Battery status prediction method, device, equipment, medium and program product

Abstract

The disclosed embodiments disclose a method, apparatus, device, medium and program product for predicting a battery state, which obtains first incremental feature data corresponding to a battery to be predicted; the first incremental feature data includes test calibration feature data, usage environment feature data, usage condition feature data, detection feature data obtained by detecting the battery to be predicted and type feature data of the battery to be predicted; based on the first incremental feature data, a battery state of the battery to be predicted in a second time period is predicted through a target state prediction model; the target state prediction model is a cross-transformation model, which is obtained by incremental transfer learning based on historical incremental feature data in a historical time period; the historical incremental feature data includes test calibration feature data, usage environment feature data, usage condition feature data, detection feature data and type feature data of at least one battery in a historical time period, thereby improving the accuracy of the battery state prediction result.

Description

Technical Field

[0001] The present disclosure relates to battery technology and artificial intelligence technology, and in particular to a method, device, equipment, medium and program product for predicting a battery state. Background Art

[0002] With the rapid development of various energy storage systems, new energy vehicles and other technologies, batteries have become an indispensable part of related technologies. Among them, the prediction of battery status such as remaining useful life (RUL) and state of health (SOH) is crucial in scenarios such as new energy vehicles and charging pile energy storage. For example, for new energy vehicles, accurate RUL prediction helps to formulate reasonable warranty policies, production plans, service strategies, etc., and can also provide an effective reference for consumers to choose cars. For another example, for charging pile energy storage, accurate RUL prediction helps to optimize charging pile maintenance strategies, improve operational efficiency, and ensure the stable operation of the power grid. In related technologies, battery RUL prediction is usually achieved based on data-driven models. Data-driven models often require a large amount of high-quality and specific battery historical data for data-driven models to learn nonlinear and complex correlations in historical data. However, due to the iteration frequency of battery technology and the current status of battery differentiation, it is difficult to obtain a large amount of historical data, especially for new batteries, resulting in low accuracy of battery RUL prediction results. Summary of the invention

[0003] Embodiments of the present disclosure provide a method, apparatus, device, medium, and program product for predicting a battery state to improve the accuracy of power load prediction results.

[0004] According to a first aspect of an embodiment of the present disclosure, a method for predicting a battery state is provided, comprising: obtaining first incremental feature data corresponding to a battery to be predicted; the first incremental feature data comprises test calibration feature data, usage environment feature data, usage condition feature data, detection feature data obtained by detecting the battery to be predicted, and type feature data of the battery to be predicted within a first time period; based on the first incremental feature data, predicting the battery state of the battery to be predicted within a second time period through a target state prediction model; the second time period is a future time period relative to the first time period; wherein the target state prediction model is a cross-transformation model; the target state prediction model is obtained by incremental transfer learning based on historical incremental feature data within a historical time period; the historical incremental feature data comprises test calibration feature data, usage environment feature data, usage condition feature data, detection feature data, and type feature data of at least one battery within the historical time period; the historical time period comprises a time period before the first time period; the test calibration feature data is determined based on data obtained from an experiment or a test environment; the usage environment feature data is determined based on data of an external environment in which the battery is located; the usage condition feature data is determined based on operating data of the battery during use; and the detection feature data is determined based on monitoring data of the battery.

[0005] According to a second aspect of the disclosed embodiment, a battery status prediction device is provided, including: an acquisition module for acquiring first incremental characteristic data corresponding to a battery to be predicted; the first incremental characteristic data includes test calibration characteristic data, usage environment characteristic data, usage condition characteristic data, detection characteristic data obtained by detecting the battery to be predicted, and type characteristic data of the battery to be predicted within a first time period; a first processing module for predicting the battery status of the battery to be predicted in a second time period based on the first incremental characteristic data through a target status prediction model; the second time period is a future time period relative to the first time period; wherein the target status prediction model It is a cross-transformation model; the target state prediction model is obtained by incremental transfer learning based on historical incremental feature data within a historical time period; the historical incremental feature data includes at least one test calibration feature data, usage environment feature data, usage condition feature data, detection feature data and type feature data of the battery within the historical time period; the historical time period includes the time period before the first time period; the test calibration feature data is determined based on data obtained from an experiment or test environment; the usage environment feature data is determined based on the external environment data of the battery; the usage condition feature data is determined based on the operating data of the battery during use; the detection feature data is determined based on the monitoring data of the battery.

[0006] According to a third aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, wherein the storage medium stores a computer program, and the computer program is used to execute the battery status prediction method described in any of the above embodiments of the present disclosure.

[0007] According to a fourth aspect of the embodiments of the present disclosure, an electronic device is provided, comprising: a processor; a memory for storing executable instructions of the processor; the processor is configured to read the executable instructions from the memory and execute the instructions to implement the battery status prediction method described in any of the above embodiments of the present disclosure; or, the electronic device comprises: the battery status prediction device provided in any of the above embodiments.

[0008] A fifth aspect of the embodiments of the present disclosure provides a computer program product. When the instructions in the computer program product are executed by a processor, the battery status prediction method provided in any of the above embodiments of the present disclosure is executed.

[0009] Based on the battery state prediction method, device, equipment, medium and program product provided by the above embodiments of the present disclosure, the target state prediction model is obtained through incremental transfer learning, and effective learning of the model can be achieved with a small number of historical samples. The historical incremental feature data based on the learning process includes battery-related test calibration feature data, usage environment feature data, usage condition feature data, detection feature data and battery type feature data. The network structure combined with the cross-transformation model is good at capturing long dependencies, global information and complex multi-scale feature associations, so that the model can effectively learn the nonlinear complex associations between the battery state and various features, which is helpful to understand the short-term patterns in the battery state changes and identify the medium- and long-term trends of battery degradation, so that the learned target state prediction model can accurately predict the battery state. Even for newly added types (i.e., new models), it is still possible to learn the association between the differentiated information of new batteries and old batteries and the state of new batteries based on a small number of historical samples of new batteries and a historical state prediction model obtained based on a large number of historical samples of old batteries through incremental transfer learning, so as to obtain a target state prediction model that can accurately predict the battery state, which can improve the efficiency of model training and the accuracy of battery state prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is an exemplary application scenario of the battery status prediction method provided by the embodiment of the present disclosure;

[0011] Figure 2 is a flowchart of a method for predicting a battery state provided by an exemplary embodiment of the present disclosure;

[0012] Figure 3 is a flowchart of a method for predicting a battery state provided by another exemplary embodiment of the present disclosure;

[0013] Figure 4 is a flowchart of a method for predicting a battery state provided by another exemplary embodiment of the present disclosure;

[0014] Figure 5 is a flowchart of obtaining a target state prediction model provided by an exemplary embodiment of the present disclosure;

[0015] Figure 6 is a flowchart of obtaining a target state prediction model provided by another exemplary embodiment of the present disclosure;

[0016] Figure 7 is a flowchart of determining the network parameter status of each encoder in a shared network provided by an exemplary embodiment of the present disclosure;

[0017] Figure 8 is a schematic diagram of battery type classification provided by an exemplary embodiment of the present disclosure;

[0018] Fig. 9 is a shared network thawing logic schematic diagram provided by an exemplary embodiment of the present disclosure;

[0019] Fig.10 is a flowchart of obtaining a target state prediction model provided by another exemplary embodiment of the present disclosure;

[0020] Fig.11 is a schematic diagram of a network structure of a state prediction model provided by an exemplary embodiment of the present disclosure;

[0021] Fig.12 is a schematic diagram of the network structure of a hybrid attention encoder provided by an exemplary embodiment of the present disclosure;

[0022] Fig.13 is a schematic diagram of the network structure of a global attention encoder provided by an exemplary embodiment of the present disclosure;

[0023] Fig.14 is an overall flow chart of model training and prediction provided by an exemplary embodiment of the present disclosure;

[0024] Fig.15 is a schematic diagram of a network structure of a state prediction model provided by another exemplary embodiment of the present disclosure;

[0025] Fig.16 is a flowchart of a method for predicting a battery state provided by another exemplary embodiment of the present disclosure;

[0026] Fig.17is a structural schematic diagram of a battery status prediction device provided by an exemplary embodiment of the present disclosure;

[0027] Fig.18 is a structural schematic diagram of a battery status prediction device provided by another exemplary embodiment of the present disclosure;

[0028] Fig.19 It is a structural diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0029] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure unless otherwise specifically stated.

[0030] Those skilled in the art can understand that the terms "first" and "second" in the embodiments of the present disclosure are only used to distinguish different steps, devices or modules, etc., and neither represent any specific technical meaning nor indicate the necessary logical order between them.

[0031] It should also be understood that in the embodiments of the present disclosure, “plurality” may refer to two or more than two, and “at least one” may refer to one, two, or more than two.

[0032] It should also be understood that any component, data or structure mentioned in the embodiments of the present disclosure can generally be understood as one or more, unless explicitly limited or otherwise indicated in the context.

[0033] In addition, the term "and / or" in the present disclosure is only a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in the present disclosure generally indicates that the associated objects before and after are in an "or" relationship.

[0034] It should also be understood that the description of the various embodiments in the present disclosure focuses on the differences between the various embodiments, and the same or similar aspects thereof can be referenced to each other, and for the sake of brevity, they will not be described one by one.

[0035] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0036] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0037] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered as part of the specification.

[0038] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0039] SUMMARY OF THE DISCLOSURE

[0040] In the process of realizing the present disclosure, the inventors found that in various energy storage systems, new energy vehicles and other related technologies, battery RUL prediction is usually implemented based on data-driven models. Data-driven models often require a large amount of high-quality and specific battery historical data for the data-driven model to learn the nonlinear and complex correlations in the historical data. However, due to the iteration frequency of battery technology and the current status of battery differentiation, especially for new batteries, the RUL prediction demand is often ahead of the data generation time, and the time scale of the battery's entire life cycle is large, making it difficult to obtain a large amount of historical data, resulting in low accuracy of the battery RUL prediction results.

[0041] Exemplary Overview

[0042] Figure 1 This is an exemplary application scenario of the battery status prediction method provided by the embodiment of the present disclosure. Figure 1As shown, the battery state prediction method provided by the embodiment of the present disclosure can be executed by the battery state prediction device 10 provided by the embodiment of the present disclosure to realize the state prediction of the battery to be predicted. Specifically, by using the battery state prediction method provided by the embodiment of the present disclosure, the first incremental feature data corresponding to the battery to be predicted can be obtained; then based on the first incremental feature data, the battery state of the battery to be predicted in the second time period is predicted by the target state prediction model. The first incremental feature data includes the test calibration feature data, the use environment feature data, the use condition feature data, the detection feature data obtained by the detection of the battery to be predicted and the type feature data of the battery to be predicted in the first time period. The second time period is a future time period relative to the first time period. The target state prediction model is a cross-transformation model. The target state prediction model is obtained by incremental transfer learning based on the historical incremental feature data in the historical time period. The historical incremental feature data includes the test calibration feature data, the use environment feature data, the use condition feature data, the detection feature data and the type feature data of at least one battery in the historical time period; the historical time period includes the time period before the first time period. The test calibration characteristic data is determined based on the data obtained in the experimental or test environment; the usage environment characteristic data is determined based on the external environment data of the battery; the usage condition characteristic data is determined based on the operating data of the battery during use; and the detection characteristic data is determined based on the monitoring data of the battery. The target state prediction model is obtained through incremental transfer learning, and effective learning of the model can be achieved with a small number of historical samples. The historical incremental feature data based on the learning process include battery-related test calibration feature data, usage environment feature data, usage condition feature data, detection feature data and battery type feature data. The network structure combined with the cross-transformation model is good at capturing long dependencies, global information and complex multi-scale feature associations, which can enable the model to effectively learn the nonlinear complex associations between battery status and various features, help understand the short-term patterns in battery state changes, and identify the medium- and long-term trends in battery degradation, so that the learned target state prediction model can accurately predict the battery status. Even for newly added types (i.e., new models), it can still be based on a small number of historical samples of new types of batteries. On the basis of the historical state prediction model obtained based on a large number of historical samples of old batteries, incremental transfer learning is used to learn the association between the differentiated information of new batteries and old batteries and the status of new batteries, thereby obtaining a target state prediction model that can accurately predict the battery status, which can improve the efficiency of model training and the accuracy of battery status prediction results.

[0043] Exemplary Methods

[0044] Figure 2FIG. 1 is a flow chart of a method for predicting a battery state provided by an exemplary embodiment of the present disclosure. The embodiment of the present disclosure can be applied to electronic devices, such as servers, terminal devices, etc. Figure 2 As shown, the method provided by the embodiment of the present disclosure may include the following steps:

[0045] Step 210, obtaining first incremental characteristic data corresponding to the battery to be predicted.

[0046] Among them, the first incremental characteristic data includes test calibration characteristic data of the battery to be predicted in the first time period, usage environment characteristic data, usage condition characteristic data, detection characteristic data obtained from the detection of the battery to be predicted, and type characteristic data of the battery to be predicted.

[0047] In some optional embodiments, the first incremental feature data may be feature data based on any time granularity. Time granularity is a basic unit for managing data. For example, if the time granularity is day (or day), the first incremental feature data may include feature data corresponding to each day in the first time period.

[0048] In some optional embodiments, the first time period may be the most recent time period in which new feature data of the battery to be predicted is generated. For example, the first time period may be the time period after the historical time period on which the target state prediction model is trained. Exemplarily, the current date is represented as day t (i.e., the tth day), and the feature data of t-1, t-2, ..., tm are used for model training, then the first time period may be day t, t and m are both positive integers, and t is greater than or equal to m. The first incremental feature data is the feature data of the battery to be predicted in the first time period.

[0049] In some optional embodiments, the first time period may include the most recent time period and part of the historical time period in which new characteristic data of the battery to be predicted is generated. For example, the first time period may include day t, and day t-1, t-2, ..., ts, where t and s are both positive integers, and t is greater than or equal to s. Optionally, s may be less than or equal to m.

[0050] Step 220 : Based on the first incremental feature data, a target state prediction model is used to predict a battery state of the battery to be predicted in a second time period.

[0051] The second time period is a future time period relative to the first time period. For example, the first time period is day t, and the second time period is day t+1. In actual applications, the lengths of the first time period and the second time period are not limited to 1 day.

[0052] In some optional embodiments of the present disclosure, the target state prediction model is a crossformer model; the target state prediction model is obtained by incremental transfer learning based on historical incremental feature data within a historical time period; the historical incremental feature data includes test calibration feature data, usage environment feature data, usage condition feature data, detection feature data and type feature data of at least one battery within a historical time period; the historical time period includes the time period before the first time period.

[0053] In some optional embodiments of the present disclosure, test calibration characteristic data is determined based on data obtained from an experimental or test environment; usage environment characteristic data is determined based on external environment data of the battery; usage condition characteristic data is determined based on operating data of the battery during use; and detection characteristic data is determined based on monitoring data of the battery.

[0054] In some optional embodiments, the historical time period includes the time period before the first time period. For example, if the first time period is day t, the historical time period includes the time period of day t-1, day t-2, ..., day tm, a total of m days.

[0055] In some optional embodiments, the historical time period and the first time period may have overlapping time. For example, the first time period is a time period of t day, t-1 day, t-2 day, ..., ts day, a total of s+1 days, and the historical time period includes a time period of t-1 day, t-2 day, ..., tm day, a total of m days, where m is greater than or equal to s.

[0056] In some optional embodiments, the battery status may include the remaining life (RUL) and health state (SOH) of the battery. In practical applications, the remaining life RUL of the battery can be carried by predicting the SOH value. For example, the predicted battery status is how long it will take for the battery to decay to 70%, what the SOH value will be after 2 years of use, etc. The specific prediction content can be set according to actual engineering needs.

[0057] In some optional embodiments, the time granularity of the historical incremental feature data within the historical time period is consistent with the time granularity of the first incremental feature data. For example, the time granularity is day.

[0058] In some optional embodiments, the Crossformer model is implemented based on a long-distance attention network (which may be called a global attention network) and a short-distance attention network (which may be called a local attention network), so as to achieve cross-scale attention, enabling the model to effectively learn long dependencies, global information, and complex multi-scale feature associations.

[0059] In some optional embodiments, incremental transfer learning means that after the basic large model is obtained through initial training, subsequent model training based on the full amount of historical data can be avoided. Instead, the previous generation model is fine-tuned or branch network training is expanded based on incremental data (or combined with partial historical data), so that the model can incrementally learn new knowledge while ensuring the validity of old knowledge.

[0060] In some optional embodiments, the test calibration feature data of each battery (e.g., a battery to be predicted, a battery for model training) may include data tested and calibrated in a laboratory or test environment, such as battery cell cycle life characteristics and cell calendar life characteristics. The usage environment feature data may include the external environment data of the battery, such as temperature, air pressure, humidity and other related information. The usage condition feature data may include the operating data of the battery during use, such as charge and discharge current, charge and discharge capacitance, charge and discharge frequency and other related information. The detection feature data may include relevant information such as the battery charging capacity, discharge capacity, battery (or battery system) temperature, etc. obtained by monitoring the battery during use. Monitoring of the battery may be achieved through sensors and / or other monitoring methods.

[0061] In some optional embodiments, the test calibration feature data may be preprocessed, and the preprocessed intermediate feature data, the use environment feature data, the use condition feature data, the detection feature data, and the type feature data together form incremental feature data (e.g., first incremental feature data, historical incremental feature data). Preprocessing may include trend statistics, data time series splitting and splicing, machine learning, neural network processing, etc., and feature mining and extraction of the test calibration feature data through preprocessing helps to learn deeper information from the test calibration feature data, so that the model can capture the test data of time series measurement.

[0062] The battery state prediction method provided by the embodiment of the present disclosure obtains the target state prediction model through incremental transfer learning, and can realize effective learning of the model with a small number of historical samples. The historical incremental feature data based on the learning process includes battery-related test calibration feature data, usage environment feature data, usage condition feature data, detection feature data and battery type feature data. The network structure combined with the cross-transformation model is good at capturing long dependencies, global information and complex multi-scale feature association relationships, so that the model can effectively learn the nonlinear complex association relationship between the battery state and various features, which is helpful to understand the short-term pattern in the battery state change and identify the medium- and long-term trend of battery degradation, so that the learned target state prediction model can accurately predict the battery state. Even for newly added types (i.e., new models), based on a small number of historical samples of the newly added types of batteries, on the basis of the historical state prediction model obtained based on a large number of historical samples of old batteries, incremental transfer learning can be used to learn the association relationship between the differentiated information of new batteries and old batteries and the state of new batteries, so as to obtain a target state prediction model that can accurately predict the battery state, which can improve the efficiency of model training and the accuracy of battery state prediction results.

[0063] Figure 3 It is a flowchart of a method for predicting a battery state provided by another exemplary embodiment of the present disclosure.

[0064] In some optional embodiments of the present disclosure, in the above Figure 2 Based on the embodiment shown, Figure 3 As shown, step 220 predicts the battery state of the battery to be predicted in the second time period based on the first incremental feature data by using the target state prediction model, which may include:

[0065] Step 2210, based on the first incremental feature data, obtain a first encoding feature through a preset number of encoders.

[0066] Among them, the preset number of encoders includes a first number of hybrid attention encoders (which can be referred to as hybrid encoders for short) and a second number of global attention encoders (which can be referred to as global encoders for short).

[0067] In some optional embodiments, the preset number can be set according to the actual requirements for model performance. For example, the preset number can be 2, 3, 4, 5, 6, 7, and so on. The sum of the first number and the second number is the preset number. For example, the first number is 3, the second number is 3, and the preset number is 6. The specific values ​​of the preset number, the first number, and the second number are not limited. The preset number of encoders are connected in series to achieve multi-level encoding of feature data. The preset number of encoders includes a global attention encoder and a hybrid attention encoder. The hybrid attention encoder may include a local window attention layer and a global window attention layer. The local window attention mechanism helps to understand the short-term patterns in battery state changes. The global attention mechanism can effectively capture long-distance dependencies and help identify medium- and long-term trends in battery degradation. The hybrid attention encoder can allow different local window attention features and global attention features to share information, which helps to enhance the model's understanding of complex interactions, and the combination of multiple attention mechanisms can effectively reduce the computational complexity of the model.

[0068] Step 2220: Based on the first coding feature, the battery state of the battery to be predicted in the second time period is obtained through the prediction head network layer.

[0069] Among them, the prediction head network layer can be a linear network layer, a convolutional network layer, a fully connected layer, etc., and the specific type of the prediction head network layer is not limited.

[0070] In the embodiments of the present disclosure, a Crossformer model constructed by a global attention encoder and a hybrid attention encoder is used as a state prediction model, because the hybrid attention encoder is implemented based on a local window attention layer and a global window attention layer. The local window attention mechanism helps to understand the short-term patterns in battery state changes. The global attention mechanism can effectively capture long-distance dependencies and help identify medium- and long-term trends in battery degradation. The hybrid attention encoder can allow different local window attention features and global attention features to share information, which helps to enhance the model's understanding of complex interactions, thereby improving the accuracy of the model's prediction results for the battery state. And the combination of multiple attention mechanisms can effectively reduce the computational complexity of the model.

[0071] In some optional embodiments of the present disclosure, a first number of hybrid attention encoders and a second number of global attention encoders are connected in series in a preset order.

[0072] The step 2210 of obtaining the first encoding feature based on the first incremental feature data by using a preset number of encoders may include:

[0073] For the i-th encoder among a preset number of encoders, based on the input data corresponding to the i-th encoder, the i-th encoding result is obtained through the i-th encoder; i=1, 2, ..., n, n is a preset number; the input data corresponding to the 1st encoder is the first incremental feature data; the input data corresponding to the j-th encoder is the j-1th encoding result of the j-1th encoder; j=2, 3, ..., n; the n-th encoding result corresponding to the n-th encoder is used as the first encoding feature.

[0074] In some optional embodiments, the preset order can be any order. For example, the preset order can be that the hybrid attention encoder is alternately connected with the global attention encoder. That is, the output of a hybrid encoder is used as the input of a global encoder, and the output of the global encoder is used as the input of another hybrid encoder, and so on. Taking 6 encoders as an example, the preset order can be expressed as hybrid encoder 1-global encoder 2-hybrid encoder 3-global encoder 4-hybrid encoder 5-global encoder 6. For another example, the preset order can be that some hybrid encoders are alternately connected with some global encoders, such as hybrid encoder 1-global encoder 2-global encoder 3-global encoder 4-hybrid encoder 5-hybrid encoder 6, or other orders.

[0075] The embodiments of the present disclosure connect the hybrid attention encoder and the global attention encoder in a certain order, so that the model can effectively capture the complex dependencies between short distances, long distances, and long and short distances, which helps the model to better understand the nonlinear relationships among various aspects of the input battery multi-source feature data, thereby improving the robustness and generalization ability of the state prediction model on the basis of improving the computational efficiency of the model.

[0076] In some optional embodiments of the present disclosure, based on any of the above embodiments, the test calibration characteristic data includes the single cell cycle life characteristics and single cell calendar life characteristics of the battery.

[0077] In some optional embodiments, the single cell cycle life characteristics include charge and discharge cycle data of the battery under different conditions. For example, charge and discharge cycle data of the battery under different temperatures and different battery capacities. For example, charge and discharge cycle data under temperature k1-battery capacity w1, charge and discharge cycle data under temperature k1-battery capacity w2, charge and discharge cycle data under temperature k2-battery capacity w1, and so on.

[0078] In some optional embodiments, the calendar life characteristics of the battery include the calendar life of the battery under different conditions. For example, the calendar life of the battery under different temperatures and different power states. Calendar life refers to the time from production to the end of the battery life when the battery is not in use. For example, the calendar life under temperature k1-power state h1, the calendar life under temperature k1-power state h2, the calendar life under temperature k2-power state h1, and so on.

[0079] In some optional embodiments of the present disclosure, the usage environment characteristic data includes the usage environment information of the battery every day in the corresponding time period. The usage environment information of each day includes at least one of the average temperature, maximum temperature difference, average air pressure, maximum air pressure difference, average relative humidity, and average relative humidity difference of the day.

[0080] In some optional embodiments of the present disclosure, the usage condition characteristic data includes the usage condition information of the battery every day in the corresponding time period. The usage condition information of each day includes at least one of the average charging current, average discharging current, average charging capacitance, average discharging capacitance, average charging start state of charge, average charging end state of charge, average discharging start state of charge, average charging end state of charge, continuous charging frequency, and continuous discharging frequency of the day.

[0081] In some optional embodiments of the present disclosure, the detection feature data includes the detection data of the battery every day in the corresponding time period. The detection data of each day includes at least one of the predicted health status of the battery corresponding to that day, the cumulative charging capacity from leaving the factory to that day, the cumulative discharge capacity, the cumulative charge and discharge capacity, the total duration, the average of the voltage difference at the end of charging on that day, the maximum temperature difference of the battery, the highest temperature, the average temperature, the lowest temperature, the proportion of time when the temperature is greater than the first quantile number, the proportion of time when the temperature is less than the first quantile number and greater than the second quantile number, and the proportion of time when the temperature is less than the second quantile number.

[0082] Among them, the health status prediction value corresponding to each day is the health status of that day predicted by the state prediction model obtained based on the learning at the historical moment before that day. The first quantile number and the second quantile number can be set according to actual needs. For example, the first quantile number is 75, and the second quantile number is 25. The temperature-related items such as the maximum temperature difference, maximum temperature, average temperature, and minimum temperature of the battery in the detection feature data are temperature-related items of the battery system obtained by battery detection, which are different from the temperature-related items in the above-mentioned usage environment feature data.

[0083] In some optional embodiments of the present disclosure, the type characteristic data includes a classification identifier of each level in a preset level classification corresponding to the battery.

[0084] The number of levels of the preset level classification can be set according to actual needs. For example, the number of levels of the preset level classification can be 2, 3, 4, ..., that is, the batteries are classified into two levels, three levels, four levels, etc. The classification identifier of each level can be set arbitrarily as long as it can uniquely identify the classification of the corresponding level.

[0085] In some optional embodiments, the preset level classification includes a primary classification of main materials, a secondary classification of battery cell differentiated design, and a tertiary classification of battery system design.

[0086] Among them, the first-level classification of main materials (which can be referred to as the first-level classification) is used to characterize the differentiated information of the main materials. Different main materials correspond to different first-level classification logos. The second-level classification of differentiated cell design (which can be referred to as the second-level classification) mainly involves the differentiated design of cells under the same main material, such as soft-pack batteries, square batteries and other batteries under different processes. Optionally, different models of batteries of the same type from one or more different manufacturers can be divided into the second-level classification, such as Model A, Model B, Model C, etc. The third-level classification of battery system design (which can be referred to as the third-level classification) mainly involves the differences in battery system design under the same second-level classification. For example, differences in series-parallel combinations, differences in software and hardware, differences in structure, etc. In practical applications, it is not limited to the above three levels.

[0087] In some optional embodiments, the classification identification of the primary classification of the main material can be determined according to the coding of the main material. The classification identification of the secondary classification of the differentiated design of the battery cell can be set according to the differentiated information of the battery cell. The classification identification of the tertiary classification of the battery system design can be set according to the differentiated information of the battery system design. For example, the type feature can be expressed as ijk, i represents the classification identification of the primary classification, j represents the classification identification of the secondary classification, and k represents the classification identification of the tertiary classification. For example, the type characteristics of a battery are 523-10-55 and 811-20-10, and the specific classification identification is not limited.

[0088] The embodiments of the present disclosure construct multi-source feature data of the battery by combining specific test calibration features, usage environment features, usage condition features, and detection features with multi-level classified battery type features for model training and battery status prediction. This allows the model to learn the complex correlation between the battery status and the refined features of the multi-source feature data, which helps to improve the model performance and thus improve the accuracy of the battery status prediction results.

[0089] Figure 4 FIG. 4 is a flow chart of a method for predicting a battery state provided by another exemplary embodiment of the present disclosure.

[0090] In some optional embodiments of the present disclosure, based on any of the above embodiments, Figure 4 As shown, step 220 predicts the battery state of the battery to be predicted in the second time period based on the first incremental feature data by using the target state prediction model, which may include:

[0091] Step 22a0: Process the test calibration feature data in the first incremental feature data based on a preset processing method to obtain processed intermediate feature data.

[0092] Among them, the preset processing method may include at least one of trend statistics, data time series splitting and splicing, machine learning, neural network processing and other processing methods. Trend statistics refers to the statistics of specified trends based on test calibration feature data, and the specified trend is, for example, the attenuation rate trend. Data time series splitting and splicing may include splicing based on capacity decay SOH value. For example, it is determined that the battery SOH value on a certain day is 80%, and the data information that decays to 80% under specified test conditions is determined according to the test calibration feature data, such as the number of cycles, test temperature, starting state of charge, and ending state of charge, as the corresponding intermediate feature data, which is used to construct the model input feature data. Both machine learning and neural network processing can use pre-trained machine learning algorithms and neural network processing models to process test calibration feature data to obtain corresponding intermediate feature data for constructing model input feature data.

[0093] Step 22b0, determining the second incremental feature data based on the intermediate feature data and the usage environment feature data, usage condition feature data, detection feature data and type feature data in the first incremental feature data.

[0094] The intermediate feature data may be fused with the use environment feature data, the use condition feature data, the detection feature data and the type feature data in a specified manner to obtain the second incremental feature data. The fusion manner includes but is not limited to concatenation, convolution, addition and the like.

[0095] In some optional embodiments, the second incremental feature data may be represented as a tensor of a preset dimension. The intermediate feature data, the use environment feature data, the use condition feature data, the detection feature data and the type feature data may form the second incremental feature data according to a preset data structure. For example, the second incremental feature data may be represented as A tensor of, T represents the time scale of the incremental characteristic data of the battery to be predicted. W represents the number of characteristic parameters. For example, the battery type feature includes 3 parameters, the intermediate features corresponding to the test calibration feature include 10 parameters, the usage environment feature includes 5 parameters, the usage condition feature includes 15 parameters, and the detection feature includes 12 parameters, then W=3+10+5+15+12=45. The specific data structure of the second incremental characteristic data is not limited to the above structure, and the embodiments of the present disclosure are not limited thereto.

[0096] Step 22c0: Based on the second incremental feature data, the battery state of the battery to be predicted in the second time period is predicted by using the target state prediction model.

[0097] In any of the above and subsequent embodiments of the present disclosure, the processed second incremental feature data may be used as the first incremental feature data for a subsequent battery state prediction process.

[0098] The embodiments of the present disclosure can perform feature mining and extraction on the test calibration feature data by further processing the test calibration feature data according to a preset processing method, which helps to learn deeper information from the test calibration feature data, enables the model to capture the test data of timing measurement, and further improves the generalization ability and robustness of the model.

[0099] Figure 5 It is a flowchart of obtaining a target state prediction model provided by an exemplary embodiment of the present disclosure.

[0100] In some optional embodiments of the present disclosure, based on any of the above embodiments, Figure 5 As shown, the method of the embodiment of the present disclosure may also include the following operations of obtaining a target state prediction model:

[0101] Step 310, obtaining historical incremental feature data within a historical time period.

[0102] The historical incremental feature data of the historical time period may include incremental feature data of one or more batteries. The historical incremental feature data may be determined based on storing multi-source historical battery data. The multi-source historical battery data includes data obtained in an experimental or test environment, data of the external environment in which the battery is located, operating data of the battery during use, and monitoring data of the battery.

[0103] Step 320: determine the state label data corresponding to the historical incremental feature data.

[0104] The state label data corresponding to the historical incremental feature data can be determined by the real state of the historical time and the misaligned correspondence of the historical incremental feature data. For example, the feature data of day t-1 corresponds to the state label of day t, the feature data of day t-2 corresponds to the state label of day t-1, and so on.

[0105] Step 330: Based on the historical incremental feature data and the state label data, incremental transfer learning is performed on the historical state prediction model to obtain a target state prediction model.

[0106] The historical state prediction model includes a shared network learned at a historical moment and at least one historical branch network. The historical state prediction model may be a model used for historical prediction. For example, the historical state prediction model may be a model used for a previous prediction.

[0107] In some optional embodiments, based on the state prediction basic large model (which may be called the first-generation state prediction model) obtained by the first training, incremental transfer learning can be performed on the basic large model through incremental data to obtain a second-generation state prediction model. Then, based on the second-generation state prediction model, incremental transfer learning is performed on the second-generation state prediction model through incremental data to obtain a third-generation state prediction model, and so on, and each generation of state prediction models can be stored. The historical state prediction model is the state prediction model of the most recent generation. The training of the basic large model can construct training samples through a large amount of existing historical stock data, and the constructed initial state prediction model is trained to obtain the basic large model. The basic large model training can be trained only once, and subsequent model iterations are continuously and dynamically updated with incremental data.

[0108] The embodiments of the present disclosure perform incremental transfer learning on the historical state prediction model learned at historical moments through the historical incremental feature data and the corresponding state label data within the historical time period to obtain a target state prediction model. This allows the model to quickly learn new battery knowledge based on the old battery knowledge that has been learned, thereby improving the accuracy of the state prediction results while ensuring learning efficiency.

[0109] Figure 6 It is a flowchart of obtaining a target state prediction model provided by another exemplary embodiment of the present disclosure.

[0110] In some optional embodiments of the present disclosure, step 330 of performing incremental transfer learning on the historical state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model may include:

[0111] Step 3310: Determine whether there is a newly added type based on the type feature data in the historical incremental feature data.

[0112] Among them, type feature data can be extracted from historical incremental feature data, matched with historical types in historical stock feature data, and whether there is a new type is determined based on the matching result. Alternatively, the learned historical types can be maintained in real time, and the type feature data in the historical incremental feature data can be matched with the historical types to determine whether there is a new type.

[0113] Step 3320, in response to the existence of a new type, determine the new branch network based on the new type and the historical state prediction model, and determine the initial parameters of the new branch network from the historical branch network of the historical state prediction model, and obtain the current state prediction model based on the historical state prediction model and the new branch network.

[0114] Among them, according to the newly added type and the historical type corresponding to each historical branch network, the target historical branch network corresponding to the target historical type closest to the newly added type can be determined from the historical branch networks, and the network parameters of the target historical branch network are used as the initial parameters of the newly added branch network. The network structures of each branch network are the same. In other words, the target historical branch network is copied to the newly added branch network.

[0115] In some optional embodiments, the current state prediction model may include a shared network, a historical branch network, and a newly added branch network of the historical state prediction model. The shared network is used to capture the shared features of different types of batteries, and different branch networks are used to capture the differentiated features (or task-specific features) of different types of batteries, and predict the battery state of each type based on the differentiated features of each type of battery.

[0116] Step 3330, based on the historical incremental feature data and state label data, the current state prediction model is trained to obtain the target state prediction model.

[0117] Among them, one or more groups of training samples and the state labels corresponding to each training sample can be determined according to preset rules based on historical incremental feature data, and each training sample can be processed using the current state prediction model to obtain the state prediction value corresponding to each training sample; based on the state prediction value and state label corresponding to each training sample, the model loss is determined; based on the model loss, the network parameters are updated to obtain an updated state prediction model. If the updated state prediction model meets the training end condition, the updated state prediction model is used as the target state prediction model. If the updated state prediction model does not meet the training end condition, the updated state prediction model is used as the current state prediction model, and iterative learning continues until the training end condition is met to obtain the target state prediction model.

[0118] In some optional embodiments, the learning rates of the shared network, each historical branch network and the newly added branch network can be determined according to whether the historical incremental feature data includes feature data of other historical types other than the newly added type, and the differentiation of the newly added type from the learned historical types. For example, for the historical branch network corresponding to the historical type where the historical incremental feature data does not exist, there is no need to adjust the network parameters of the historical branch network. For the historical branch network of the historical type that exists in the historical incremental feature data, a smaller learning rate is used to adjust the network parameters, and a larger learning rate is used to adjust the network parameters for the newly added branch network. The shared network may include an encoder corresponding to each level in the feature type, and the encoders of each level of the shared network are frozen or unfrozen according to the degree of differentiation between the newly added type and the historical type. For the frozen encoder, a smaller learning rate is used to adjust the network parameters, and for the unfrozen encoder, a larger learning rate is used to adjust the network parameters. For example, in the case where the main material in the newly added type is the same as the main material of the learned historical type, the encoder corresponding to the primary classification of the main material in the shared network is frozen, and the network parameters of the encoder are fine-tuned at a smaller learning rate. If the main material in the newly added type is different from the main material of the learned historical type, the encoders corresponding to the three levels are all unfrozen and the network parameters are adjusted with a larger learning rate.

[0119] In some optional embodiments, the current state prediction model can be trained in a K-fold traversal manner (K ​​is a positive integer) to obtain a target state prediction model. Specifically, the historical incremental feature data and the corresponding state label data can be divided into K equal parts, and K-1 parts are used as training sample sets each time, and the remaining 1 part is used as a test set. This is repeated K times, and the average of the K results is taken as the performance evaluation result of the model. Whether to end the training is determined based on the performance evaluation result. For example, using the mean absolute error (MAE) as the loss function, if the MAE values ​​of K times are all less than the preset threshold, it is determined that the model has converged and the training is ended.

[0120] In some optional embodiments, step 330 of performing incremental transfer learning on the historical state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model may also include:

[0121] Step 3340, in response to the absence of a newly added type, fine-tune the historical state prediction model based on the historical incremental feature data and the state label data to obtain a target state prediction model.

[0122] Among them, if there is no new type, it means that the historical incremental feature data is the incremental feature data of the historical type of battery that the model has learned, and the historical branch network corresponding to the corresponding type of battery already exists in the historical state prediction model. This learning can be based on the historical incremental feature data and state label data. Fine-tune the historical state prediction model to further improve the accuracy of the model's state prediction results for this type of battery. The principle of fine-tuning is similar to the learning process of the above-mentioned new branch network, which will not be repeated here.

[0123] In some optional embodiments, corresponding to the first incremental feature data, the test calibration feature data in the historical incremental feature data can be processed based on a preset processing method to obtain processed intermediate feature data, and the target historical incremental feature data is constructed based on the intermediate feature data and the usage environment feature data, usage condition feature data, detection feature data and type feature data in the historical incremental feature data, which is used for incremental transfer learning to obtain a target state prediction model. The specific processing process can be referred to the aforementioned embodiments.

[0124] In some optional embodiments, a K (K is a positive integer) fold traversal method may be used to fine-tune the historical state prediction model to obtain a target state prediction model. For the specific principle, refer to the above embodiment.

[0125] In the embodiments of the present disclosure, when there is a new type, a new branch network is added on the basis of the historical state prediction model, and the network parameters of the historical type branch network closest to the new type are copied as the initial parameters of the new branch network. This allows the model to learn the differential information of the feature data of the new type and the historical type based on the learned old battery knowledge, thereby quickly learning the correlation between the battery state of the new type and the incremental feature data, and obtaining an effective target state prediction model with fewer training samples, thereby ensuring the accuracy of the battery state prediction results for the new type.

[0126] In some optional embodiments of the present disclosure, step 3330 of training the current state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model may include:

[0127] In response to the historical incremental feature data only including feature data of the newly added type, based on the historical incremental feature data and the state label data, the network parameters of the shared network and the newly added branch network in the current state prediction model are adjusted to obtain a target state prediction model. In response to the historical incremental feature data including feature data of the newly added type and feature data of the historical type, based on the historical incremental feature data and the state label data, the parameters of the shared network, the historical branch network and the newly added branch network are adjusted to obtain a target state prediction model; wherein, the network parameters of the historical branch network are adjusted based on a first learning rate, and the network parameters of the newly added branch network are adjusted based on a second learning rate; the second learning rate is greater than the first learning rate; each historical type corresponds to a historical branch network; the historical type refers to the battery type that has been learned by the historical state prediction model.

[0128] Among them, if the historical incremental feature data only includes feature data of the newly added type and does not include feature data of the historical type, the network parameters of the historical branch network do not need to be adjusted, and only the network parameters of the shared network and the newly added branch network need to be adjusted. The learning rate of each encoder of the shared network can be determined according to the degree of differentiation between the newly added type and the learned historical type, and the newly added branch network uses a larger learning rate (which can be a preset value) to adjust the network parameters.

[0129] If the historical incremental feature data includes both feature data of the newly added type and feature data of one or more historical types, it is necessary to fine-tune the network parameters of the corresponding historical branch network (i.e., fine-tune with a smaller first learning rate) and adjust the network parameters of the newly added branch network (i.e., adjust with a second learning rate greater than the first learning rate). The learning rate of each encoder of the shared network is determined according to the degree of differentiation between the newly added type and the learned historical type.

[0130] Exemplarily, the historical state prediction model is obtained by training based on the historical feature data of the historical type 523-10-55. The new type of 811-20-10 appears in the historical incremental feature data this time. The battery of this new type is different from the historical type. In this case, a new branch network is added for the new type during model training. The new branch network is adjusted based on the historical incremental feature data of 811-20-10 with a larger learning rate. If only feature data of the new type 811-20-10 exists in the historical incremental feature data, the network parameters of the historical branch network of the historical type 523-10-55 are not adjusted. If the historical incremental feature data also includes feature data of the historical type 523-10-55 on the basis of the feature data of 811-20-10, the historical branch network corresponding to the historical type fine-tunes the network parameters with a smaller first learning rate, and the new branch network adjusts the network parameters with a larger second learning rate.

[0131] The embodiments of the present disclosure expand independent branch networks according to the differences in battery types in incremental data, and use different learning rates to adjust network parameters for the newly added branch networks and historical branch networks. This improves the accuracy of prediction results for newly added types of batteries while ensuring the accuracy of state prediction for historical types of batteries.

[0132] In some optional embodiments of the present disclosure, the shared network includes a third number of encoders; the type feature data includes a classification identifier for each level in a preset level classification; and the third number is the number of levels in the preset level classification. That is, the third number can be set according to the number of levels of the battery classification. Optionally, the third number of encoders can include at least one hybrid attention encoder and / or at least one global attention encoder.

[0133] In some optional embodiments, step 3330 of training the current state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model may include:

[0134] According to the classification identifier of each level of the newly added type, determine whether the network parameters of each encoder in the shared network are frozen; the frozen state includes two states: frozen and unfrozen; according to the state of whether the network parameters of each encoder are frozen, determine the learning rate corresponding to each encoder; the learning rate corresponding to the frozen encoder is the third learning rate, and the learning rate corresponding to the unfrozen encoder is the fourth learning rate; the third learning rate is less than the fourth learning rate; based on the historical incremental feature data and the state label data, according to the learning rate corresponding to each encoder, train the current state prediction model to obtain the target state prediction model.

[0135] Among them, the network parameters of each encoder in the shared network can be determined according to the preset logic. For example, first determine whether the classification identifier of the primary classification of the main material in the newly added type (which can be referred to as the primary classification identifier or the primary identifier) ​​is the same as the learned historical type, that is, determine whether there is a historical type in the historical type that is the same as the main material of the newly added type. If the primary classification identifier is not the same, that is, the model sharing network has not learned the relevant knowledge of the battery of the newly added type at all, then all levels of encoders in the shared network need to be unfrozen. If the primary classification identifier is the same, the encoder corresponding to the primary classification in the shared network is frozen, that is, the model has learned the knowledge of the main material, and the corresponding encoder can be fine-tuned. In this case, it is necessary to continue to determine whether the classification identifier of the secondary classification (which can be referred to as the secondary classification identifier or the secondary identifier) ​​is the same as the historical type. If not, the encoder corresponding to the secondary classification and the encoder corresponding to the third-level classification need to be unfrozen. If the secondary classification identifier is the same as the historical type, the encoder corresponding to the secondary classification is frozen. And so on.

[0136] In the embodiments of the present disclosure, the network parameters of the shared network are adjusted at different learning rates as needed according to the degree of differentiation between the newly added types and the historical types. Combined with the expansion of the independent branch network, it is possible to avoid excessive adjustment of the network parameters of the shared network during the incremental transfer learning process, thereby ensuring the prediction accuracy of the model for the original historical type batteries, and improving the high generalization and prediction accuracy of the model after transfer learning for multiple types of batteries and the entire life cycle.

[0137] Figure 7 It is a flowchart of determining the network parameter status of each encoder in a shared network provided by an exemplary embodiment of the present disclosure.

[0138] In some optional embodiments of the present disclosure, based on any of the above embodiments, the preset level classification includes primary classification of main materials, secondary classification of differentiated cell design, and tertiary classification of battery system design. The shared network includes a first encoder corresponding to the primary classification of main materials, a second encoder corresponding to the secondary classification of differentiated cell design, and a third encoder corresponding to the tertiary classification of battery system design.

[0139] According to the classification identifier of each level of the newly added type, determining whether the network parameters of each encoder in the shared network are frozen may include:

[0140] Step 410 , in response to the historical classifications learned by the historical state prediction model not including the primary classification identifier of the newly added type, determining that the states of the first encoder, the second encoder, and the third encoder are all unfrozen.

[0141] Among them, if the learned historical classification does not include the first-level classification identifier of the newly added type, it means that the shared network has not yet learned the shared features of the newly added type of battery and the historical type of battery. Therefore, it is necessary to unfreeze the encoders corresponding to each level in the shared network and adjust the network parameters with a larger learning rate.

[0142] Step 420: In response to the historical classification including a primary classification identifier of a newly added type, determining that the state of the network parameter of the first encoder is frozen.

[0143] Among them, if the historical classification includes a newly added type of first-level classification identifier, it means that the shared network has learned the shared features related to the first-level classification. The first encoder corresponding to the first-level classification can be frozen and the network parameters can be fine-tuned at a smaller learning rate.

[0144] Step 430: In a case where the historical classification includes the primary classification identifier of the newly added type, in response to the historical classification not including the secondary classification identifier of the newly added type, determining that the network parameters of the second encoder and the third encoder are unfrozen.

[0145] Among them, if the historical classification does not include the secondary classification identifier of the newly added type, it means that although the shared network has learned the shared features related to the first-level classification of the newly added type, it has not learned the shared features related to the second-level classification of the newly added type. The second encoder corresponding to the secondary classification and the third encoder corresponding to the tertiary classification need to be unfrozen and the network parameters need to be adjusted with a larger learning rate.

[0146] Step 440: In response to the historical classification including a secondary classification identifier of a newly added type, determining that the state of the network parameter of the second encoder is frozen.

[0147] Among them, if the historical classification includes a secondary classification mark of a newly added type, it means that the shared network has learned the shared features related to the main material and battery cell design of the newly added type, and the second encoder corresponding to the secondary classification can fine-tune the network parameters at a smaller learning rate.

[0148] Step 450: In a case where the historical classification includes the secondary classification identifier of the newly added type, in response to the historical classification not including the tertiary classification identifier of the newly added type, determining that the state of the network parameter of the third encoder is unfrozen.

[0149] The principle of determining the state of the third encoder is similar to that of the aforementioned second encoder.

[0150] The embodiments of the present disclosure, by splitting the network architecture of the model into a shared network and a specific task branch network, adjust the network parameters of the shared network on demand, in an orderly and gradual manner based on the differences in battery types in the incremental data, and expand the independent branch network, can avoid excessive adjustment of the network parameters of the shared network during the incremental transfer learning process, thereby ensuring the model's prediction accuracy for the original old batteries, while improving the model's high generalization and prediction accuracy for various types of batteries and the entire life cycle.

[0151] In some optional embodiments, Figure 8 FIG. 1 is a schematic diagram of battery type classification provided by an exemplary embodiment of the present disclosure. Figure 8As shown, the battery classification takes three levels as an example. The first-level classification of main materials may include M types. The second-level classification of differentiated cell design under each first-level classification of main materials may include multiple types. The figure only shows the second-level classification corresponding to main material type 1, including S types. Taking cell design type 1 as an example, the third-level classification of battery system design corresponding to cell design type 1 includes L battery system types. M, S, and L are all positive integers. The type characteristics of the battery are expressed as ijk according to the battery classification setting, where i corresponds to the classification identification of the first-level classification of the main material of the battery, j corresponds to the classification identification of the second-level classification of the differentiated cell design of the battery, and k corresponds to the classification identification of the third-level classification of the system design of the battery. In actual applications, the classification method of batteries and the representation of type characteristics are not limited to the methods in the figure. This is only an exemplary description and is not a limitation of the embodiments of the present disclosure.

[0152] In some optional embodiments, Fig. 9 FIG. 1 is a schematic diagram of a shared network unfreezing logic provided by an exemplary embodiment of the present disclosure. Fig. 9 As shown, for the newly added type, if the main material is different from the learned historical type, the first encoder, the second encoder, and the third encoder are unfrozen. If the main material is the same as the historical type, the first encoder remains frozen, and the battery cell design is further determined. If the battery cell design is different from the historical type, the second encoder and the third encoder are unfrozen. If the battery cell design is the same as the historical type, the second encoder remains frozen. The battery system design is further determined. If the battery system design is different from the historical type, the third encoder is unfrozen. If the battery system design is the same as the historical type, it means that there is no newly added type. The first encoder, the second encoder, and the third encoder are all frozen.

[0153] In some optional embodiments of the present disclosure, each branch network includes a fourth number of encoders; the fourth number of encoders includes at least one hybrid attention encoder and / or at least one global attention encoder.

[0154] The fourth number can be set to any number according to actual needs. For example, the fourth number can be 1, 2, 3, 4, etc.

[0155] In some optional embodiments, the fourth number of encoders may include a fifth number of mixed attention encoders and a sixth number of global attention encoders. The fourth number is the sum of the fifth number and the sixth number.

[0156] In the embodiment of the present disclosure, a specific task branch network corresponding to a battery type is formed by a hybrid attention encoder and a global attention encoder, so that the model can capture the long-distance dependency and short-distance dependency of a specific task through the hybrid attention encoder, and allow information sharing between long-distance attention features and short-distance attention features, which helps to enhance the model's understanding of the complex interactions of specific tasks, effectively capture the long-distance dependency of specific tasks through the global attention mechanism, and help identify the medium- and long-term trends of degradation of specific battery types. Thereby, the accuracy of the model's prediction results for a specific type of battery state can be improved.

[0157] Fig.10 It is a flowchart of obtaining a target state prediction model provided by another exemplary embodiment of the present disclosure.

[0158] In some optional embodiments of the present disclosure, based on any of the above embodiments, Fig.10 As shown, step 3330 trains the current state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model, including:

[0159] Step 33310, based on the historical incremental feature data and state label data, determine the training sample data and the corresponding state label data.

[0160] In some optional embodiments, the historical incremental feature data and state label data can be divided into K equal parts based on a K-fold traversal method, and K-1 parts are used as training sample sets each time. The training sample set includes the training sample data and the corresponding state label data.

[0161] Step 33320, based on the training sample data, obtain the state prediction result corresponding to the training sample data through the current state prediction model.

[0162] The training sample data may be used as input data of the current state prediction model, and the state prediction result corresponding to the training sample data may be obtained through prediction by the current state prediction model.

[0163] Step 33330, determine the first loss based on the state prediction result and the state label data.

[0164] The first loss may be determined based on a preconfigured loss function, which may be, for example, one or more of a mean absolute error loss function, a mean square error loss function, and other loss functions.

[0165] Step 33340, determine the comprehensive loss based on the first loss, the second loss corresponding to the historical state prediction model, and the pre-configured loss weight.

[0166] Among them, the second loss corresponding to the historical state prediction model is the loss when training the historical branch network.

[0167] In some optional embodiments, the first loss and the second loss may be fused according to preconfigured weights to obtain a comprehensive loss (also referred to as a multi-task combined loss).

[0168] Exemplarily, the first loss is represented by L1, the historical state prediction model is the S-generation model, and the corresponding branch network training loss function is represented by L2. Then the comprehensive loss L can be expressed as follows:

[0169]

[0170] Where a and b are pre-configured weights.

[0171] In some optional embodiments, the first loss and the second loss may be fused based on preconfigured weights and regularization terms to obtain a comprehensive loss.

[0172] For example, the comprehensive loss can be expressed as follows:

[0173]

[0174] Among them, a and b are pre-configured weights, c represents the weight of the regularization term, Representation and network parameters Related regularization terms.

[0175] Step 33350, based on the comprehensive loss, update the network parameters of the current state prediction model to obtain an updated state prediction model.

[0176] Wherein, based on the comprehensive loss, a pre-configured optimization algorithm can be used to update the network parameters of the model to obtain an updated state prediction model. Wherein, the learning rate used by each part when the network parameters are updated can be obtained according to the determination method of the above embodiment.

[0177] In some optional embodiments, if the model cannot automatically reach the specified target, the third encoder can be unfrozen and tried to be trained one by one during the incremental transfer learning process to ensure the accuracy and generalization ability of the model. For example, the third encoder is unfrozen first, the first encoder and the second encoder remain frozen, and training is performed based on the training sample set to determine whether the model can reach the specified target. If the specified target cannot be reached, the second encoder is unfrozen again, the first encoder remains frozen, and training is performed again, and so on, until the specified target is reached and the target state prediction model is obtained.

[0178] Step 33360, determine the target state prediction model based on the updated state prediction model.

[0179] If the updated state prediction model meets the training end condition, the updated state prediction model is used as the target state prediction model. Otherwise, the updated state prediction model is used as the current state prediction model, and the above steps 33310 to 33360 are repeated until the target state prediction model is obtained.

[0180] In the embodiments of the present disclosure, a multi-task combination loss function is used to update network parameters, so that in the process of incremental transfer learning, the performance index of the newly trained model can be ensured to be no lower than that of the previous generation model.

[0181] In some optional embodiments of the present disclosure, the target state prediction model includes a first number of hybrid attention encoders and a second number of global attention encoders.

[0182] In some optional embodiments, each hybrid attention encoder includes at least one local window attention layer, a global window attention layer, and a fusion network layer.

[0183] The number of local window attention layers can be set to any number, for example, the number of local window attention layers can be 1, 2, 3, ... The fusion network layer can be a splicing layer or a fusion layer in other ways.

[0184] In some optional embodiments, each local window attention layer is used to perform a local window attention operation on the first input feature corresponding to the layer to obtain a local attention result corresponding to the local window attention layer; the first input feature is the output feature of the previous network layer.

[0185] The previous network layer may be, for example, an Add&Norm layer.

[0186] In some optional embodiments, each hybrid attention encoder may further include other related network layers after the fusion network layer. The other related network layers may include, for example, an addition and normalization (Add&Norm) layer, a feedforward neural network (FFN) layer, etc.

[0187] In some optional embodiments, the global window attention layer is used to perform a global window attention operation on the first input feature to obtain a first global attention result.

[0188] In some optional embodiments, the fusion network layer is used to fuse each local attention result with the first global attention result to obtain a first fusion feature, and the first fusion feature is used to determine the encoding result of the hybrid attention encoder.

[0189] In some optional embodiments, each global attention encoder includes a global window attention layer, which is used to perform a global window attention operation on a second input feature corresponding to the layer to obtain a second global attention result; the second global attention result is used to determine the encoding result corresponding to the global attention encoder.

[0190] Among them, each global attention encoder can also include other related network layers, such as Add&Norm layer, FFN layer, etc., and the specific structure is not limited.

[0191] In the embodiments of the present disclosure, a hybrid attention encoder is formed by a local window attention layer and a global window attention layer, wherein the local window attention mechanism helps to understand the short-term patterns in battery state changes. The global window attention mechanism can effectively capture long-distance dependencies and help identify medium- and long-term trends in battery degradation. The hybrid attention encoder can allow different local window attention features and global attention features to share information, which helps to enhance the model's understanding of complex interactions, thereby improving the accuracy of the model's prediction results for the battery state. And the combination of multiple attention mechanisms can effectively reduce the computational complexity of the model.

[0192] In some optional embodiments, Fig.11 Schematic diagram of the network structure of a state prediction model provided by an exemplary embodiment of the present disclosure. Fig.11 As shown, the input feature data can be feature data in the training phase (e.g., training sample data) or feature data in the prediction phase (e.g., first incremental feature data or second incremental feature data). In the figure, 6 encoders are taken as an example. The hybrid attention encoder and the global attention encoder are alternately connected. The global attention encoder-6 outputs the encoding feature (e.g., the first encoding feature in the prediction phase) and obtains the output result through the prediction head network layer. In the training result, the output result is the state prediction result corresponding to the training sample data, which is used to compare with the state label data to determine the loss. In the prediction phase, the output result is the predicted battery state of the battery to be predicted in the second time period.

[0193] In some optional embodiments, Fig.12 is a schematic diagram of the network structure of a hybrid attention encoder provided by an exemplary embodiment of the present disclosure. Fig.12As shown, the hybrid attention encoder takes two local window attention layers and one global window attention layer as an example. The input data X represents the feature data input to the hybrid attention encoder, which can be the initial feature data (such as the training sample data in the training phase, the first incremental feature data in the prediction phase, the second incremental feature data, etc.). The input data X is processed by the Add&Norm layer to obtain normalized feature data. The first local window attention layer divides the normalized feature data into M local windows, represented as x1, x2, ..., xM, and the second local window attention layer divides the normalized feature data into N local windows, represented as x1, x2, ..., xN. The attention operation (i.e., local window attention operation) is performed on each local window to obtain the local attention results Z1i and Z2j corresponding to each local window, i=1, 2, ..., M, j=1, 2, ..., N. The global window attention layer performs a global window attention operation on the normalized feature data to obtain a global attention result (i.e., the first global attention result) Z3. The feature fusion layer (the concatenation layer (which can be expressed as Concatenate-Z) is used as an example in the figure) to concatenate and fuse the local attention results corresponding to each local window and the first global attention results to obtain the fused feature (i.e., the first fused feature) Z. By fusing the attention results of multiple windows, a comprehensive feature representation containing local and global information can be obtained, which helps to capture more comprehensive information. The first fused feature Z is added and normalized with the normalized feature data of the first Add&Norm layer (which can be called the first normalized feature data) through the second Add&Norm layer to obtain the second normalized feature data. The second normalized feature data is then processed through the feedforward neural network (FFN) layer to obtain intermediate data. The fused second normalized feature data can be processed at one time through FFN, which can reduce the computational complexity. The intermediate data is then added and normalized with the first normalized feature data to obtain the third normalized feature data as the encoding result of the hybrid attention encoder.

[0194] In some optional embodiments, Fig.13 is a schematic diagram of the network structure of a global attention encoder provided by an exemplary embodiment of the present disclosure. Fig.13As shown, the global attention encoder includes an addition and normalization (Add&Norm) layer, a global window attention layer, a feature fusion layer (for example, a concatenation layer Concatenate-Z), an addition and normalization (Add&Norm) layer, a feedforward neural network (FFN) layer, and an addition and normalization (Add&Norm) layer in the direction of data flow. The difference between the global attention encoder and the above-mentioned hybrid attention encoder is that the global attention encoder does not include the above-mentioned local window attention layer. The functions of each network layer in the global attention encoder refer to the above-mentioned hybrid attention encoder and are not repeated here.

[0195] By combining the horizontal and vertical encoder types of the hybrid attention encoder and the global attention encoder, the complex dependencies between short distance and long distance, as well as between long and short distance can be effectively captured, which helps the model to better understand the nonlinear relationships in various aspects of the input battery multi-source data, and can improve the robustness and generalization ability of the model while improving computational efficiency.

[0196] In some optional embodiments, Fig.14 FIG. 1 is a flowchart of the overall process of model training and prediction provided by an exemplary embodiment of the present disclosure. Fig.14 As shown, the initial state prediction model can be obtained based on the incremental migration cross-transformation (Crossformer) architecture, and the initial state prediction model can be trained based on the historical stock battery multi-source data to obtain a basic large model as the old state prediction model (i.e., the historical state prediction model) for subsequent training. In the subsequent model training, the model can be dynamically adjusted through incremental migration learning based on the basic large model. In the figure, the incremental feature data of day t-1 is taken as an example. If the incremental feature data of day t-1 does not contain the feature data of the newly added type, the "parameter adjustment" migration learning of the t-1 model is performed on the basis of the old version of the state prediction model, that is, no new branch network is added, and only the parameters of the historical branch network are adjusted to obtain the t-1 version of the state prediction model. If the incremental feature data of day t-1 contains the feature data of the newly added type, the "parameter adjustment + expansion branch network" migration learning of the t-1 model is performed on the basis of the old version of the state prediction model, that is, an independent new branch network is added and the network parameters are adjusted to obtain the t-1 version of the state prediction model. The t-1 version of the state prediction model is deployed to the engineering side, and the battery state prediction on the engineering side is performed based on the t-day characteristic data of the battery to be predicted, and the state prediction result on day t+1 is obtained. Optionally, the state prediction result on day t+1 may include the t+1 short-term battery state, the t+1 long-term state trend, etc.

[0197] Optionally, the incremental feature data on day t-1 can be concatenated with some historical stock feature data (such as t-2, t-3, ...) as the incremental feature data on day t-1 for incremental transfer learning of the model.

[0198] Optionally, the characteristic data on day t may be concatenated with part of the historical characteristic data of the battery to be predicted, and the concatenated characteristic data may be used as the characteristic data on day t to predict the battery status at t+1.

[0199] Optionally, for t+1 short-term battery status prediction, the model can be deployed and used by daily and scheduled scheduling to generate the t+1 day predicted SOH value of the new energy or energy storage system, which can be used for abnormal monitoring on the business side. For t+1 long-term status trend (i.e., long-time scale trend) prediction, the cloud-based "cycle-increment" offline calculation can also be used on a daily and scheduled basis, and the input samples are the corresponding long-time scale fitting data set and long-time scale test calibration data set. The fitting data set refers to generating future feature parameters based on experience, manual prediction, etc., and constructing the corresponding full-life feature set to achieve long-time scale trend prediction for the entire life cycle.

[0200] In some optional embodiments, Fig.15 FIG. 1 is a schematic diagram of a network structure of a state prediction model provided by another exemplary embodiment of the present disclosure. Fig.15 As shown, the shared network takes three encoders as an example, including global attention encoder-1, hybrid attention encoder-2 and global attention encoder-3. Each branch network (only 2 branch networks are shown in the figure) takes three encoders as an example, including hybrid attention encoder-4, global attention encoder-5 and hybrid attention encoder-6. The basic model branch A is a branch network of the old version of the battery type of the basic large model (i.e., the old version of the state prediction model) obtained by initial training. The incremental migration model branch B is a newly added branch network obtained by incremental migration learning. In practical applications, multiple incremental migration model branches can be obtained through multiple incremental migration learning based on the basic model branch A. The latest added branch network is the newly added branch network, and the previous branch networks are all called historical branch networks (or old version branch networks).

[0201] In practical applications, the shared network is not limited to the network structure in the figure. For example, the shared network may include a hybrid attention encoder-1, a global attention encoder-2, and a hybrid attention encoder-3.

[0202] Fig.16 It is a flowchart of a method for predicting a battery state provided by yet another exemplary embodiment of the present disclosure.

[0203] In some optional embodiments of the present disclosure, predicting the battery state of the battery to be predicted in the second time period by using the target state prediction model based on the first incremental feature data in step 220 may include:

[0204] Step 2201, obtaining historical characteristic data corresponding to the battery to be predicted.

[0205] The historical characteristic data corresponding to the battery to be predicted refers to the characteristic data of all or part of the historical time periods before the first time period. For example, the first time period is day t, and the historical characteristic data includes the characteristic data of day t-1, day t-2, ..., day ts.

[0206] Step 2202: Determine current feature data based on the first incremental feature data and the historical feature data.

[0207] The first incremental feature data and the historical feature data may be fused to obtain the current feature data. The fusion method may include concatenation, convolution, etc.

[0208] Step 2203: Based on the current feature data, the battery state of the battery to be predicted in the second time period is predicted by using the target state prediction model.

[0209] The specific principle of step 2203 is similar to the prediction principle based on the first incremental feature data or the second incremental feature data in the aforementioned embodiment, and is not elaborated here.

[0210] The embodiments of the present disclosure fuse incremental feature data with historical feature data for predicting future battery status, so that the model can refer to the characteristics of a certain time series of the battery to be predicted, and combine the relationship between the battery status and the long time series (or long distance) learned by the model, thereby improving the prediction accuracy of the battery status.

[0211] The battery status prediction method provided by the embodiments of the present disclosure can be applied to the status prediction of batteries or battery packs used in various energy storage systems and new energy vehicles, especially for the transition stage from old batteries to new batteries. It can effectively use existing data for model training and quickly adapt to the characteristics of new batteries through a small amount of new data, thereby accurately predicting the battery status.

[0212] In addition to data-driven models, related technologies also include mechanism models, equivalent circuit models, empirical models, etc. Among them, the mechanism model is mainly constructed based on the detailed physical and chemical equations of the internal chemical reaction process of the battery, involving knowledge of electrochemical reactions, material transfer, thermodynamics and kinetics, etc. It requires a deep understanding of electrochemical reactions and the construction of simulation models that can truly reflect complex chemical mechanisms. It requires a large number of design, material and other characteristic parameters, and is difficult to implement. The equivalent circuit model simulates the electrical behavior of the battery by using circuit elements, but the equivalent circuit cannot solve the problem of RUL prediction on the engineering side caused by the complex nonlinear performance characteristics of the battery system, the multi-factor dependence of performance and the diversity of applications. Empirical models are usually only valid under specific conditions and are easily affected by changes in the environment, working conditions, etc., resulting in low accuracy of the prediction results.

[0213] Due to the rapid iteration of battery technology, especially for newly launched battery systems, it is often difficult to fully cover the complex working conditions and environments of the entire life cycle and the entire time and space domain due to laboratory testing and calibration data. Through semi-empirical simulation, it is difficult to make differentiated and refined predictions of the subsequent battery attenuation rate and trend; at the same time, there is no accumulation of a large amount of user or operator data on battery usage, which makes it impossible to effectively predict future unknown trend changes. The method of the embodiment of the present disclosure fully utilizes the existing battery big data to train the basic large model, and then performs incremental transfer learning based on various new battery performance parameters and data to establish a state prediction large model that can be applied to all or most batteries. It is of great significance to the car companies, energy storage companies and corresponding operating companies corresponding to the research and development and operation of battery systems. Especially when the battery system has a small amount of accumulated battery data, it is still possible to accurately predict the battery status based on the training results of the transfer learning model. Due to the obvious timing dependence and long time scale characteristics of battery state decay, the SOH value of each stage is affected by historical usage, aging process and current performance status, and has typical nonlinear characteristics of multiple factors. The embodiment of the present disclosure is based on the Crossformer architecture, through the hybrid architecture of local window attention mechanism, global attention mechanism, cross-window attention mechanism and the sequential combination of global and hybrid attention encoders. The use of local window design in the hybrid attention encoder is very effective for understanding short-term patterns in battery changes. The global attention mechanism can effectively capture long-distance dependencies and can effectively identify medium- and long-term trends in battery degradation. The cross-attention mechanism can allow different local window attentions and global window attentions to share information, enhance the model's understanding of complex interactions, and combine multiple attention mechanisms to reduce the computational complexity of the model. During the incremental transfer learning process, in order to ensure the accuracy of the large model in predicting each type of battery status and its wide applicability throughout the entire life cycle, a shared network and a specific task branch network are set up according to the impact of differences in battery main materials, battery cell design, and battery system design on performance reliability. Based on the differentiated needs of batteries, the parameters of the shared network are adjusted sequentially, orderly, and step by step, and independent branch networks are expanded. During the training process, the multi-task learning combined loss function is combined to collaboratively ensure that the core needs of status prediction are met.

[0214] The above-mentioned embodiments of the present disclosure may be implemented separately or in any combination without conflict. The specific configuration may be based on actual needs and the present disclosure does not limit this.

[0215] Any battery status prediction method provided in the embodiments of the present disclosure can be executed by any appropriate electronic device with data processing capabilities, including but not limited to: electronic devices such as terminal devices and servers. Alternatively, any battery status prediction method provided in the embodiments of the present disclosure can be executed by a processor, such as the processor executing any battery status prediction method mentioned in the embodiments of the present disclosure by calling the corresponding instructions stored in the memory. This will not be repeated below.

[0216] Exemplary Devices

[0217] Fig.17 is a schematic diagram of the structure of a battery state prediction device provided by an exemplary embodiment of the present disclosure. The device of this embodiment can be used to implement the corresponding method embodiment of the present disclosure, such as Fig.17 The device shown may include: an acquisition module 51 and a first processing module 52 .

[0218] The acquisition module 51 is used to acquire first incremental characteristic data corresponding to the battery to be predicted.

[0219] Among them, the first incremental characteristic data includes test calibration characteristic data of the battery to be predicted in the first time period, usage environment characteristic data, usage condition characteristic data, detection characteristic data obtained from the detection of the battery to be predicted, and type characteristic data of the battery to be predicted.

[0220] The first processing module 52 is used to predict the battery state of the battery to be predicted in the second time period based on the first incremental feature data and through the target state prediction model.

[0221] The second time period is a future time period relative to the first time period. The target state prediction model is a cross-transformation model; the target state prediction model is obtained by incremental transfer learning based on historical incremental feature data in a historical time period; the historical incremental feature data includes at least one battery test calibration feature data, use environment feature data, use condition feature data, detection feature data and type feature data in a historical time period; the historical time period includes the time period before the first time period;

[0222] The test calibration characteristic data is determined based on the data obtained in the experimental or test environment; the usage environment characteristic data is determined based on the external environment data of the battery; the usage condition characteristic data is determined based on the operating data of the battery during use; and the detection characteristic data is determined based on the monitoring data of the battery.

[0223] In some optional embodiments of the present disclosure, in the above Fig.17On the basis of the illustrated embodiment, the first processing module 52 is specifically used to: obtain a first coding feature through a preset number of encoders based on the first incremental feature data. The preset number of encoders includes a first number of mixed attention encoders and a second number of global attention encoders. Based on the first coding feature, the battery state of the battery to be predicted in the second time period is obtained through the prediction head network layer.

[0224] In some optional embodiments of the present disclosure, a first number of hybrid attention encoders and a second number of global attention encoders are connected in series in a preset order.

[0225] The first processing module 52 is specifically used for: for the i-th encoder among a preset number of encoders, based on the input data corresponding to the i-th encoder, obtaining the i-th encoding result through the i-th encoder; i=1, 2, ..., n, n is a preset number; the input data corresponding to the 1st encoder is the first incremental feature data; the input data corresponding to the j-th encoder is the j-1th encoding result of the j-1th encoder; j=2, 3, ..., n; the n-th encoding result corresponding to the n-th encoder is used as the first encoding feature.

[0226] In some optional embodiments of the present disclosure, based on any of the above embodiments, the test calibration characteristic data includes the single cell cycle life characteristics and single cell calendar life characteristics of the battery.

[0227] In some optional embodiments, the cell cycle life characteristics include charge and discharge cycle data of the battery under different conditions.

[0228] In some optional embodiments, the cell calendar life characteristics include the calendar life of the battery under different conditions.

[0229] In some optional embodiments of the present disclosure, the usage environment characteristic data includes the usage environment information of the battery every day in the corresponding time period. The usage environment information of each day includes at least one of the average temperature, maximum temperature difference, average air pressure, maximum air pressure difference, average relative humidity, and average relative humidity difference of the day.

[0230] In some optional embodiments of the present disclosure, the usage condition characteristic data includes the usage condition information of the battery every day in the corresponding time period. The usage condition information of each day includes at least one of the average charging current, average discharging current, average charging capacitance, average discharging capacitance, average charging start state of charge, average charging end state of charge, average discharging start state of charge, average charging end state of charge, continuous charging frequency, and continuous discharging frequency of the day.

[0231] In some optional embodiments of the present disclosure, the detection feature data includes the detection data of the battery every day in the corresponding time period. The detection data of each day includes at least one of the predicted health status of the battery corresponding to that day, the cumulative charging capacity from leaving the factory to that day, the cumulative discharge capacity, the cumulative charge and discharge capacity, the total duration, the average of the voltage difference at the end of charging on that day, the maximum temperature difference of the battery, the highest temperature, the average temperature, the lowest temperature, the proportion of time when the temperature is greater than the first quantile number, the proportion of time when the temperature is less than the first quantile number and greater than the second quantile number, and the proportion of time when the temperature is less than the second quantile number.

[0232] In some optional embodiments of the present disclosure, the type characteristic data includes a classification identifier for each level in a preset level classification corresponding to the battery; the preset level classification includes a first-level classification of main materials, a second-level classification of battery cell differentiated design, and a third-level classification of battery system design.

[0233] In some optional embodiments of the present disclosure, based on any of the above embodiments, the first processing module 52 is specifically used to: process the test calibration feature data in the first incremental feature data based on a preset processing method to obtain processed intermediate feature data. Determine the second incremental feature data based on the intermediate feature data and the usage environment feature data, usage condition feature data, detection feature data and type feature data in the first incremental feature data. Based on the second incremental feature data, the battery state of the battery to be predicted in the second time period is predicted through the target state prediction model.

[0234] Fig.18 It is a structural schematic diagram of a battery status prediction device provided by another exemplary embodiment of the present disclosure.

[0235] In some optional embodiments of the present disclosure, based on any of the above embodiments, Fig.18 As shown, the apparatus of the embodiment of the present disclosure may further include a third processing module 53, which is used to perform the following operations to obtain a target state prediction model: obtain historical incremental feature data within a historical time period. Determine the state label data corresponding to the historical incremental feature data. Based on the historical incremental feature data and the state label data, perform incremental transfer learning on the historical state prediction model to obtain a target state prediction model. The historical state prediction model includes a shared network learned at a historical moment and at least one historical branch network.

[0236] In some optional embodiments of the present disclosure, the third processing module 53 is specifically used to: determine whether there is a new type according to the type feature data in the historical incremental feature data. In response to the existence of the new type, determine the new branch network according to the new type and the historical state prediction model, and determine the initial parameters of the new branch network from the historical branch network of the historical state prediction model, and obtain the current state prediction model based on the historical state prediction model and the new branch network. Train the current state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model.

[0237] In some optional embodiments, the third processing module 53 is further used to: in response to the absence of a newly added type, fine-tune the historical state prediction model based on the historical incremental feature data and the state label data to obtain a target state prediction model.

[0238] In some optional embodiments of the present disclosure, the third processing module 53 is specifically used to: in response to the historical incremental feature data only including feature data of the newly added type, based on the historical incremental feature data and the state label data, adjust the network parameters of the shared network and the newly added branch network in the current state prediction model to obtain the target state prediction model. In response to the historical incremental feature data including feature data of the newly added type and feature data of the historical type, based on the historical incremental feature data and the state label data, adjust the parameters of the shared network, the historical branch network and the newly added branch network to obtain the target state prediction model; wherein, the network parameters of the historical branch network are adjusted based on the first learning rate, and the network parameters of the newly added branch network are adjusted based on the second learning rate; the second learning rate is greater than the first learning rate; each historical type corresponds to a historical branch network; the historical type refers to the battery type that has been learned by the historical state prediction model.

[0239] In some optional embodiments of the present disclosure, the shared network includes a third number of encoders; the type feature data includes a classification identifier of each level in the preset level classification; the third number is the number of levels in the preset level classification. The third processing module 53 is specifically used to: determine whether the network parameters of each encoder in the shared network are frozen according to the classification identifier of each level of the newly added type; the frozen state includes two states: frozen and unfrozen; determine the learning rate corresponding to each encoder according to whether the network parameters of each encoder are frozen; the learning rate corresponding to the frozen encoder is the third learning rate, and the learning rate corresponding to the unfrozen encoder is the fourth learning rate; the third learning rate is less than the fourth learning rate; based on the historical incremental feature data and the state label data, the current state prediction model is trained according to the learning rate corresponding to each encoder to obtain the target state prediction model.

[0240] In some optional embodiments of the present disclosure, based on any of the above embodiments, the preset level classification includes primary classification of main materials, secondary classification of differentiated cell design, and tertiary classification of battery system design. The shared network includes a first encoder corresponding to the primary classification of main materials, a second encoder corresponding to the secondary classification of differentiated cell design, and a third encoder corresponding to the tertiary classification of battery system design.

[0241] The third processing module 53 is specifically used for: in response to the historical classification that has been learned by the historical state prediction model not including the first-level classification mark of the newly added type, determining that the states of the first encoder, the second encoder and the third encoder are all thawed. In response to the historical classification including the first-level classification mark of the newly added type, determining that the state of the network parameters of the first encoder is frozen. In the case where the historical classification includes the first-level classification mark of the newly added type, in response to the historical classification not including the second-level classification mark of the newly added type, determining that the network parameters of the second encoder and the third encoder are thawed. In response to the historical classification including the second-level classification mark of the newly added type, determining that the state of the network parameters of the second encoder is frozen. In the case where the historical classification includes the second-level classification mark of the newly added type, in response to the historical classification not including the third-level classification mark of the newly added type, determining that the state of the network parameters of the third encoder is thawed.

[0242] In some optional embodiments of the present disclosure, each branch network includes a fourth number of encoders; the fourth number of encoders includes at least one hybrid attention encoder and / or at least one global attention encoder.

[0243] In some optional embodiments of the present disclosure, based on any of the above embodiments, the third processing module 53 is specifically used to: determine the training sample data and the corresponding state label data based on the historical incremental feature data and the state label data. Based on the training sample data, obtain the state prediction result corresponding to the training sample data through the current state prediction model. Determine the first loss based on the state prediction result and the state label data. Determine the comprehensive loss based on the first loss, the second loss corresponding to the historical state prediction model, and the pre-configured loss weight. Based on the comprehensive loss, update the network parameters of the current state prediction model to obtain an updated state prediction model. Determine the target state prediction model based on the updated state prediction model.

[0244] In some optional embodiments of the present disclosure, the target state prediction model includes a first number of hybrid attention encoders and a second number of global attention encoders.

[0245] In some optional embodiments, each hybrid attention encoder includes at least one local window attention layer, a global window attention layer, and a fusion network layer.

[0246] In some optional embodiments, each local window attention layer is used to perform a local window attention operation on the first input feature corresponding to the layer to obtain a local attention result corresponding to the local window attention layer; the first input feature is the output feature of the previous network layer.

[0247] In some optional embodiments, the global window attention layer is used to perform a global window attention operation on the first input feature to obtain a first global attention result.

[0248] In some optional embodiments, the fusion network layer is used to fuse each local attention result with the first global attention result to obtain a first fusion feature, and the first fusion feature is used to determine the encoding result of the hybrid attention encoder.

[0249] In some optional embodiments, each global attention encoder includes a global window attention layer, which is used to perform a global window attention operation on a second input feature corresponding to the layer to obtain a second global attention result; the second global attention result is used to determine the encoding result corresponding to the global attention encoder.

[0250] In some optional embodiments of the present disclosure, the first processing module 52 is specifically used to: obtain historical feature data corresponding to the battery to be predicted. Determine current feature data based on the first incremental feature data and the historical feature data. Based on the current feature data, predict the battery state of the battery to be predicted in the second time period through the target state prediction model.

[0251] The above-mentioned embodiments of the present disclosure may be implemented separately or in any combination without conflict. The specific configuration may be based on actual needs and the present disclosure does not limit this.

[0252] The beneficial technical effects corresponding to the exemplary embodiment of the present device can be found in the corresponding beneficial technical effects of the above exemplary method section, which will not be repeated here.

[0253] Exemplary Electronic Devices

[0254] Fig.19 It is a structural diagram of an electronic device provided by an embodiment of the present disclosure, including at least one processor 91 and a memory 92.

[0255] The processor 91 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 90 to perform desired functions.

[0256] The memory 92 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 91 may execute one or more computer program instructions to implement the methods and / or other desired functions of the various embodiments of the present disclosure described above.

[0257] In one example, the electronic device 90 may further include: an input device 93 and an output device 94 , and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0258] The input device 93 may also include, for example, a keyboard, a mouse, a touch screen, a microphone, various sensors, and the like. Sensors may include, for example, image sensors (such as cameras, cameras, etc.), laser radars, millimeter-wave radars, ultrasonic radars, positioning sensors, pressure sensors, air quality sensors, temperature sensors, and the like. Image sensors, laser radars, millimeter-wave radars, ultrasonic radars, and the like may be used to sense the surrounding environment, that is, to detect dynamic and static objects in the surrounding environment. Dynamic and static objects may include, for example, static objects such as lane lines, curbs, arrows, signs, trees, and buildings, as well as dynamic objects such as surrounding vehicles, pedestrians, and cyclists. Positioning sensors are used to realize the positioning of movable devices (such as self-vehicles, robots, and the like) where electronic devices are located. Positioning sensors may include, for example, inertial measurement units (IMUs), global positioning systems (GPSs), and the like. Pressure sensors may be used to detect seat pressure. Temperature sensors may be used to detect the temperature in the vehicle cabin. Air quality sensors may be used to detect the air quality in the vehicle cabin.

[0259] The output device 94 can output various information to the outside, and may include, for example, a display, a speaker, a communication network and a remote output device connected thereto, and the like.

[0260] Of course, to simplify, Fig.19 Only some of the components related to the present disclosure in the electronic device 90 are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, according to specific application situations, the electronic device 90 may also include any other appropriate components.

[0261] Exemplary computer program products and computer-readable storage media

[0262] In addition to the above methods and devices, embodiments of the present disclosure may also provide a computer program product, including computer program instructions, which, when executed by a processor, enable the processor to execute the steps of the methods of various embodiments of the present disclosure described in the above “Exemplary Methods” section.

[0263] The computer program product may be written in any combination of one or more programming languages ​​to write program code for performing the operations of the disclosed embodiments, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0264] In addition, an embodiment of the present disclosure may also be a computer-readable storage medium having computer program instructions stored thereon. When the computer program instructions are executed by a processor, the processor executes the steps in the method of various embodiments of the present disclosure described in the above “Exemplary Method” section.

[0265] Computer readable storage media can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium is, for example, but not limited to, a system, device or device including electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0266] The basic principles of the present disclosure are described above in conjunction with specific embodiments. However, the advantages, strengths, effects, etc. mentioned in the present disclosure are only examples and not limitations, and cannot be considered as necessary for each embodiment of the present disclosure. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, rather than limitation, and the above details do not limit the present disclosure to being implemented by adopting the above specific details.

[0267] Those skilled in the art may make various changes and modifications to the present disclosure without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present disclosure claims and their equivalents, the present disclosure is also intended to include these modifications and variations.

Claims

1. A method for predicting a battery state, comprising: Acquire first incremental characteristic data corresponding to the battery to be predicted; The first incremental characteristic data includes test calibration characteristic data of the battery to be predicted within a first time period, usage environment characteristic data, usage condition characteristic data, detection characteristic data obtained by detecting the battery to be predicted, and type characteristic data of the battery to be predicted; Based on the first incremental feature data, a battery state of the battery to be predicted in a second time period is predicted by a target state prediction model; the second time period is a future time period relative to the first time period; Wherein, the target state prediction model is a cross-transformation model; the target state prediction model is obtained by incremental transfer learning based on historical incremental feature data in a historical time period; the historical incremental feature data includes test calibration feature data, usage environment feature data, usage condition feature data, detection feature data and type feature data of at least one battery in the historical time period; the historical time period includes the time period before the first time period; The test calibration characteristic data is determined based on the data obtained in the experiment or test environment; the use environment characteristic data is determined based on the external environment data of the battery; the use condition characteristic data is determined based on the operation data of the battery during use; the detection characteristic data is determined based on the monitoring data of the battery; The method further includes the following operations of obtaining the target state prediction model: Acquire the historical incremental feature data within the historical time period; Determine the state label data corresponding to the historical incremental feature data; Based on the historical incremental feature data and the state label data, incremental transfer learning is performed on the historical state prediction model to obtain the target state prediction model; the historical state prediction model includes a shared network learned at a historical moment and at least one historical branch network; The step of performing incremental transfer learning on the historical state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model includes: Determining whether there is a new type based on the type feature data in the historical incremental feature data; In response to the presence of a newly added type, a newly added branch network is determined according to the newly added type and the historical state prediction model, and initial parameters of the newly added branch network are determined from the historical branch network of the historical state prediction model, and a current state prediction model is obtained based on the historical state prediction model and the newly added branch network; Based on the historical incremental feature data and the state label data, the current state prediction model is trained to obtain the target state prediction model; or, In response to the absence of the newly added type, fine-tuning the historical state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model; The shared network includes a third number of encoders; the type feature data includes a classification identifier of each level in a preset level classification; the third number is the number of levels in the preset level classification; The training of the current state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model includes: Determine, according to the classification identifier of each level of the newly added type, whether the network parameters of each encoder in the shared network are frozen; the freezing status includes two states: frozen and unfrozen; Determine the learning rate corresponding to each encoder according to whether the network parameters of each encoder are frozen; the learning rate corresponding to the frozen encoder is a third learning rate, and the learning rate corresponding to the unfrozen encoder is a fourth learning rate; the third learning rate is less than the fourth learning rate; Based on the historical incremental feature data and the state label data, the current state prediction model is trained according to the learning rates corresponding to the encoders to obtain the target state prediction model.

2. The method according to claim 1, wherein: The step of predicting the battery state of the battery to be predicted in the second time period based on the first incremental feature data by using a target state prediction model includes: Based on the first incremental feature data, obtaining a first encoding feature through a preset number of encoders; the preset number of encoders includes a first number of hybrid attention encoders and a second number of global attention encoders; Based on the first coding feature, the battery state of the battery to be predicted in the second time period is obtained through the prediction head network layer.

3. The method according to claim 2, wherein: The first number of hybrid attention encoders and the second number of global attention encoders are connected in series according to a preset order; The obtaining of a first coding feature based on the first incremental feature data by using a preset number of encoders includes: For the i-th encoder among the preset number of encoders, based on the input data corresponding to the i-th encoder, the i-th encoding result is obtained through the i-th encoder; i=1, 2, ..., n, n is a preset number; the input data corresponding to the 1st encoder is the first incremental feature data; the input data corresponding to the j-th encoder is the j-1th encoding result of the j-1th encoder; j=2, 3, ..., n; the n-th encoding result corresponding to the n-th encoder is used as the first encoding feature.

4. The method according to claim 1, wherein: The test calibration characteristic data includes the single cell cycle life characteristics and single cell calendar life characteristics of the battery; The single cell cycle life characteristics include charge and discharge cycle data of the battery under different conditions; the single cell calendar life characteristics include calendar life of the battery under different conditions; and / or, The usage environment characteristic data includes the usage environment information of the battery every day in the corresponding time period; the usage environment information of each day includes at least one of the average temperature, maximum temperature difference, average air pressure, maximum air pressure difference, average relative humidity, and average relative humidity difference of the day; and / or, The usage condition characteristic data includes the usage condition information of the battery every day in the corresponding time period; the usage condition information of each day includes at least one of the average charging current, average discharging current, average charging capacitance, average discharging capacitance, average charge start state of charge, average charge end state of charge, average discharge start state of charge, average charge end state of charge, continuous charging frequency, and continuous discharging frequency of the day; and / or, The detection feature data includes the detection data of the battery for each day in the corresponding time period; the detection data of each day includes at least one of the predicted health status of the battery corresponding to that day, the cumulative charging capacity from leaving the factory to that day, the cumulative discharging capacity, the cumulative charging and discharging capacity, the total duration, the average of the charging terminal pressure difference for that day, the maximum temperature difference of the battery, the highest temperature, the average temperature, the lowest temperature, the proportion of the time when the temperature is greater than the first quantile number, the proportion of the time when the temperature is less than the first quantile number and greater than the second quantile number, and the proportion of the time when the temperature is less than the second quantile number; and / or, The type characteristic data includes the classification identification of each level in the preset level classification corresponding to the battery; the preset level classification includes the first-level classification of main materials, the second-level classification of battery cell differentiated design, and the third-level classification of battery system design.

5. The method according to claim 1, wherein: The step of predicting the battery state of the battery to be predicted in the second time period based on the first incremental feature data by using a target state prediction model includes: Processing the test calibration feature data in the first incremental feature data based on a preset processing method to obtain processed intermediate feature data; Determine second incremental feature data based on the intermediate feature data and the usage environment feature data, the usage condition feature data, the detection feature data, and the type feature data in the first incremental feature data; Based on the second incremental feature data, the battery state of the battery to be predicted in the second time period is predicted by the target state prediction model.

6. The method according to claim 1, wherein: The training of the current state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model includes: In response to the historical incremental feature data only including feature data of a newly added type, adjusting network parameters of a shared network and a newly added branch network in the current state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model; In response to the fact that the historical incremental feature data includes feature data of a newly added type and feature data of a historical type, the parameters of the shared network, the historical branch network and the newly added branch network are adjusted based on the historical incremental feature data and the state label data to obtain the target state prediction model; wherein the network parameters of the historical branch network are adjusted based on a first learning rate, and the network parameters of the newly added branch network are adjusted based on a second learning rate; the second learning rate is greater than the first learning rate; each historical type corresponds to a historical branch network; the historical type refers to the battery type that has been learned by the historical state prediction model.

7. The method according to claim 1, wherein: The preset level classification includes the primary classification of main materials, the secondary classification of battery cell differentiated design, and the tertiary classification of battery system design; the shared network includes a first encoder corresponding to the primary classification of main materials, a second encoder corresponding to the secondary classification of battery cell differentiated design, and a third encoder corresponding to the tertiary classification of battery system design; The determining, according to the classification identifier of each level of the newly added type, whether the network parameters of each encoder in the shared network are frozen includes: In response to the historical classifications learned by the historical state prediction model not including the primary classification identifier of the newly added type, determining that the states of the first encoder, the second encoder, and the third encoder are all unfrozen; In response to the historical classification including the primary classification identifier of the newly added type, determining that the state of the network parameter of the first encoder is frozen; In a case where the historical classification includes the primary classification identifier of the newly added type, in response to the historical classification not including the secondary classification identifier of the newly added type, determining that the network parameters of the second encoder and the third encoder are unfrozen; In response to the historical classification including the secondary classification identifier of the newly added type, determining that the state of the network parameter of the second encoder is frozen; In a case where the historical classification includes the secondary classification identifier of the newly added type, in response to the historical classification not including the tertiary classification identifier of the newly added type, it is determined that the state of the network parameter of the third encoder is unfrozen.

8. The method according to claim 1, wherein: Each branch network includes a fourth number of encoders; the fourth number of encoders includes at least one hybrid attention encoder and / or at least one global attention encoder.

9. The method according to claim 1, wherein: The training of the current state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model includes: Determine training sample data and corresponding state label data based on the historical incremental feature data and the state label data; Based on the training sample data, obtaining a state prediction result corresponding to the training sample data through the current state prediction model; Determining a first loss based on the state prediction result and the state label data; Determine a comprehensive loss based on the first loss, the second loss corresponding to the historical state prediction model, and a preconfigured loss weight; Based on the comprehensive loss, updating the network parameters of the current state prediction model to obtain an updated state prediction model; Based on the updated state prediction model, the target state prediction model is determined.

10. The method according to any one of claims 1 to 5, wherein: The target state prediction model includes a first number of hybrid attention encoders and a second number of global attention encoders; Each hybrid attention encoder includes at least one local window attention layer, a global window attention layer, and a fusion network layer; Each of the local window attention layers is used to perform a local window attention operation on a first input feature corresponding to the layer to obtain a local attention result corresponding to the local window attention layer; the first input feature is an output feature of a previous network layer; The global window attention layer is used to perform a global window attention operation on the first input feature to obtain a first global attention result; The fusion network layer is used to fuse each of the local attention results and the first global attention result to obtain a first fusion feature, and the first fusion feature is used to determine the encoding result of the hybrid attention encoder; and / or, Each of the global attention encoders includes a global window attention layer, which is used to perform a global window attention operation on a second input feature corresponding to the layer to obtain a second global attention result; The second global attention result is used to determine the encoding result corresponding to the global attention encoder.

11. The method according to any one of claims 1 to 5, wherein: The step of predicting the battery state of the battery to be predicted in the second time period based on the first incremental feature data by using a target state prediction model includes: Acquire historical characteristic data corresponding to the battery to be predicted; Determining current feature data based on the first incremental feature data and the historical feature data; Based on the current feature data, the battery state of the battery to be predicted in the second time period is predicted by the target state prediction model.

12. A battery status prediction device, comprising: An acquisition module, used for acquiring first incremental characteristic data corresponding to the battery to be predicted; The first incremental characteristic data includes test calibration characteristic data of the battery to be predicted within a first time period, usage environment characteristic data, usage condition characteristic data, detection characteristic data obtained by detecting the battery to be predicted, and type characteristic data of the battery to be predicted; A first processing module is configured to predict the battery state of the battery to be predicted in a second time period based on the first incremental feature data and through a target state prediction model; the second time period is a future time period relative to the first time period; Wherein, the target state prediction model is a cross-transformation model; the target state prediction model is obtained by incremental transfer learning based on historical incremental feature data in a historical time period; the historical incremental feature data includes test calibration feature data, usage environment feature data, usage condition feature data, detection feature data and type feature data of at least one battery in the historical time period; the historical time period includes the time period before the first time period; The test calibration characteristic data is determined based on the data obtained in the experiment or test environment; the use environment characteristic data is determined based on the external environment data of the battery; the use condition characteristic data is determined based on the operation data of the battery during use; the detection characteristic data is determined based on the monitoring data of the battery; The third processing module is used to perform the following operations to obtain the target state prediction model: Acquire the historical incremental feature data within the historical time period; determine the state label data corresponding to the historical incremental feature data; perform incremental transfer learning on the historical state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model; the historical state prediction model includes a shared network learned at a historical moment and at least one historical branch network; The third processing module is specifically used for: Determine whether there is a new type according to the type feature data in the historical incremental feature data; in response to the existence of the new type, determine the new branch network according to the new type and the historical state prediction model, and determine the initial parameters of the new branch network from the historical branch network of the historical state prediction model, and obtain the current state prediction model based on the historical state prediction model and the new branch network; train the current state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model; or, in response to the absence of the new type, fine-tune the historical state prediction model based on the historical incremental feature data and the state label data to obtain the target state prediction model; The shared network includes a third number of encoders; the type feature data includes a classification identifier of each level in a preset level classification; the third number is the number of levels in the preset level classification; The third processing module is specifically used for: According to the classification identifier of each level of the newly added type, determine whether the network parameters of each encoder in the shared network are frozen; the frozen state includes two states: frozen and unfrozen; according to the frozen state of each encoder, determine the learning rate corresponding to each encoder; the learning rate corresponding to the frozen encoder is the third learning rate, and the learning rate corresponding to the unfrozen encoder is the fourth learning rate; the third learning rate is less than the fourth learning rate; based on the historical incremental feature data and the state label data, the current state prediction model is trained according to the learning rate corresponding to each encoder to obtain the target state prediction model.

13. An electronic device, comprising: processor; a memory for storing instructions executable by the processor; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the method according to any one of claims 1 to 11; or, The electronic device comprises: the battery status prediction device as claimed in claim 12 above.

14. A computer-readable storage medium storing a computer program, wherein the computer program is used to execute the method according to any one of claims 1 to 11.

15. A computer program product, when instructions in the computer program product are executed by a processor, performs the method according to any one of claims 1 to 11 of the present disclosure.

Images