Battery state parameter prediction method and device, prediction model training method and device, energy storage equipment and computer readable storage medium

Abstract

The invention provides a battery state parameter prediction method and device, a prediction model training method and device, energy storage equipment and a computer readable storage medium. The battery state parameter prediction method comprises the steps that a time characteristic sequence of a battery pack is acquired, the time characteristic sequence comprises data of at least one observation parameter of the battery pack in a historical time period, and the observation parameter is related to a battery state parameter; and processing the time feature sequence based on a prediction model to obtain a battery state parameter prediction value at the current moment, the prediction model comprising a first network structure and a second network structure, the first network structure being based on a cross-variable attention mechanism Crossform architecture, and the second network structure being based on a feedforward neural network. According to the battery state parameter prediction method provided by the invention, the accuracy of battery state parameter prediction can be improved.

Description

Technical Field

[0001] The present application relates to the field of battery management technology, and in particular to a battery state parameter prediction method, a prediction model training method, an apparatus, an energy storage device, and a computer-readable storage medium. Background Art

[0002] With the development of electronic technology, the application of batteries is becoming more and more extensive. For example, lithium-ion batteries have become the first choice for electric vehicles, smart phones, portable electronic devices and other fields due to their high energy density, fast charging, long life, low self-discharge rate and other advantages. In the battery management system (BMS), various battery status parameters, such as battery state of charge (SOC), battery health (SOH), battery power (SOP) and battery energy (SOE), are one of the key monitoring indicators and are crucial to the safe use of batteries. Accurate estimation of battery status parameters not only provides BMS with real-time battery remaining capacity, real-time battery health status and power information, but also reduces the probability of battery overcharge and overdischarge, prolongs battery life, and prevents explosions caused by thermal reactions, thereby improving the safety of battery use.

[0003] Related technologies mainly predict battery status parameters based on recurrent neural network models such as LSTM and Transformer models based on attention mechanisms. However, recurrent neural network models cannot capture long-term dependencies of longer time series, and their training and reasoning efficiency is limited due to the difficulty in achieving effective parallelization. The Transformer model performs well in time series prediction tasks due to its self-attention mechanism. However, for multivariate time series, the Transformer model cannot effectively capture the dependencies of variable dimensions. The prediction of battery status parameters is closely related to multiple variables such as voltage, current, and temperature. Summary of the invention

[0004] In view of this, the present application provides a battery state parameter prediction method, a prediction model training method, an apparatus, an energy storage device and a computer-readable storage medium to improve the accuracy of battery state parameter prediction.

[0005] The present application provides a first aspect of a battery status parameter prediction method, which is applied to a battery pack. The method includes: obtaining a time feature sequence of the battery pack, the time feature sequence includes data of at least one observed parameter of the battery pack in a historical time period, and the observed parameter is related to the battery status parameter; processing the time feature sequence based on a prediction model to obtain a predicted value of the battery status parameter at the current moment, wherein the prediction model includes a first network structure and a second network structure, the first network structure is based on a cross-variable attention mechanism Crossformer architecture, and the second network structure is based on a feedforward neural network.

[0006] The second aspect of the present application provides a training method for a prediction model, including: determining a training set based on historical data of various observed parameters of a battery pack within a historical time period; determining a verification set based on historical data of battery status parameters of the battery pack within a historical time period; training a model to be tested based on the training set to output a prediction result, and when the loss function value determined according to the verification set and the prediction result reaches a preset threshold, a trained prediction model is obtained, and the prediction model includes a first network structure and a second network structure, the first network structure is based on a cross-variable attention mechanism Crossformer architecture; the second network structure is based on a feedforward neural network.

[0007] The third aspect of the present application provides a prediction device, including: an acquisition module, used to obtain a time feature sequence of a battery pack, the time feature sequence includes data of at least one observation parameter of the battery pack in a historical time period, and the observation parameter is related to a battery status parameter; a prediction module, used to process the time feature sequence based on a prediction model to obtain a prediction value of the battery status parameter, wherein the prediction model includes a first network structure and a second network structure, the first network structure is based on a cross-variable attention mechanism Crossformer architecture, and the second network structure is based on a feedforward neural network.

[0008] The fourth aspect of the present application provides a training device, including: a first determination module, which determines a training set based on historical data of various observation parameters of a battery pack within a historical time period, and the observation parameters are related to battery status parameters; a second determination module, which determines a verification set based on historical data of battery status parameters of the battery pack within a historical time period; a training module, which trains a model to be tested based on the training set to output a prediction result, and when the loss function value obtained according to the verification set and the prediction result reaches a preset threshold, a trained prediction model is obtained, and the prediction model includes a first network structure and a second network structure, the first network structure is based on a cross-variable attention mechanism Crossformer architecture; the second network structure is based on a feedforward neural network.

[0009] In a fifth aspect, the present application provides an energy storage device, including a battery pack and a controller. The controller is used to execute a battery state parameter prediction method.

[0010] In a sixth aspect, the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a controller, the controller implements a battery state parameter prediction method or a prediction model training method.

[0011] The battery status parameter prediction method provided in the present application processes the time characteristic sequence based on the prediction model to obtain the predicted value of the battery status parameter at the current moment. Among them, the time characteristic sequence includes data of at least one observation parameter, and the observation parameter is related to the battery status parameter. The prediction model includes a first network structure and a second network structure. The first network structure is based on the Crossformer architecture, so that the time dependence and variable correlation between the observation parameter and the battery status parameter can be effectively captured through the first network structure, so as to first predict based on the observation parameter. Then, the prediction is made based on the second network structure to obtain the predicted value of the battery status parameter at the current moment, so as to further improve the accuracy of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present application and should not be regarded as limiting the scope of protection of the present application. In each of the drawings, similar components are numbered similarly.

[0013] Figure 1 A schematic diagram of the first application scenario of the battery state parameter prediction method provided in this application.

[0014] Figure 2 A schematic diagram of a second application scenario of the battery state parameter prediction method provided in this application.

[0015] Figure 3 A flowchart of a battery status parameter prediction method provided in an embodiment of the present application.

[0016] Figure 4 Schematic diagram of the flow of sub-steps of step S302 in one embodiment of the present application.

[0017] Figure 5 Schematic diagram of the flow of sub-steps of step S404 in one embodiment of the present application.

[0018] Figure 6 Schematic diagram of the flow of sub-steps of step S501 in an embodiment of the present application.

[0019] Figure 7 Schematic diagram of the flow of sub-steps of step S603 in one embodiment of the present application.

[0020] Figure 8A flowchart of a method for training a prediction model provided in accordance with an embodiment of the present application.

[0021] Fig. 9 This is a structural block diagram of a prediction device provided in one embodiment of the present application.

[0022] Fig.10 This is a structural block diagram of a training device provided in one embodiment of the present application.

[0023] Fig.11 A structural block diagram of an electronic device provided in one embodiment of the present application.

[0024] Fig.12 A structural block diagram of a computer storage medium provided in one embodiment of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0026] It is understandable that the connection relationship described in this application refers to direct or indirect connection. For example, the connection between A and B can be either a direct connection between A and B or an indirect connection between A and B through one or more other electrical components. For example, A and C can be directly connected, and C and B can be directly connected, so that A and B are connected through C. It is also understandable that the "A connects B" described in this application can be a direct connection between A and B, or an indirect connection between A and B through one or more other electrical components.

[0027] In the description of this application, unless otherwise specified, " / " means "or", for example, A / B can mean A or B. "And / or" in this article is only a way to describe the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0028] In the description of this application, words such as "first" and "second" are only used to distinguish different objects, and do not limit the quantity and execution order, and words such as "first" and "second" do not necessarily limit differences. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0030] Some embodiments will be described below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments may be combined with each other.

[0031] With the development of electronic technology, the application of batteries is becoming more and more extensive. For example, lithium-ion batteries have become the first choice for electric vehicles, smart phones, portable electronic devices and other fields due to their high energy density, fast charging, long life, low self-discharge rate and other advantages. In the battery management system (BMS), various battery status parameters, such as battery state of charge (SOC), battery health (SOH), battery power (SOP) and battery energy (SOE), are one of the key monitoring indicators and are crucial to the safe use of batteries. Accurate estimation of battery status parameters not only provides BMS with real-time battery remaining capacity, real-time battery health status and power information, but also reduces the probability of battery overcharge and overdischarge, prolongs battery life, and prevents explosions caused by thermal reactions, thereby improving the safety of battery use.

[0032] Related technologies mainly predict battery status parameters based on recurrent neural network models such as LSTM and Transformer models based on attention mechanisms. However, recurrent neural network models cannot capture long-term dependencies of longer time series, and their training and reasoning efficiency is limited due to the difficulty in achieving effective parallelization. The Transformer model performs well in time series prediction tasks due to its self-attention mechanism. However, for multivariate time series, the Transformer model cannot effectively capture the dependencies of variable dimensions. The prediction of battery status parameters is closely related to multiple variables such as voltage, current, and temperature.

[0033] In view of this, the present application provides a battery state parameter prediction method, a prediction model training method, a training device and an energy storage device to improve the accuracy of battery state parameter prediction.

[0034] See also Figure 1 , Figure 1Schematic diagram of the first application scenario of the battery status parameter prediction method provided in the present application. The energy storage device 10 includes a battery pack 11 and a controller 12. The battery pack 11 includes at least one battery cell. The battery pack 11 is used to store electrical energy, and the battery pack 11 can be charged and discharged. The controller 12 can obtain various parameters of the battery pack 11 to monitor and manage the battery pack 11. In some embodiments, a battery management system (Battery Management System, BMS) may be loaded on the controller 12. It is understandable that the energy storage device 10 can be any electronic device equipped with a battery pack 11. The present application does not limit the specific functions of the energy storage device 10. In some embodiments, the battery status parameter prediction method can be executed by the controller 12.

[0035] Please continue reading Figure 2 , Figure 2 A schematic diagram of a second application scenario of the battery state parameter prediction method provided in this application. Figure 2 In the scenario shown, the prediction device 20 can obtain the corresponding time feature sequence, and use the time feature sequence to predict the battery state parameters to output the battery state parameter prediction value. The prediction device 20 includes but is not limited to fixed devices and / or mobile devices. For example, fixed devices include but are not limited to servers, and the server can be a cloud server or an ordinary server or an edge server. For example, mobile devices include but are not limited to mobile phones, tablet computers, vehicle-mounted terminals, etc. In some possible implementations, the battery state parameter prediction method can also be implemented by a processor calling computer-readable instructions stored in a memory.

[0036] Please continue reading Figure 3 , Figure 3 A flowchart of a method for predicting battery state parameters provided in an embodiment of the present application is provided. The method comprises the following steps:

[0037] Step S301: Acquire a time characteristic sequence of a battery pack, where the time characteristic sequence includes data of at least one observed parameter of the battery pack within a historical time period, where the observed parameter is related to a battery state parameter.

[0038] Among them, the battery pack 12 can generate multiple parameters, including but not limited to current, voltage, temperature, battery state of charge SOC, battery state of health SOH, battery state of power SOP and battery state of energy SOE. Among them, the multiple parameters can be divided into observation parameters and battery state parameters. Among them, the battery state parameter can be any one of the parameters such as battery state of charge SOC, battery state of health SOH, battery state of power SOP and battery state of energy SOE, and the battery state parameter can be used as a parameter to be predicted in this application. In some embodiments, among the multiple parameters of the battery pack 12, parameters other than the parameters to be predicted can be used as observation parameters. For example, when the battery state parameter is SOC, the observation parameter can be current, voltage, temperature, SOH, etc.; when the battery state parameter is SOH, the observation parameter can be current, voltage, temperature, SOC, etc.

[0039] Among them, the observation parameter is related to the battery state parameter, and the observation parameter is used to participate in the prediction of the battery state parameter. The observation parameter can be a parameter with a high degree of correlation with the battery state parameter. The degree of correlation can be represented by a statistical value, such as a correlation coefficient, a correlation ratio, etc. For example, in some embodiments, the observation parameter can be a parameter whose absolute value of the correlation coefficient with the battery state parameter is greater than or equal to a preset threshold. This application does not limit the type and number of observation parameters.

[0040] Understandably, the time feature sequence is used to represent the characteristics of the battery pack 11 within the historical time period. The time feature sequence may include data of at least one observed parameter of the battery pack 11 within the historical time period. Specifically, the time feature sequence may be a data set formed by arranging at least one observed parameter of the battery pack 11 within the historical time period in the order of the time when its data was generated. For example, the time feature sequence can be represented by a T*P matrix, where T represents time and P represents the observed parameter. When the time feature sequence is a 10*3 matrix, the time feature sequence represents a feature sequence formed by data generated at 10 consecutive time points for the three observed parameters of the battery pack 11 within the historical time period.

[0041] In some embodiments, the time feature sequence can be generated by collecting data of the observed parameters of the battery pack 11 within a historical time period and performing data cleaning, such as filtering, on the collected data.

[0042] Step S302: Process the time feature sequence based on the prediction model to obtain the predicted value of the battery status parameter at the current moment, wherein the prediction model includes a first network structure and a second network structure, the first network structure is based on the cross-variable attention mechanism Crossformer architecture, and the second network structure is based on a feedforward neural network.

[0043] Understandably, the Crossformer architecture mines the feature mapping relationship of the input sequence from the time scale and the variable scale respectively, which is conducive to making full use of information at different time scales. The Crossformer architecture takes into account both time dependency and variable correlation, which is conducive to improving the accuracy of parameter prediction. The feedforward neural network (FNN) includes multiple layers of neurons to form an input layer, a hidden layer and an output layer. FNN is suitable for a variety of tasks such as classification and regression. The present application does not limit the number of neuron levels included in the second network structure, and the second network structure can apply different feedforward neural networks. For example, in some embodiments, the architecture of the second network structure can be a multilayer perceptron architecture.

[0044] In summary, the battery status parameter prediction method provided in the present application processes the time characteristic sequence based on the prediction model to obtain the predicted value of the battery status parameter at the current moment. Among them, the time characteristic sequence includes data of at least one observation parameter, and the observation parameter is related to the battery status parameter. The prediction model includes a first network structure and a second network structure. The first network structure is based on the Crossformer architecture, so that the time dependence and variable correlation between the observation parameters and the battery status parameters can be effectively captured through the first network structure, so as to first predict based on the observation parameters. Then, prediction is made based on the second network structure to obtain the predicted value of the battery status parameter at the current moment, so as to further improve the accuracy of the prediction.

[0045] Please continue reading Figure 4 In some embodiments, step S302 includes the following sub-steps:

[0046] Step S401: Processing the time feature sequence based on the first network structure to obtain a battery state parameter feature vector.

[0047] In some embodiments, the battery state parameter feature vector may be a feature vector of the battery state parameter output by the first network structure based on the time feature sequence.

[0048] Step S402: Processing the battery state parameter feature vector according to the second network structure to obtain a battery state parameter reference value at the current moment.

[0049] The battery state parameter reference value is used to characterize the predicted value of the battery state parameter output by the second network structure at the current moment.

[0050] Step S403: Obtain the observed value change and the reference value change of the battery status parameter observation value and the battery status parameter reference value within a preset time interval.

[0051] Among them, the battery status parameter observation value refers to the measured value or calculated value of the battery status parameter. That is to say, the battery status parameter observation value can be a battery status parameter measurement value obtained by conventional measurement means or calculation means. This application does not limit the specific duration of the preset interval duration. For example, the preset interval duration can be the interval duration between the current moment and the previous moment.

[0052] In some embodiments, the observed value change amount may be the difference between the battery state parameter observed value at the last moment of the preset interval and the battery state parameter observed value at the start moment of the preset interval. In other embodiments, the observed value change amount may also be the difference between the battery state parameter observed value at the start moment of the preset interval and the battery state parameter observed value at the last moment of the preset interval. This application is not limited to this.

[0053] Accordingly, in some embodiments, the reference value variation may be the difference between the reference value of the battery status parameter at the last moment of the preset interval duration output by the second network structure and the reference value of the battery status parameter at the starting moment of the preset interval duration output by the second network structure. In other embodiments, the reference value variation may also be the difference between the reference value of the battery status parameter at the starting moment of the preset interval duration output by the second network structure and the reference value of the battery status parameter at the last moment of the preset interval duration output by the second network structure. This application is not limited to this.

[0054] Step S404: Determine the predicted value of the battery state parameter at the current moment according to the predicted value of the battery state parameter at the previous moment, the change in the observed value and the change in the reference value.

[0055] Due to the complexity of the structure and parameters in the neural network model, the error is likely to be gradually amplified as the neural network model runs. In addition, since the change trend of the battery state parameters is relatively stable in a short period of time, the observed value change in a short period of time can have a certain reference value when representing the change trend of the battery state parameters. Therefore, in step S404, the change trend of the battery state parameter prediction value can be jointly corrected based on the observed value change and the reference value change, so as to determine the battery state parameter prediction value at the current moment based on the battery state parameter prediction value at the previous moment, thereby improving the accuracy of the battery state parameter prediction value at the current moment.

[0056] It is understandable that, at the initial moment of running the battery state parameter prediction method, the battery state parameter prediction value at the previous moment may be a preset battery state parameter value.

[0057] In summary, by executing steps S401 to S402, the accuracy of the battery state parameter prediction value at the current moment can be improved by combining the observed value change and the battery state parameter reference value change output by the prediction model.

[0058] Please continue reading Figure 5 In some embodiments, step S404 includes the following sub-steps:

[0059] Step S501: Determine the adjustment coefficient according to the observed value change and the reference value change.

[0060] The adjustment coefficient can be used to adjust the growth trend of the predicted value of the battery state parameter.

[0061] In some embodiments, the observed value change can be used as a reference, and the multiple of the reference value change compared to the observed value change can be calculated as the adjustment coefficient. In other embodiments, the multiple of the reference value change compared to the observed value change can also be calculated, and the multiple can be adjusted based on the historical battery state parameter data to obtain the adjustment coefficient. This application does not limit the specific calculation method for determining the adjustment coefficient.

[0062] Step S502: Obtain the predicted value of the battery state parameter at the current moment according to the predicted value of the battery state parameter at the previous moment, the variation of the observed value and the adjustment coefficient.

[0063] In some embodiments, the predicted value of the battery status parameter at the previous moment can be used as a reference value, and the product of the observed value change and the adjustment coefficient is calculated as the predicted value change from the previous moment to the current moment, and then the predicted value of the battery status parameter at the previous moment and the predicted value change are added to obtain the predicted value of the battery status parameter at the current moment.

[0064] In some other embodiments, the first adjustment parameter and the second adjustment parameter may be set respectively to calculate the product of the first adjustment parameter and the predicted value of the battery state parameter at the previous moment, plus the product of the second adjustment parameter and the predicted value change as the predicted value of the battery state parameter at the current moment. The first adjustment parameter and the second adjustment parameter may be negative or positive. The present application does not limit the specific calculation method for calculating the predicted value of the state parameter at the current moment according to the predicted value of the battery state parameter at the previous moment, the observed value change and the adjustment coefficient in step S502.

[0065] In some embodiments, the battery state parameter is taken as the battery state of charge SOC as an example to further illustrate the battery state parameter prediction method provided by the present application. Figure 6 In some embodiments, step S501 may include the following sub-steps:

[0066] Step S601: Obtain the battery current at the current moment and the fully charged battery capacity of the battery pack to determine the change in the observed value.

[0067] In some embodiments, a current sampling circuit is provided at the output port of the battery pack 11. Thus, the battery current at the current moment can be obtained through the current sampling circuit.

[0068] The fully charged battery capacity may be a preset value. For example, in some embodiments, the fully charged battery capacity may be preset in the energy storage device 10. The fully charged battery capacity may also be a variable value, in which case the fully charged battery capacity may be calculated based on a preset algorithm. This application does not limit the specific algorithm for calculating the fully charged battery capacity.

[0069] Understandably, when the battery state parameter is SOC, the battery state parameter observation value can be calculated based on the ampere-hour integration method. In this way, the observed value change between the current moment and the previous moment can be obtained based on the battery current at the current moment, the preset interval time and the full-charge battery capacity of the battery pack, that is, the SOC change calculated based on the ampere-hour integration method.

[0070] In some embodiments, the observed value change can be determined based on the following formula (1):

[0071]

[0072] In the above formula (1), D soC1 Indicates the change in observed value, I t represents the battery current at the current moment; Δt represents the interval duration; Q represents the fully charged battery capacity.

[0073] In other embodiments, the observed value variation can also be calculated based on other formulas. The present application does not limit the specific calculation method for calculating the observed value variation.

[0074] Step S602: determining a reference value variation according to a battery state parameter reference value at a current moment and a battery state parameter reference value at a previous moment.

[0075] In step S602, the difference between the reference value of the battery state parameter output by the second network structure at the previous moment and the reference value of the battery state parameter output by the second network structure at the current moment may be calculated as the reference value variation.

[0076] In some embodiments, after the second network structure outputs the battery state parameter reference value, it can also perform smoothing filtering based on the battery state parameter reference value at the current moment and the battery state parameter reference value obtained in the historical time period to filter out abnormal values, thereby updating all battery state parameter reference values. This can reduce the impact of noise on the reference value variation calculated in step S602.

[0077] Step S603: Determine the reference coefficient according to the observed value change and the reference value change.

[0078] The reference coefficient may be a ratio between a change in a reference value and a change in an observed value, and is used to characterize a multiple of a change in a battery state parameter predicted by a prediction model compared to a change in a battery state parameter obtained by an observation method.

[0079] In some embodiments, the reference coefficient may be determined according to the following formula (2):

[0080]

[0081] In the above formula (2), λ t Denotes the reference coefficient, D soc2 Indicates the reference value change.

[0082] Step S604: determining an adjustment coefficient according to the reference coefficient, and the adjustment coefficient is positively correlated with the reference coefficient.

[0083] It can be understood that since the reference value is obtained based on the prediction of the first network structure and the second network structure, in order to avoid the battery status parameter predicted by the prediction model (i.e., the battery SOC in this embodiment) from changing too much or too little, it is possible to determine whether the change trend of the battery status parameter predicted by the prediction model is abnormal based on the benchmark coefficient, and adjust it based on the benchmark coefficient to obtain the adjustment coefficient.

[0084] Among them, since the adjustment coefficient is positively correlated with the reference coefficient, when the predicted value of the battery status parameter at the current moment is determined according to the adjustment coefficient, the change trend of the battery status parameter predicted by the prediction model can be retained, and the change amplitude of the battery status parameter prediction value can be avoided to be too large or too small and inaccurate.

[0085] Please continue reading Figure 7 In some embodiments, step S603 includes the following sub-steps:

[0086] Step S701: Determine the weight coefficient according to the reference coefficient.

[0087] The weight coefficient is used to adjust the influence of the reference coefficient on the adjustment coefficient. In some embodiments, the weight coefficient can be determined based on the size of the reference coefficient. For example, when the reference coefficient is within the normal range, the weight coefficient is a first value; when the reference coefficient is greater than the maximum value within the normal range or less than the minimum value within the normal range, the weight coefficient is a second value, and the first value is greater than the first value.

[0088] Step S702: Determine the adjustment coefficient according to the reference coefficient and the weight coefficient.

[0089] In some embodiments, another weight coefficient of the change trend of the observed value change obtained by the ampere-hour integration method to the predicted value of the battery state parameter can be determined according to the weight coefficient. In this way, the adjustment coefficient can be determined according to the two weight coefficients, that is, the adjustment coefficient is determined based on the reference value change obtained by the prediction model and the observed value change obtained by the ampere-hour integration method, so that when the predicted value of the battery state parameter at the current moment is determined according to the adjustment coefficient, the predicted value of the battery state parameter at the current moment can be determined by the prediction model and the ampere-hour integration method at the same time, thereby improving the prediction accuracy.

[0090] In some embodiments, the adjustment coefficient may be determined according to the following formula (3):

[0091] λ ′ t =αλ t +(1-α) (3)

[0092]

[0093] In the above formula (3), λ′ t represents the adjustment coefficient, α represents the weight coefficient, 1-α represents another weight coefficient, λ t Represents the base coefficient.

[0094] In this way, by executing the above steps S701 to S702, the adjustment coefficient can be determined.

[0095] In some embodiments, step S502 includes:

[0096] The predicted value of the battery state parameter at the previous moment is integrated in ampere-hours according to the adjustment coefficient, the battery current at the current moment, and the fully charged battery capacity, so as to obtain the predicted value of the battery state parameter at the current moment.

[0097] In one embodiment, the battery state parameter reference value at the current moment may be used as a benchmark value for ampere-hour integration, and the change trend during the ampere-hour integration process may be adjusted according to an adjustment coefficient to obtain a predicted value of the battery state parameter at the current moment.

[0098] For example, in some embodiments, the predicted value of the battery state parameter at the current moment may be determined based on the following formula (4):

[0099]

[0100] In the above formula (4), SOC′ t Represents the predicted value of the battery state parameter at the current moment, SOC′ t-1 Indicates the predicted value of the battery status parameter at the previous moment.

[0101] In this way, by executing the above steps, the predicted value of the battery state parameter at the current moment can be obtained.

[0102] In some embodiments, the first network structure includes a dimensionally segmented embedding layer, an encoder, a decoder, and a linear mapping layer.

[0103] Step S301 includes:

[0104] Step S3011: Input the time feature sequence into the dimension segmentation embedding layer to generate a first feature vector, the dimension of the first feature vector is smaller than the dimension of the time feature sequence.

[0105] In some embodiments, the dimension segment embedding layer is used to slice each variable in the time feature sequence in the time dimension, and then embed the variable time segment to obtain a two-dimensional vector h i,d , which is the first eigenvector. Where d represents the dth variable and i represents the i-th time segment.

[0106] Understandably, various observation parameters of the battery may be collected, and then data cleaning, filtering and other operations may be performed on the data collected for each observation parameter to obtain a time feature sequence.

[0107] Step S3012: input the first feature vector into the encoder to generate a second feature vector.

[0108] Among them, the encoder is used to fuse adjacent two-dimensional vectors in the time dimension, and use the cross-time and cross-variable attention module to obtain the encoded output, that is, the second eigenvector.

[0109] In some embodiments, the cross-temporal and cross-variable attention module is

[0110]

[0111] Where l represents the lth encoding layer, TSA(·) represents the cross-time and cross-variable attention mechanism, M[·] represents the fusion function, is the encoded output of the lth layer.

[0112] Step S3013: input the second feature vector into the decoder to generate a third feature vector.

[0113] In some embodiments, the decoder includes a cross-time and cross-variable attention module, a multi-head self-attention module, a residual connection module, a layer normalization module, and a feed-forward network module.

[0114] Among them, the decoder can be calculated based on the following formula:

[0115]

[0116] Z3 =LayerNorm(Z 1 +Z 2 )

[0117] Z dec,l =LayerNorm(Z 3 +MLP(Z 3 ))

[0118] Among them, E pos is the position code, Z 1 ,Z 2 ,Z 3 is the intermediate calculation result, is the encoding output of the lth layer, MSA(·) is the multi-head self-attention mechanism function, LayerNorm(·) is the layer normalization, and MLP(·) is the feedforward network layer.

[0119] The decoding stage outputs Z to the previous layer dec,l-1 Or position code E pos , the intermediate calculation result Z is calculated by using the cross-time and cross-variable attention module 1 , and then use the output of the corresponding encoding layer As the input of the multi-head self-attention mechanism, the intermediate calculation result Z is calculated 2 , and apply the residual connection structure and layer normalization LayerNorm(·) calculation to obtain the intermediate calculation result Z 3 Finally, the decoded output Z is obtained through the feedforward network MLP(·), residual connection structure, and layer normalization calculation dec,l .

[0120] Step S804: input the third feature vector into the linear mapping layer to generate a battery state parameter feature vector.

[0121] After each decoding layer output is obtained in step S803, each decoding layer output is linearly mapped and summed based on the following formula to obtain the final output:

[0122]

[0123] Among them, W l is the network parameter of the l-th layer linear mapping, is the battery state parameter feature vector.

[0124] In some embodiments, the cross-time and cross-variable attention module mentioned above is divided into two stages, namely the cross-time attention stage and the cross-variable attention stage. In the cross-time attention stage, for each variable, the multi-head self-attention mechanism function MSA is used in the time dimension. time(·), and apply the residual structure and layer normalization to calculate the intermediate calculation result Z 1 , and then through the feedforward network, residual connection structure, and layer normalization calculation, the cross-time attention stage output Z is obtained time :

[0125] Z 1 =LayerNorm(Z :,d +MSA time (Z :,d ,Z :,d ,Z :,d ))

[0126] Z time =LayerNorm(Z 1 +MLP(Z 1 ))

[0127] In the cross-variable attention stage, the router mechanism is used to pass information through the router to reduce the high complexity caused by directly applying the multi-head self-attention mechanism calculation. Specifically, for each time segment, the multi-head self-attention mechanism function is first used in the variable dimension. Compress the output of the attention stage across time into information B i,: , and then use the multi-head self-attention mechanism function Restore it, and apply the residual structure and layer normalization to obtain the intermediate calculation result Z 1 Finally, through the feedforward network, residual connection structure, and layer normalization calculation, the cross-variable attention stage output Z is obtained dim :

[0128]

[0129] Z dim =LayerNorm(Z 1 +MLP(Z 1 ))

[0130] Among them, R i,: are learnable router parameters.

[0131] See also Figure 8 , an embodiment of the present application further provides a prediction model training method, the training method comprising:

[0132] Step S801: Determine a training set based on historical data of various observed parameters of the battery pack within a historical time period.

[0133] Among them, the observed parameter may be a parameter with a high degree of correlation with the battery state parameter. In some embodiments, the degree of correlation between other parameters of the battery pack other than the battery state parameter and the battery state parameter may be calculated, so as to determine each observed parameter according to the degree of correlation. The training set may include multiple sets of training time feature sequences generated in different historical time periods.

[0134] In some embodiments, data cleaning and filtering operations may be performed on various observation parameters to form a training set including various observation parameters of different historical time periods.

[0135] This application does not limit the specific length of the historical time period. For example, the historical time period can be the time period when the battery pack is discharged from a fully charged state to a preset threshold, or the historical time period can be one hour when the battery pack is in a charging state or a discharging state.

[0136] Step S802: Determine a verification set based on historical data of the battery status parameters of the battery pack within a historical time period.

[0137] The validation set may include real data of battery status parameters generated by the battery pack in a historical time period corresponding to that in the training set.

[0138] Step S803: Train the model to be tested based on the training set to output the prediction result. When the loss function value determined according to the verification set and the prediction result reaches a preset threshold, a trained prediction model is obtained. The prediction model includes a first network structure and a second network structure. The first network structure is based on the cross-variable attention mechanism Crossformer architecture; the second network structure is based on a feedforward neural network.

[0139] The present application does not limit the loss function selected in step S903. For example, in some embodiments, the loss function may include at least one of a mean square error loss (MSE Loss) function, a mean absolute error loss (MAE Loss) function, and the like.

[0140] In some embodiments, the first network structure includes a dimension segment embedding layer, an encoder, a decoder, and a linear mapping layer, and step S803 includes the following sub-steps:

[0141] Inputting the training time feature sequence about the observation parameters in the training set into the dimension segmentation embedding layer to generate a first training feature vector, wherein the dimension of the first training feature vector is smaller than the dimension of the training time feature sequence;

[0142] Inputting the first training feature vector sequence into an encoder to generate a second training feature vector;

[0143] Inputting the second training feature vector into a decoder to generate a third training feature vector;

[0144] Inputting the third training feature vector into the linear mapping layer to generate a state parameter training feature vector;

[0145] The state parameter training feature vector is processed based on the second network structure to output a prediction result.

[0146] In some embodiments, the decoder includes a cross-time and cross-variable attention module, a multi-head self-attention module, a residual connection module, a layer normalization module, and a feed-forward network module.

[0147] In the embodiment of the present application, a more detailed functional description of each of the above modules can be referred to the corresponding content of the aforementioned part, which will not be repeated here.

[0148] See also Fig. 9 , an embodiment of the present application further provides a prediction device 100, comprising:

[0149] The acquisition module 110 is used to acquire a time feature sequence of the battery pack, where the time feature sequence includes data of at least one observed parameter of the battery pack within a historical time period, and the observed parameter is related to the battery state parameter.

[0150] The prediction module 120 is used to process the time feature sequence based on the prediction model to obtain the predicted value of the battery state parameter at the current moment, wherein the prediction model includes a first network structure and a second network structure, the first network structure is based on the cross-variable attention mechanism Crossformer architecture, and the second network structure is based on a feedforward neural network.

[0151] In the embodiment of the present application, a more detailed functional description of each of the above modules can be referred to the corresponding content of the aforementioned part, which will not be repeated here.

[0152] See also Fig.10 , an embodiment of the present application further provides a training device 200, comprising:

[0153] The first determination module 210 determines a training set based on historical data of various observed parameters of the battery pack within a historical time period, and the observed parameters are related to the battery state parameters.

[0154] The second determination module 220 determines a verification set based on historical data of the battery status parameters of the battery pack within a historical time period.

[0155] The training module 230 trains the model to be tested based on the training set to output the prediction result. When the loss function value obtained according to the verification set and the prediction result reaches a preset threshold, a trained prediction model is obtained. The prediction model includes a first network structure and a second network structure. The first network structure is based on the Crossformer architecture of the cross-variable attention mechanism; the second network structure is based on a feedforward neural network.

[0156] Please refer again Figure 1 The present application provides an energy storage device 10, including a battery pack 11 and a controller 12, and the controller 12 is used to execute the battery state parameter prediction method described in any of the above embodiments.

[0157] See also Fig.11 The present application provides an electronic device 20, including a processor 21, and a memory 22 for storing a program. The program includes instructions, which, when executed by the processor 21, cause the processor 21 to execute the battery state parameter prediction method or the prediction model training method as described in any of the above embodiments. The electronic device 20 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device 20 can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0158] See also Fig.12 The embodiment of the present application also provides a computer storage medium 30, which stores a computer program 31. When the computer program 31 is executed by the processor, the processor executes the above-mentioned battery state parameter prediction method or prediction model training method.

[0159] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer storage medium or transmitted through the computer storage medium. The computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (Digital Subscriber Line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode. The computer storage medium can be any available medium that a computer can access or a data storage device such as a server, data center, etc. that contains one or more available media integrated. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a digital versatile disc (DVD)), or a semiconductor medium (eg, a solid state disk (SSD)).

[0160] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other media that can store program codes. In the absence of conflict, the technical features in this embodiment and the implementation scheme can be combined arbitrarily.

[0161] The present application is not limited to the above specific implementations. A person skilled in the art can easily understand that there are many alternatives to the test fixture of the present application without departing from the principle and scope of the present application. The scope of protection of the present application shall be subject to the content of the claims.

Claims

1. A battery state parameter prediction method, applied to a battery pack, characterized in that: The method comprises: Acquire a time characteristic sequence of the battery pack, wherein the time characteristic sequence includes data of at least one observed parameter of the battery pack in a historical time period, wherein the observed parameter is related to the battery state parameter; The time feature sequence is processed based on a prediction model to obtain a predicted value of a battery status parameter at a current moment, wherein the prediction model includes a first network structure and a second network structure, the first network structure is based on a cross-variable attention mechanism Crossformer architecture, and the second network structure is based on a feedforward neural network.

2. The method according to claim 1, characterized in that The processing of the time feature sequence based on the prediction model to obtain a predicted value of the battery state parameter at the current moment includes: Processing the time feature sequence based on the first network structure to obtain a battery state parameter feature vector; Processing the battery state parameter feature vector according to the second network structure to obtain a battery state parameter reference value at a current moment; Obtaining a battery status parameter observation value and a battery status parameter reference value within a preset time interval. The predicted value of the battery state parameter at the current moment is determined according to the predicted value of the battery state parameter at the previous moment, the observed value change and the reference value change.

3. The method according to claim 2, characterized in that The step of determining the predicted value of the battery state parameter at the current moment according to the predicted value of the battery state parameter at the previous moment, the observed value change, and the reference value change includes: Determine an adjustment coefficient according to the observed value change and the reference value change; The predicted value of the battery state parameter at the current moment is obtained according to the predicted value of the battery state parameter at the previous moment, the change in the observed value and the adjustment coefficient.

4. The method according to claim 3, characterized in that: When the battery state parameter is the battery state of charge SOC, determining the adjustment coefficient according to the observed value change and the reference value change includes: Obtaining the battery current at the current moment and the fully charged battery capacity of the battery pack to determine the change in the observed value; Determine a reference value change according to a battery state parameter reference value at a current moment and a battery state parameter reference value at a previous moment; Determine a reference coefficient according to the observed value change and the reference value change; The adjustment coefficient is determined according to the reference coefficient, and the adjustment coefficient is positively correlated with the reference coefficient.

5. The method according to claim 4, characterized in that The step of determining the adjustment coefficient according to the reference coefficient comprises: Determining a weight coefficient according to the reference coefficient; The adjustment coefficient is determined according to the reference coefficient and the weight coefficient.

6. The method according to claim 4, characterized in that The method of obtaining the predicted value of the battery status parameter at the current moment based on the predicted value of the battery status parameter at the previous moment, the change in the observed value and the adjustment coefficient includes: performing ampere-hour integration on the predicted value of the battery status parameter at the previous moment based on the adjustment coefficient, the battery current at the current moment and the fully charged battery capacity to obtain the predicted value of the battery status parameter at the current moment.

7. The method according to claim 1, characterized in that The first network structure includes a dimensional segment embedding layer, an encoder, a decoder, and a linear mapping layer. The processing of the time feature sequence based on the first network structure to obtain a battery state parameter feature vector includes: Inputting the time feature sequence into the dimension segmentation embedding layer to generate a first feature vector, wherein the dimension of the first feature vector is smaller than the dimension of the time feature sequence; Inputting the first feature vector into the encoder to generate a second feature vector; Inputting the second feature vector into the decoder to generate a third feature vector; The third feature vector is input into the linear mapping layer to generate the battery state parameter feature vector.

8. The method according to claim 7, characterized in that The decoder includes a cross-time and cross-variable attention module, a multi-head self-attention module, a residual connection module, a layer normalization module and a feedforward network module.

9. A method for training a prediction model, characterized in that: The training method comprises: Determine the training set based on historical data of various observed parameters of the battery pack within a historical time period; Determine a validation set based on historical data of the battery status parameters of the battery pack within the historical time period; The model to be tested is trained based on the training set to output a prediction result. When the loss function value determined according to the verification set and the prediction result reaches a preset threshold, the trained prediction model is obtained. The prediction model includes a first network structure and a second network structure. The first network structure is based on a cross-variable attention mechanism Crossformer architecture; the second network structure is based on a feedforward neural network.

10. The training method according to claim 9, characterized in that: The first network structure includes a dimension segment embedding layer, an encoder, a decoder, and a linear mapping layer, and the training of the model to be tested based on the training set includes: Inputting the training time feature sequence about the observation parameter in the training set into the dimension segmentation embedding layer to generate a first training feature vector, wherein the dimension of the first training feature vector is smaller than the dimension of the time feature training sequence; Inputting the first training feature vector into the encoder to generate a second training feature vector; Inputting the second training feature vector into the decoder to generate a third training feature vector; Inputting the third training feature vector into the linear mapping layer to generate the state parameter training feature vector; The state parameter training feature vector is processed based on the second network structure to output the prediction result.

11. The training method according to claim 10, characterized in that: The decoder includes a cross-time and cross-variable attention module, a multi-head self-attention module, a residual connection module, a layer normalization module and a feedforward network module.

12. A prediction device, characterized in that: include: An acquisition module, used to acquire a time characteristic sequence of a battery pack, wherein the time characteristic sequence includes data of at least one observation parameter of the battery pack within a historical time period, wherein the observation parameter is related to a battery state parameter; A prediction module is used to process the time feature series based on a prediction model to obtain a predicted value of a battery state parameter, wherein the prediction model includes a first network structure and a second network structure, the first network structure is based on a cross-variable attention mechanism Crossformer architecture, and the second network structure is based on a feedforward neural network.

13. A training device, characterized in that: include: A first determination module determines a training set based on historical data of various observed parameters of the battery pack within a historical time period, wherein the observed parameters are related to the battery state parameters; A second determination module determines a verification set based on historical data of the battery status parameter of the battery pack within the historical time period; A training module trains the model to be tested based on the training set to output a prediction result. When the loss function value obtained according to the verification set and the prediction result reaches a preset threshold, a trained prediction model is obtained. The prediction model includes a first network structure and a second network structure. The first network structure is based on a cross-variable attention mechanism Crossformer architecture; the second network structure is based on a feedforward neural network.

14. An energy storage device, characterized in that: It comprises a battery pack and a controller, wherein the controller is used to execute the battery status parameter prediction method as described in any one of claims 1 to 8.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a controller, enables the controller to implement the battery state parameter prediction method according to any one of claims 1 to 8 or the training method according to claims 9 to 11.

Images