data processing mainly includes standardizing and managing the time-series data of the training battery pack. Standardization ensures that each feature has equal importance in the training of the bidirectional temporal neural network model and scales feature values ​​to a similar numerical range, avoiding convergence problems caused by differences in feature scale during gradient updates and improving the stability of the bidirectional temporal neural network model.

[0051] For example, the Z-score standardization method is used to standardize the sample data in the sample dataset. The Z-score standardization method can be expressed as: ; in, This represents the standardized result. Represents the corresponding eigenvalues. The average value of the feature. The standard deviation of a feature.

[0052] Through this transformation, all features are converted into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0053] However, due to the inherent randomness of electric vehicle charging processes—specifically, the varying lengths of charging time-series data for different samples—bidirectional temporal neural network models are difficult to train directly. Related techniques typically pad the longest sequence with zeros. However, this wastes computational resources, and the padding interferes with the bidirectional temporal neural network model's learning of the real sequence, leading to a performance degradation.

[0054] In this embodiment, a 95th quantile-based adaptive padding sequence is used to truncate only 5% of the excessively long sequences, preserving the complete information of the vast majority of samples. Furthermore, a masking technique is used to process the time-series data, explicitly identifying the validity of each time step; for example, a value of 1 represents real data, and a value of 0 represents a padding position. The mask allows the bidirectional temporal neural network model to ignore the padding portion. In subsequent calculations, the mask ensures that only valid time steps are computed, avoiding wasted resources on padding positions. Simultaneously, in the subsequent attention weight calculation stage, the mask sets the attention score for padding positions to negative numbers, ensuring that the bidirectional temporal neural network model only focuses on the valid sequence portion, thus enabling the model to more fully learn the valid information from the sample data.

[0055] Specifically, in this embodiment, a bidirectional temporal neural network model based on an attention mechanism is constructed, specifically designed to learn and predict charging charge differences (i.e., charging Ah differences) from historical time-series operational data of the training battery pack under historical charging conditions. This model employs a multi-layered deep architecture, effectively capturing complex temporal patterns during the charging process. This facilitates subsequent training to obtain the optimal target charging charge difference assessment model, thereby improving the accuracy of charging Ah difference prediction.

[0056] In some embodiments, the bidirectional temporal neural network model includes: a bidirectional gated recurrent unit encoding layer, an attention mechanism layer, a context vector generation layer, and a fully connected regression layer; the historical time-series running data includes historical running data at multiple time steps.

[0057] Figure 2 This is a schematic diagram of the overall framework of a bidirectional temporal neural network model based on an attention mechanism provided in an embodiment of this application. For example, see [link to relevant documentation]. Figure 2The bidirectional temporal neural network model includes: an input layer, a bidirectional gated recurrent unit (GRU) encoding layer, an attention mechanism layer, a context vector generation layer, and a fully connected regression layer (including a fully connected regressor and an output layer). Among them, the bidirectional GRU encoding layer has a bidirectional structure.

[0058] The target charge difference assessment model is obtained by training a bidirectional temporal neural network model based on an attention mechanism using a sample dataset, including the following steps: Step 1: Determine the target hiding state at each time step based on the bidirectional gated loop unit encoding layer, historical running data of each time step, and preset time sequence.

[0059] In some embodiments, the bidirectional gated loop unit encoding layer includes a forward gated loop unit and a backward gated loop unit; the preset time sequence includes a first preset time sequence and a second preset time sequence; determining the target hidden state of each time step based on the bidirectional gated loop unit encoding layer, the historical running data of each time step, and the preset time sequence includes: processing the historical running data of each time step according to the first preset time sequence through the forward gated loop unit to obtain a first hidden state of each time step along the first preset time sequence; processing the historical running data of each time step according to the second preset time sequence through the backward gated loop unit to obtain a second hidden state of each time step along the second preset time sequence; and determining the target hidden state of each time step based on the first hidden state and the second hidden state of each time step.

[0060] The first preset time sequence and the second preset time sequence are opposite time sequences. The first preset time sequence follows the chronological order, for example, from the past to the future (i.e., forward). The second preset time sequence follows the reverse chronological order, for example, from the future to the past (i.e., backward).

[0061] The bidirectional gated recurrent unit (GRU) encoding layer is composed of a special type of recurrent neural network that can capture forward and backward information in time series data. Unlike the traditional unidirectional GRU, which can only process sequences from front to back, the bidirectional GRU processes sequences from two directions simultaneously: one from past to future (forward) and the other from future to past (backward). In this way, for any element in the sequence, the bidirectional GRU can utilize all the information before and after it.

[0062] The bidirectional gated cyclic unit encoding layer receives time-series execution data, for example, a certain time series data. It serves as the input to the bidirectional gated recurrent unit (ROU) encoding layer. Here, T is the sequence length. It consists of two hidden layers, each with 128 hidden units, enabling it to learn more complex operating conditions.

[0063] Figure 3 This is a schematic diagram of the structure of the bidirectional gated cyclic unit coding layer provided in the embodiments of this application. For example, see [link to relevant documentation]. Figure 3 The bidirectional gated recurrent unit (GRU) encoding layer includes a forward GRU and a backward GRU. The forward GRU processes the input time-series running data sequentially in time, while the backward GRU processes the input time-series running data in reverse time, and finally outputs the hidden state of each time step.

[0064] Specifically, the process of determining the target hidden state at each time step through the bidirectional gated loop unit encoding layer, historical running data at each time step, and preset time order is as follows: the forward gated loop unit processes the time series data according to the first preset time order. To capture the historical dependencies of the sequence data, i.e., from arrive At each time step t, not only the current input is considered. It will also combine the hidden state generated in the previous time step. That is, based on the current input The hidden state generated at the previous time step To update the current hidden state Therefore, through the processing of the forward-gated loop unit, the first hidden state along the first preset time sequence can be obtained for each time step. Simultaneously, the backward-gated loop unit processes the time series data according to the second preset time sequence. To predict the future trend information of the sequence data, that is, from arrive Similarly, at each time step t, not only the current input is considered. It will also combine the hidden state generated in the previous time step. That is, based on the current input The hidden state generated at the previous time step To update the current hidden state Therefore, through the processing of the backward-gated loop unit, the second hidden state along the second preset time sequence can be obtained for each time step. Then, based on the first and second hidden states of each time step, the target hidden state of each time step can be obtained, for example, [ , Meanwhile, the bidirectional gated recurrent unit encoding layer adopts a two-layer stacked structure, which enhances feature representation ability through hierarchical feature extraction, and the gating mechanism effectively alleviates the gradient vanishing problem. In the layer structure, a random deactivation technique with a probability of 0.2 is used between layers to prevent overfitting.

[0065] Step 2: Output the predicted current charge / charge difference of the training battery pack based on the attention mechanism layer, target hidden state, context vector generation layer, and fully connected regression layer.

[0066] The attention mechanism layer is used for dimension adaptation and dropout regularization; the context vector generation layer is used to set the weighted feature fusion, batch size, and hidden layer dimensions; and the fully connected regression layer is used to set the hidden layer dimensions, activation function, and dropout regularization.

[0067] In some embodiments, the current charge-to-power difference prediction value of the training battery pack is output based on the attention mechanism layer, the target hidden state, the context vector generation layer, and the fully connected regression layer, including: determining the target attention weight of the target hidden state based on the attention mechanism layer and the target hidden state; determining the context vector of the target dimension based on the context vector generation layer, each target hidden state, and the target attention weight corresponding to each target hidden state; and outputting the current charge-to-power difference prediction value based on the fully connected regression layer and the context vector of the target dimension.

[0068] Each time step includes corresponding runtime data, and the feature dimensions of each time step are fixed, such as including basic parameter features, derived features, and rolling features. The context vector for the target dimension is a vector of fixed dimensions.

[0069] Specifically, the attention mechanism layer receives the complete hidden state sequence (i.e., the target hidden state at each time step) generated by the bidirectional gated recurrent unit encoding layer at all time steps, combining forward and backward information. It then calculates the importance weights of the features at each time step using a multilayer perceptron. This attention mechanism layer first maps the features to the attention dimension through a linear transformation, then introduces non-linearity using a hyperbolic tangent activation function, followed by regularization through a random deactivation layer. Finally, it generates normalized attention weights through a linear layer and a normalized exponential function along the time dimension. This yields the target attention weights corresponding to the target hidden state at each time step.

[0070] After obtaining the target attention weights corresponding to each target hidden state, the context vector generation layer performs a weighted summation of the target hidden states at each time step output by the bidirectional GRU encoding layer based on each target attention weight, generating a target-dimensional context vector. This target-dimensional context vector integrates the effective information of the entire sequence, providing a global feature representation for subsequent regression prediction. For example, the calculation formula for the target-dimensional context vector is: Target-dimensional context vector = Target attention weight 1 × Target hidden state at time step 1 + Target attention weight 2 × Target hidden state at time step 2 + ... + Target attention weight T × Target hidden state at time step t.

[0071] Finally, the fully connected regression layer progressively compresses the dimension of the context vector of the target dimension and outputs the predicted value of the current charging load difference.

[0072] It should be noted that during the calculation of attention weights for each target, the attention weights at the filling positions are set to negative infinity using a mask sequence packing technique to ensure that the bidirectional temporal neural network model only focuses on the effective time steps.

[0073] In some embodiments, the fully connected regression layer includes a first fully connected sublayer, a second fully connected sublayer, and an output layer; outputting the current charging load difference prediction value based on the fully connected regression layer and the target dimension context vector includes: performing dimensionality reduction processing on the target dimension context vector through the first fully connected sublayer to obtain a first dimension context vector; performing dimensionality reduction processing on the first dimension context vector through the second fully connected sublayer to obtain a second dimension context vector; and performing a linear transformation on the second dimension context vector through the output layer to obtain the current charging load difference prediction value.

[0074] The fully connected regression layer consists of three fully connected sublayers: a first fully connected sublayer, a second fully connected sublayer, and an output layer. These sublayers progressively compress the dimensionality of the target-dimensional context vector. The first fully connected sublayer performs a first dimensionality reduction on the target-dimensional context vector, resulting in a first-dimensional context vector. For example, it reduces the dimension of the target-dimensional context vector to 64 dimensions. The second fully connected sublayer performs a second dimensionality reduction on the first-dimensional context vector, resulting in a second-dimensional context vector, for example, further reducing it to 32 dimensions. Finally, the output layer performs a linear transformation to generate a single-valued prediction result for the charging Ah difference. The operation performed by the output layer on the vector after the second dimensionality reduction is a linear transformation (e.g., weighted summation + bias), compressing the vector from d dimensions (e.g., 32 dimensions) to a 1-dimensional scalar, which serves as the final regression prediction result.

[0075] Each fully connected sublayer is activated by a linear rectified function, and a random deactivation technique with a probability of 0.2 is introduced between layers.

[0076] Step 3: Determine the current training loss of the bidirectional temporal neural network model based on the preset loss function, the predicted value of the current charging load difference, and the actual load. Then, return to execute the step of determining the target hidden state at each time step until the current training loss reaches the preset training loss, and obtain the target load difference evaluation model.

[0077] The bidirectional temporal neural network model training process employs a length-based sequence packing technique. First, the sample data is sorted in descending order of actual sequence length. Then, the sorted sequences are packed together, and calculations are performed only on valid time steps to avoid wasting computational resources on filling positions. The packed sequences are input into a bidirectional gated recurrent unit for forward propagation, and the output results are then unpacked to restore the original batch order. To prevent gradient explosion, gradient pruning is used during training to limit the gradient norm within a preset threshold. This preset threshold is the upper limit of the L2 norm (the square root of the sum of the squares of all gradient elements). When the calculated L2 norm exceeds this threshold, the system scales the gradient proportionally to make its L2 norm equal to the threshold. For example, this preset threshold is 2; the specific value can be set according to actual conditions and is not specifically limited here.

[0078] The default loss function is the mean squared error (MSE) loss.

[0079] Specifically, based on the constructed feature sample data (i.e., sample dataset), the training loss of the bidirectional temporal neural network model is calculated using the mean squared error (MSE) value. The squared difference between the predicted charging Ah value and the actual Ah value is measured to calculate the current training loss. The process is then repeated until the training loss of the bidirectional temporal neural network model reaches the predetermined training loss, resulting in a completed bidirectional temporal neural network model (i.e., target charge difference assessment model).

[0080] During the testing phase of the bidirectional temporal neural network model, reserved test set data was used. Evaluation metrics included root mean square error (MSE), mean absolute error (MAE), coefficient of determination, and mean absolute percentage error. A k-fold cross-validation strategy was employed, dividing the dataset into k subsets. Training was performed using k-1 subsets in rotation, and testing was conducted using one subset, ensuring the stability of the evaluation results.

[0081] It should be noted that the core innovation of the technical approach proposed in this application lies in the overall methodology of predicting Ah differences by analyzing time-series data of the battery's full charging process and evaluating SOC inconsistencies accordingly, including a specific feature engineering system and data processing flow. In addition to the gated recurrent unit (GRU) used in this application embodiment, the specific neural network model for implementing the time-series feature extraction function may also include at least one of the following: Recurrent Neural Network (RNN), Long Short-Term Memory Neural Network (LSTM), and Transformer encoder based on a self-attention mechanism.

[0082] Among them, RNN, as a basic sequence model, processes data step by step through recurrent connections, but it has the limitation of difficulty in learning long-range dependencies. LSTM, by introducing a sophisticated gating mechanism, can more effectively capture long-term dynamic correlations in the charging process. Transformer, on the other hand, relies entirely on a self-attention mechanism, which can compute in parallel and globally perceive the interrelationships between all time steps, but its implementation requires the additional introduction of position encoding to preserve temporal information.

[0083] Although these models differ in their internal structure and timing information processing mechanisms, they all follow the same input-output specifications and data processing logic in achieving the purpose of this application. Specifically, they receive battery charging timing data that has undergone the same preprocessing and feature engineering, and are all compatible with masking techniques to handle variable-length sequences, ultimately outputting Ah difference prediction values. Therefore, the core of this application's protection lies in its overall technical approach and data processing method, rather than being specific to any particular model implementation. This demonstrates the strong versatility and wide applicability of this technical solution.

[0084] Step 130: Determine the current charging capacity difference of the target battery pack based on the current time-series operating data and the target charge capacity difference assessment model.

[0085] Specifically, the system collects and preprocesses the time-series operational data (i.e., current time-series operational data) for each vehicle during its current charging process. Then, it inputs this data into a trained time-series bidirectional neural network model (i.e., the target charge-to-weight ratio difference assessment model). This model automatically extracts feature vectors from the entire charging process and outputs the predicted charging Ah difference (i.e., the current charging charge-to-weight ratio difference). This difference value reflects the deviation between the actual charging capacity and the ideal charging capacity.

[0086] Step 140: Determine the inconsistency of the state of charge of the target battery pack based on the difference between the current charging capacity and the theoretical charging capacity.

[0087] The theoretical charge capacity is the ideal charge capacity. For example, in this embodiment of the application, considering practical applications, the average charge capacity of the battery pack is taken as the ideal charge capacity.

[0088] In some embodiments, determining the inconsistency of the state of charge (SOC) of the target battery pack based on the difference in current charge capacity and the theoretical charge capacity includes: The inconsistency of the state of charge of the target battery pack is determined by the ratio of the current difference in charge capacity to the theoretical charge capacity.

[0089] Based on the predicted differences in charging Ah, the SOC inconsistency is calculated using the following formula. : ; in, This represents the difference in charging Ah predicted by the target charge difference assessment model. This indicates the ideal charge capacity (i.e., the theoretical charge capacity).

[0090] It is understood that the embodiments of this application obtain a target charge difference assessment model by training a bidirectional temporal neural network model based on an attention mechanism, so as to quantitatively assess the current charging charge difference of the target battery pack based on the target charge difference assessment model and the current temporal operation data of the target battery pack, thereby realizing the quantitative assessment of the inconsistency of the state of charge of the target battery pack and improving the accuracy of the assessment.

[0091] In some embodiments, the battery pack state of charge assessment method further includes: establishing a multi-level inconsistency assessment system based on the inconsistency of the state of charge of the target battery pack; the multi-level inconsistency assessment system is used to quantify the inconsistency of the state of charge of the target battery pack.

[0092] Specifically, based on the magnitude of the SOC differences, a multi-level inconsistency assessment system can be established, as follows: Slight inconsistencies: SOC < 3% indicates that the battery pack has good consistency and can be used normally, and should be monitored regularly.

[0093] Moderate inconsistency: 3% ≤ If the SOC is less than 8%, it indicates a significant inconsistency in the battery pack, requiring closer monitoring or manual intervention.

[0094] Serious inconsistency: If the SOC is ≥8%, it indicates a danger signal for the battery pack, requiring immediate professional inspection or replacement.

[0095] Based on the above multi-level inconsistency assessment system, the SOC inconsistency of individual battery cells can be accurately quantified.

[0096] In some embodiments, the battery pack state of charge assessment method further includes: constructing a model of the evolution trend of inconsistent state of charge.

[0097] Specifically, based on the evaluation results of multiple charging processes, a SOC inconsistency evolution trend model is established, as follows: Stable trend: Small range of inconsistency fluctuations indicates that the battery pack is in a stable state.

[0098] Worsening trend: The increasing inconsistency indicates a need for timely intervention.

[0099] Improvement trend: The inconsistency is gradually decreasing, indicating that the battery cell balancing or recharging measures are effective.

[0100] Figure 4 This is a schematic diagram of the overall process of a battery pack inconsistency evaluation method provided in the embodiments of this application. For example, see [link to relevant documentation]. Figure 4 The overall process of the battery pack inconsistency assessment method is as follows: Step S101, collect all charging time-series operation data of the vehicle throughout its entire life cycle and perform data preprocessing. Step S102, feature engineering construction, divide the time-series operation data of each vehicle for each charge into segments, and train the model based on the segmented charging time-series operation data. Step S103, standardize the battery sample dataset and process variable-length sample data using mask sequence packaging technology. Step S104, construct a bidirectional time-series neural network model, input the processed sample dataset into the model for training, and test the model reliability. Step S105, predict the difference in vehicle charging Ah based on the trained model, and assess SOC inconsistency based on the magnitude of the difference in charging Ah. It can be seen that this application, by analyzing the time-series characteristics of the entire charging process, can detect the development trend of inconsistency in advance, providing a decision-making basis for the balanced maintenance of the battery pack. Furthermore, this application can perform early identification and quantitative assessment of SOC inconsistency in power battery packs, improve the accuracy of SOC inconsistency assessment and early warning capability, and provide technical support for preventive maintenance of battery packs.

[0101] As can be seen, the battery pack state-of-charge (SOC) assessment method provided in this application proposes a new paradigm for indirect SOC inconsistency assessment based on differences in charging Ah, breaking through the technical bottleneck of traditional direct SOC estimation. It establishes a technical route of "charging behavior analysis - Ah difference prediction - SOC inconsistency assessment," circumventing the SOC estimation problem in the smooth voltage plateau region of lithium iron phosphate batteries. Furthermore, it introduces time-series pattern recognition technology into the field of battery inconsistency assessment, realizing a shift from static parameter comparison to dynamic process analysis. Moreover, it designs a multi-timescale time-series feature engineering method, including forward rolling features and backward rolling features, effectively capturing the dynamic characteristics of the charging process. A deep learning architecture integrating a bidirectional GRU encoding layer and an attention mechanism is constructed, simultaneously considering long-term dependency modeling and focusing on key time-point information. Masked sequence packaging processing technology is used to achieve efficient processing of variable-length charging sequences and optimization of computational resources.

[0102] The battery pack state of charge assessment method provided in this application is a SOC inconsistency assessment method based on Ah quantity difference. It assesses the SOC inconsistency of the battery pack by predicting the difference between the actual and ideal charging capacity during the charging process. It is an implementation path that transforms the assessment of Ah quantity difference into SOC difference and establishes a multi-level inconsistency assessment system.

[0103] This application also constructs a temporal feature system and a bidirectional neural network model architecture. The multi-timescale temporal feature system designed for the battery charging process can comprehensively describe the dynamic characteristics of the battery charging process. The designed neural network model uses a bidirectional GRU model and is combined with an attention mechanism, allowing it to pay more attention to the effective information in the charging process. For charging process data with variable-length sequences, masking technology is used for processing.

[0104] In summary, the battery pack state-of-charge assessment method provided in this application has the following advantages: First, this application addresses the difficulty in accurately estimating the State of Charge (SOC) of power batteries in the voltage plateau region, solving the industry problem of insensitivity in traditional methods. Especially for lithium iron phosphate batteries, whose open-circuit voltage curves are very flat, direct SOC estimation has always been a challenge, leading to instability in assessing and warning of battery pack SOC consistency. Furthermore, compared to direct voltage evaluation, its accuracy is significantly affected by data acquisition precision and environmental factors such as noise. Therefore, this invention bypasses accurate SOC estimation, using charge quantity to assess battery pack SOC consistency and accurately quantifies this consistency difference.

[0105] Secondly, this application fully considers the temporal information of the data, capturing its dynamic characteristics across multiple time scales to describe abnormal battery behavior. Related methods, based on fragmented data or statistical features, do not fully utilize information. This application employs a bidirectional GRU, which can simultaneously consider historical and future information, and its attention mechanism automatically focuses on key time points. Masking techniques effectively handle variable-length sequences, avoiding wasted computational resources and enhancing the model's generalization ability.

[0106] Figure 5 This is a schematic diagram of the principle structure of a battery pack state-of-charge assessment system provided in the embodiments of this application. On the other hand, the embodiments of this application provide a battery pack state-of-charge assessment system, see below. Figure 5 The battery pack state of charge assessment system 100 includes: a first acquisition module 10, used to acquire the current time-series operating data of the target battery pack under the current charging condition; a second acquisition module 20, used to acquire the theoretical charge capacity of the target battery pack under the current time-series operating data under the current charging condition; a third acquisition module 30, used to acquire the target charge capacity difference assessment model; the target charge capacity difference assessment model is trained based on a bidirectional time-series neural network model with an attention mechanism; a first determination module 40, used to determine the current charge capacity difference of the target battery pack based on the current time-series operating data and the target charge capacity difference assessment model; and a second determination module 50, used to determine the inconsistency of the state of charge of the target battery pack based on the current charge capacity difference and the theoretical charge capacity.

[0107] The technical solution of this application provides a battery pack state of charge assessment system. It obtains a target charge difference assessment model by training a bidirectional temporal neural network model based on an attention mechanism. The system then quantifies the current charging charge difference of the target battery pack based on the target charge difference assessment model and the current temporal operating data of the target battery pack, thereby achieving a quantitative assessment of the inconsistency of the target battery pack's state of charge and improving the accuracy of the assessment.

[0108] In some embodiments, the third acquisition module 30 is further configured to: Acquire historical time-series operational data of the training battery pack under historical charging conditions; A sample dataset was constructed based on historical time-series operational data; A bidirectional temporal neural network model based on an attention mechanism is trained using a sample dataset to obtain a target charge difference assessment model.

[0109] In some embodiments, the bidirectional temporal neural network model includes: a bidirectional gated recurrent unit encoding layer, an attention mechanism layer, a context vector generation layer, and a fully connected regression layer; the historical time-series running data includes historical running data at multiple time steps; The third acquisition module 30 is also used for: The target hiding state at each time step is determined based on the bidirectional gated cyclic unit encoding layer, the historical running data of each time step, and the preset time sequence. The training battery pack's current charge-to-capacity difference prediction is output based on the attention mechanism layer, target hidden state, context vector generation layer, and fully connected regression layer. The current training loss of the bidirectional temporal neural network model is determined based on the preset loss function, the predicted value of the current charging load difference, and the actual charge. The model then returns to the step of determining the target hidden state at each time step until the current training loss reaches the preset training loss, thus obtaining the target charge difference evaluation model.

[0110] In some embodiments, the bidirectional gated loop unit encoding layer includes a forward gated loop unit and a backward gated loop unit; the preset time sequence includes a first preset time sequence and a second preset time sequence; The third acquisition module 30 is also used for: The historical running data of each time step is processed by the forward gated loop unit according to the first preset time order to obtain the first hidden state of each time step along the first preset time order. The historical running data of each time step is processed by the backward gated loop unit according to the second preset time order to obtain the second hidden state of each time step along the second preset time order. Determine the target hidden state for each time step based on the first and second hidden states.

[0111] In some embodiments, the third acquisition module 30 is further configured to: The target attention weights for the target hidden state are determined based on the attention mechanism layer and the target hidden state. The context vector for the target dimension is determined based on the context vector generation layer, each target hidden state, and the target attention weights corresponding to each target hidden state. The current charge-to-charge difference is predicted based on the context vector of the fully connected regression layer and the target dimension.

[0112] In some embodiments, the fully connected regression layer includes a first fully connected sublayer, a second fully connected sublayer, and an output layer; The third acquisition module 30 is also used for: The first fully connected sublayer performs the first dimensionality reduction on the context vector of the target dimension to obtain the context vector of the first dimension. The first-dimensional context vector is reduced in dimensionality by a second fully connected sublayer to obtain the second-dimensional context vector. The context vector of the second dimension is linearly transformed through the output layer to obtain the predicted value of the current charging load difference.

[0113] In some embodiments, the second determining module 50 is further configured to: The inconsistency of the state of charge of the target battery pack is determined by the ratio of the current difference in charge capacity to the theoretical charge capacity.

[0114] This embodiment also provides a vehicle that includes a battery pack state of charge assessment system as described in any embodiment of this application.

[0115] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to perform the steps of any of the methods in the above embodiments.

[0116] In the embodiments of this application, the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0117] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0118] The foregoing has provided a detailed description of a battery pack state-of-charge assessment method, system, vehicle, and storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for assessing the state of charge of a battery pack, characterized in that, Includes the following steps: Obtain the current time-series operating data and theoretical charge capacity of the target battery pack under the current charging conditions; Obtain the target charge difference assessment model; The target charge difference assessment model is trained based on a bidirectional temporal neural network model with an attention mechanism; The current charge-to-capacity difference of the target battery pack is determined based on the current time-series operating data and the target charge-to-capacity difference assessment model. The inconsistency in the state of charge of the target battery pack is determined based on the difference in the current charging capacity and the theoretical charging capacity.

2. The method according to claim 1, characterized in that, The method for obtaining the target charge difference assessment model includes: Acquire historical time-series operational data of the training battery pack under historical charging conditions; A sample dataset was constructed based on historical time-series operational data; The bidirectional temporal neural network model based on the attention mechanism is trained using the sample dataset to obtain the target charge difference evaluation model.

3. The method according to claim 2, characterized in that, The bidirectional temporal neural network model includes: a bidirectional gated recurrent unit encoding layer, an attention mechanism layer, a context vector generation layer, and a fully connected regression layer; the historical time-series running data includes historical running data at multiple time steps; The step of training the attention-based bidirectional temporal neural network model based on the sample dataset to obtain the target charge difference assessment model includes: The target hiding state of each time step is determined based on the bidirectional gated loop unit encoding layer, the historical running data of each time step, and the preset time sequence. The current charge-to-capacity difference prediction value of the training battery pack is output based on the attention mechanism layer, the target hidden state, the context vector generation layer, and the fully connected regression layer. The current training loss of the bidirectional temporal neural network model is determined based on the preset loss function, the predicted value of the current charging load difference, and the actual charge. The model then returns to the step of determining the target hidden state at each time step until the current training loss reaches the preset training loss, thus obtaining the target charge difference evaluation model.

4. The method according to claim 3, characterized in that, The bidirectional gated loop unit encoding layer includes a forward gated loop unit and a backward gated loop unit; the preset time sequence includes a first preset time sequence and a second preset time sequence; The step of determining the target hiding state at each time step based on the bidirectional gated loop unit encoding layer, the historical running data at each time step, and the preset time sequence includes: The forward-gated loop unit processes the historical running data of each time step according to the first preset time order to obtain the first hidden state of each time step along the first preset time order. The backward gated loop unit processes the historical running data of each time step according to the second preset time order to obtain the second hidden state of each time step along the second preset time order; The target hidden state of each time step is determined based on the first hidden state and the second hidden state of each time step.

5. The method according to claim 3, characterized in that, The step of outputting the predicted current charge-to-weight ratio difference of the trained battery pack based on the attention mechanism layer, the target hidden state, the context vector generation layer, and the fully connected regression layer includes: The target attention weight of the target hidden state is determined based on the attention mechanism layer and the target hidden state. The context vector of the target dimension is determined based on the context vector generation layer, each target hidden state, and the target attention weights corresponding to each target hidden state; The predicted value of the current charging load difference is output based on the context vector of the fully connected regression layer and the target dimension.

6. The method according to claim 5, characterized in that, The fully connected regression layer includes a first fully connected sub-layer, a second fully connected sub-layer, and an output layer; The step of outputting the predicted value of the current charging load difference based on the context vector of the fully connected regression layer and the target dimension includes: The first fully connected sublayer performs dimensionality reduction on the context vector of the target dimension to obtain the context vector of the first dimension. The second fully connected sublayer is used to reduce the dimensionality of the first-dimensional context vector to obtain the second-dimensional context vector. The context vector of the second dimension is linearly transformed through the output layer to obtain the predicted value of the current charging load difference.

7. The method according to claim 1, characterized in that, The inconsistency in determining the state of charge of the target battery pack based on the difference in current charging capacity and the theoretical charging capacity includes: The inconsistency of the state of charge of the target battery pack is determined based on the ratio of the difference between the current charging capacity and the theoretical charging capacity.

8. A battery pack state-of-charge assessment system, characterized in that, include: The first acquisition module is used to acquire the current timing operation data of the target battery pack under the current charging condition; The second acquisition module is used to acquire the theoretical charge capacity of the target battery pack under the current time-series operating data in the current charging condition. The third acquisition module is used to acquire the target charge difference assessment model; The target charge difference assessment model is trained based on a bidirectional temporal neural network model with an attention mechanism; The first determining module is used to determine the current charging charge difference of the target battery pack based on the current time-series operating data and the target charge difference evaluation model. The second determining module is used to determine the inconsistency of the state of charge of the target battery pack based on the difference in the current charging capacity and the theoretical charging capacity.

9. A vehicle, characterized in that, Includes the battery pack state of charge assessment system as described in claim 8.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which is loaded by a processor to perform the steps of the method as described in any one of claims 1-7.

Images