Knowledge-data fusion driven power battery available energy prediction method and system

Abstract

The application provides a knowledge-data fusion driven power battery available energy prediction method and system, a battery mechanism model, a battery characteristic prediction model and a data driven model of reliable data are sequentially established according to the basic principle of the power battery, the available energy prediction of the battery in an open loop mode is realized through the mechanism model and the battery characteristic prediction model, the available energy correction under the condition of reliable data is realized through the data driven model, so that the power battery available energy prediction of the fusion of the knowledge driven method and the data driven method is realized, the method and the system can improve the available energy prediction precision and the scene coverage of the battery, and reduce the unexplainability of the data driven model.

Description

Technical Field

[0001] This invention relates to the field of power battery testing, and specifically to a knowledge-data fusion-driven method and system for predicting the available energy of power batteries. Background Technology

[0002] As the production, sales and ownership of new energy vehicles continue to increase, the consumer market's requirements for new energy vehicles are constantly rising. As a key component of new energy vehicles, the performance and control of power batteries directly affect the user experience of new energy vehicle customers.

[0003] However, the state estimation and energy consumption prediction of vehicle-mounted power batteries often lack widely applicable methods due to the difficulty in balancing computational load and accuracy in their models. Existing open-loop prediction accuracy of available energy for power batteries is poor. Therefore, new solutions are needed to address the complex operating conditions of vehicle-mounted power batteries and the lack of continuous adaptive correction methods. Summary of the Invention

[0004] To address the issues of poor accuracy in predicting available energy of existing power batteries, this invention proposes a knowledge-data fusion-driven method for predicting available energy of power batteries. This method establishes a battery mechanistic model, a battery characteristic prediction model, and a data-driven model based on reliable data, all conforming to the fundamental principles of power batteries. It achieves open-loop available energy prediction through the mechanistic and battery characteristic prediction models, and corrects available energy under reliable data conditions through the data-driven model. This method integrates knowledge-driven and data-driven approaches, improving the accuracy and scenario coverage of battery available energy prediction while reducing the uninterpretability of the data-driven model. This invention also relates to a knowledge-data fusion-driven system for predicting available energy of power batteries.

[0005] The technical solution of the present invention is as follows:

[0006] A knowledge-data fusion-driven method for predicting the available energy of a power battery, characterized by comprising the following steps;

[0007] S1. Obtain cloud data uploaded by the vehicle-side power battery system to the cloud platform. On the cloud platform, perform data analysis on the acquired cloud data to extract the operating condition characteristics of the power battery. Use a data dimensionality reduction algorithm to reduce the dimensionality of the high-dimensional characteristics of the power battery operating conditions. Perform correlation analysis on the dimensionality-reduced characteristics based on physical knowledge. Then, use a data-driven operating condition characteristic classification method to classify the characteristics. Calculate the probability map of the power battery switching between different operating conditions. Finally, use an operating condition generation algorithm to establish the operating condition characteristics of the power battery.

[0008] S2. Based on the operating principles of power batteries, a mechanistic model of the power battery is established on the cloud platform. The input conditions of the mechanistic model are charging / discharging current / power, temperature, and initial SOC parameters. The output conditions of the mechanistic model are the positive and negative electrode ion concentrations and electrolyte ion concentrations of the power battery. Based on the mechanistic model, a battery characteristic prediction model integrating metaheuristic algorithms and physical information neural networks is established. The battery characteristic prediction model simulates the characteristic operating conditions of the input power battery to predict the positive and negative electrode ion concentrations and electrolyte ion concentrations of the power battery under given input conditions. The predicted positive and negative electrode ion concentrations and the calibrated power battery ion concentrations are then analyzed on the cloud platform. The SOC of the positive and negative electrodes is calculated based on the maximum ion concentration of the negative electrode. Then, the potential of the positive and negative electrodes is calculated based on the calculated SOC and the calibrated SOC-OCV curves of the positive and negative electrodes. The electrolyte potential is calculated using the predicted electrolyte ion concentration. Then, the negative electrode potential, the electrolyte potential, and the ohmic voltage drop obtained based on the characteristic operating conditions of the power battery are subtracted from the positive electrode potential to obtain the power battery terminal voltage. The power battery power is obtained by using the power battery terminal voltage and the current obtained from the characteristic operating conditions of the power battery. When the power battery is continuously virtually discharged to the cutoff voltage, the available energy of the power battery is calculated by power-time integration, which is used as the model to predict the available energy.

[0009] S3. Conduct offline testing experiments, including capacity testing and dynamic charge-discharge testing, to measure the actual usable energy of the power battery under the same initial conditions of charge-discharge current / power, temperature, and initial SOC parameters.

[0010] S4. Using a data-driven approach on the cloud platform, the model-predicted available energy obtained in step S2 is used as input, and the actual test available energy obtained in step S3 is used as output to jointly establish a training set and a test set, and construct a data-driven model. The data-driven model establishes a mapping relationship between the model-predicted available energy and the actual test available energy through the data-driven approach, thereby realizing the correction of the model-predicted available energy.

[0011] Preferably, the feature is that step S1 acquires cloud data uploaded by the vehicle-side power battery system to the cloud platform, including voltage, current, time, temperature, SOC, charge / discharge status, battery output power, battery fault status, and vehicle speed information; on the cloud platform, data analysis is performed on the acquired cloud data to extract the operating condition characteristics of the power battery, including charging frequency, charge / discharge depth, acceleration time, deceleration time, average current, average voltage, average power, maximum power, minimum power, and power distribution curve characteristics; an operating condition generation algorithm is used to establish the operating condition characteristics of the power battery, wherein the battery operating condition characteristics include a current-time operating condition curve, and the curve length is selected as a characteristic operating condition of less than 1800s.

[0012] Preferably, the data dimensionality reduction algorithm used in step S1 includes principal component analysis, partial principal component analysis, singular value decomposition, or autoencoder algorithm; the data-driven working condition feature classification method used includes K-means clustering algorithm, random forest algorithm, and support vector machine algorithm; and the working condition generation algorithm used includes Monte Carlo simulation method, decision tree algorithm, and decision forest algorithm.

[0013] Preferably, in step S2, a power battery mechanism model is established on the cloud platform based on the power battery usage principle, including an equivalent circuit model, an electrochemical model, or a heterogeneous model. The equivalent circuit model is a first-order equivalent circuit model, a second-order equivalent circuit model, a fractional-order equivalent circuit model, or other electrochemically reduced-order equivalent circuit models. The electrochemical model is a pseudo-two-dimensional electrochemical model, a single-particle model, or a three-dimensional electrochemical model. The pseudo-two-dimensional electrochemical model integrates solid-phase lithium-ion concentration calculation, liquid-phase lithium-ion concentration calculation, solid-phase potential calculation, liquid-phase potential calculation, and lithium insertion / extraction / deintercalation calculation.

[0014] Preferably, step S2 establishes a battery characteristic prediction model based on the mechanistic model of the power battery, which integrates a metaheuristic algorithm and a physical information neural network. This involves introducing the prediction bias of the mechanistic model of the power battery into the penalty function of the physical information neural network algorithm and assigning it a weight greater than the weight threshold. Then, during the training process, a metaheuristic algorithm is used to optimize the parameters. The metaheuristic algorithm used is a particle swarm optimization-gray wolf optimization coupled algorithm, a chaotic search algorithm, or a particle swarm optimization-simulated annealing coupled algorithm. The physical information neural network algorithm is a long short-term memory neural network algorithm, a graph neural network algorithm, or a bidirectional long short-term memory neural network algorithm with an attention mechanism.

[0015] Preferably, step S3 involves conducting offline testing experiments to measure the actual usable energy of the power battery under the same initial conditions of charge / discharge current / power, temperature, and initial SOC parameters. This includes the actual usable energy of the power battery system under NEDC or CLTC test conditions, or the actual usable energy of the power battery system under characteristic operating conditions, or the actual usable energy of a single power battery cell under DST or FUDS test conditions.

[0016] Preferably, the data-driven model constructed in step S4 establishes a mapping relationship between the model's predicted available energy and the actual tested available energy through a data-driven method. The data-driven method employs extreme learning machine algorithm, BP neural network, recurrent neural network, ELMAN neural network, or LSTM neural network.

[0017] A knowledge-data fusion-driven power battery available energy prediction system is characterized by comprising a power battery usage characteristic condition establishment module, a model prediction available energy calculation module, an actual test available energy testing module, and a data-driven model construction module, all located on a cloud platform. The power battery usage characteristic condition establishment module is connected to the model prediction available energy calculation module, and both the model prediction available energy calculation module and the actual test available energy testing module are connected to the data-driven model construction module.

[0018] The power battery usage characteristic condition establishment module acquires cloud data uploaded by the vehicle-side power battery system to the cloud platform, extracts the power battery usage condition characteristics by performing data analysis on the acquired cloud data, uses a data dimensionality reduction algorithm to reduce the dimensionality of the high-dimensional features of the power battery usage condition, performs correlation analysis on the dimensionality-reduced features based on physical knowledge, classifies them using a data-driven condition feature classification method, calculates the probability map of the power battery switching between different conditions, and uses a condition generation algorithm to establish the power battery usage characteristic conditions.

[0019] The model-predicted available energy calculation module includes a power battery mechanistic model construction module, a battery characteristic prediction model construction module, and a model-predicted available energy calculation processing module connected in sequence. The power battery mechanistic model construction module establishes a power battery mechanistic model based on the power battery's operating principles. The input conditions of the power battery mechanistic model are charging / discharging current / power, temperature, and initial SOC parameters. The output conditions of the power battery mechanistic model are the positive and negative electrode ion concentrations and the electrolyte ion concentration. The battery characteristic prediction model construction module establishes a battery characteristic prediction model based on the power battery mechanistic model, fusing a metaheuristic algorithm with a physical information neural network. This model simulates the characteristic operating conditions of the input power battery to predict the positive and negative electrode ion concentrations under given input conditions. The model predicts available energy calculation and processing module calculates the SOC of the positive and negative electrodes based on the predicted positive and negative electrode ion concentrations and the calibrated maximum ion concentrations of the positive and negative electrodes of the power battery. Then, it calculates the positive electrode potential and negative electrode potential based on the calculated SOC of the positive and negative electrodes and the calibrated SOC-OCV curves of the positive and negative electrodes. It calculates the electrolyte potential using the predicted electrolyte ion concentration. Then, it subtracts the negative electrode potential, electrolyte potential, and ohmic voltage drop obtained based on the characteristic operating conditions of the power battery from the positive electrode potential to obtain the power battery terminal voltage. The power battery power is obtained through the power battery terminal voltage and the current obtained from the characteristic operating conditions of the power battery. When the power battery is continuously virtually discharged to the cutoff voltage, the available energy of the power battery is calculated by power-time integration as the model predicted available energy.

[0020] The actual test usable energy test module conducts offline testing experiments, including capacity testing and dynamic charge-discharge testing, to measure the actual test usable energy of the power battery under the same initial conditions of charge-discharge current / power, temperature, and initial SOC parameters.

[0021] The data-driven model building module adopts a data-driven approach, using the model's predicted available energy as input and the actual tested available energy as output to jointly establish a training set and a test set, thereby constructing a data-driven model. The data-driven model establishes a mapping relationship between the model's predicted available energy and the actual tested available energy through the data-driven approach, thereby correcting the model's predicted available energy.

[0022] Preferably, the power battery usage characteristic condition establishment module acquires cloud data uploaded by the vehicle-side power battery system to the cloud platform, including voltage, current, time, temperature, SOC, charge / discharge status, battery output power, battery fault status, and vehicle speed information; it extracts power battery usage characteristic features through data analysis of the acquired cloud data, including charging frequency, charge / discharge depth, acceleration time, deceleration time, average current, average voltage, average power, maximum power, minimum power, and power distribution curve features; the data dimensionality reduction algorithms used include principal component analysis, partial principal component analysis, singular value decomposition, or autoencoder algorithm; the data-driven operating condition feature classification methods used include K-means clustering algorithm, random forest algorithm, and support vector machine algorithm; and the operating condition generation algorithm is used to establish the power battery usage characteristic conditions, including Monte Carlo simulation method, decision tree algorithm, and decision forest algorithm. The battery usage characteristic conditions include current-time operating condition curves, with curve lengths selected to be less than 1800s.

[0023] Preferably, the battery characteristic prediction model construction module in the model prediction available energy calculation module establishes a battery characteristic prediction model based on the power battery mechanistic model, which integrates a metaheuristic algorithm and a physical information neural network. This involves introducing the prediction bias of the power battery mechanistic model into the penalty function of the physical information neural network algorithm and assigning it a weight greater than a weight threshold. Then, during training, a metaheuristic algorithm is used for parameter optimization. The metaheuristic algorithm used is a particle swarm optimization-gray wolf optimization coupled algorithm, a chaotic search algorithm, or a particle swarm optimization-simulated annealing coupled algorithm. The physical information neural network algorithm is a long short-term memory neural network algorithm, a graph neural network algorithm, or a bidirectional long short-term memory neural network algorithm with an attention mechanism.

[0024] The beneficial effects of this invention are as follows:

[0025] This invention proposes a knowledge-data fusion-driven method for predicting the available energy of a power battery. First, it acquires cloud data uploaded from the vehicle's power battery system to a cloud platform, such as voltage, current, and time information. Based on the cloud platform, it extracts the battery's operating condition characteristics through data analysis. A data dimensionality reduction algorithm is used to reduce the dimensionality of these high-dimensional characteristics. Correlation analysis is performed on the reduced features based on physical knowledge, and data-driven methods such as clustering analysis and support vector machines are used for classification. The probability graph of battery transitions between different operating conditions is calculated, and then a Monte Carlo simulation method is used to establish the characteristic operating conditions of the power battery. Second, based on the operating principles of the power battery, a mechanistic model of the power battery, such as an equivalent circuit or electrochemical model, is established. Finally, a battery characteristic prediction model integrating a metaheuristic algorithm and a physical information neural network is constructed to predict the positive and negative electrode ions of the power battery under given input operating conditions. Based on the concentrations of the positive and negative electrodes and the electrolyte ion concentrations, and combined with parameters such as the calibrated maximum ion concentrations of the positive and negative electrodes and the SOC-OCV curves of the positive and negative electrodes, the terminal voltage of the power battery is calculated using a series of logical calculations. This results in a simulated predicted available energy, which is then combined with the actual available energy (or actual test available energy) of the power battery under the same initial conditions measured in offline experiments. Data-driven methods such as neural networks and regression analysis can be used to construct a data-driven model, establishing a mapping relationship between the model's predicted available energy and the actual test available energy. In actual use, based on the established battery operating conditions, the available energy (model-predicted available energy) is predicted under that battery state using a mechanistic model and a battery characteristic prediction model. Data-driven methods are then used to correct the model's predicted available energy, achieving accurate available energy prediction. This invention is a data-driven model that integrates knowledge-driven and data-driven methods to predict the available energy of a power battery. It establishes a battery mechanistic model, a battery characteristic prediction model, and a reliable data model that conform to the basic principles of power batteries. The mechanistic model and the battery characteristic prediction model enable open-loop prediction of battery available energy, while the data-driven model enables correction of available energy under reliable data conditions. This method can improve the accuracy and scenario coverage of battery available energy prediction, enhance the interpretability of the data-driven model, and improve the problem of poor open-loop prediction accuracy of the mechanistic model. It has significant application value in improving the accuracy of range estimation for new energy vehicles.

[0026] This invention also relates to a knowledge-data fusion-driven power battery available energy prediction system, corresponding to the aforementioned knowledge-data fusion-driven power battery available energy prediction method. It can be understood as a system that implements the aforementioned knowledge-data fusion-driven power battery available energy prediction method. This system includes a power battery usage characteristic condition establishment module, a model prediction available energy calculation module, an actual test available energy testing module, and a data-driven model construction module, all located on a cloud platform. These modules work together systematically, employing data dimensionality reduction algorithms, physical knowledge-driven methods, correlation analysis, data-driven condition feature classification methods, condition generation algorithms, power battery mechanistic model modeling, battery characteristic prediction model modeling, and model prediction available energy calculation processing. Combined with the construction of a data-driven model, it achieves available energy correction under reliable data conditions through data-driven methods, thereby integrating knowledge-driven and data-driven methods to improve battery available energy prediction accuracy and scenario coverage, and reduce the uninterpretability of the data-driven model. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the knowledge-data fusion-driven power battery available energy prediction method of this invention.

[0028] Figure 2 This is a preferred flowchart of a power battery availability prediction method driven by invention knowledge-data fusion.

[0029] Figure 3 These are the characteristic operating conditions of the power battery in the embodiments of the present invention.

[0030] Figure 4 This is a flowchart of the battery characteristic prediction model established in this invention.

[0031] Figure 5 This is a flowchart of the model prediction and available energy calculation process of the present invention.

[0032] Figure 6 This is an offline testing diagram of the power battery operating conditions according to an embodiment of the present invention.

[0033] Figure 7 This is a comparison diagram of the neural network before and after correction in an embodiment of the present invention.

[0034] Figure 8 This is a block diagram of the knowledge-data fusion-driven power battery available energy prediction system of the present invention. Detailed Implementation

[0035] The present invention will now be described with reference to the accompanying drawings.

[0036] This invention relates to a knowledge-data fusion-driven method for predicting the available energy of power batteries. The principle of this method is as follows: Figure 1As shown, this is a method for predicting the available energy of a power battery by acquiring cloud data of the power battery, establishing usage characteristic conditions, and then establishing a battery model, or a battery mechanism model and a battery characteristic prediction model that conform to the basic principles of power batteries. This enables the prediction of available energy of the battery (i.e., model-predicted available energy or model simulation-predicted available energy). Offline testing experiments are then conducted to obtain the actual available energy. The model simulation-predicted available energy is used as input, and the experimentally obtained actual available energy is used as output. Machine learning training is then performed to build a machine learning training model (or a data-driven model). This allows for the correction of the model simulation-predicted available energy, thereby integrating knowledge-driven and data-driven methods for predicting the available energy of a power battery.

[0037] Specifically, the flowchart of this knowledge-data fusion-driven method for predicting the available energy of power batteries is as follows: Figure 2 As shown, it includes the following steps;

[0038] S1: Obtain cloud data and establish characteristic operating conditions, which can also be referred to as the steps for establishing characteristic operating conditions of power batteries.

[0039] The system acquires cloud data uploaded from the vehicle's power battery system to the cloud platform. On the cloud platform, it performs data analysis on the acquired cloud data to extract the operating condition characteristics of the power battery. It then uses a data dimensionality reduction algorithm to reduce the dimensionality of the high-dimensional features of the power battery operating conditions. Based on physical knowledge, it performs correlation analysis on the dimensionality-reduced features and then uses a data-driven operating condition feature classification method to classify them. It calculates the probability map of the power battery switching between different operating conditions and uses an operating condition generation algorithm to establish the operating condition characteristics of the power battery.

[0040] Furthermore, cloud data such as voltage, current, time, temperature, SOC, charge / discharge status, battery output power, battery fault status, and vehicle speed information of the power battery are obtained through the cloud platform and used as real application data. Initial values ​​are input, and the above data features are used as characteristic data to form battery operating conditions. Then, data analysis is performed to extract the operating condition characteristics of the power battery, including charging frequency, depth of charge and discharge, acceleration time, deceleration time, average current, average voltage, average power, maximum power, minimum power, and power distribution curve characteristics. The high-dimensional features of the power battery's operating conditions are reduced using data dimensionality reduction algorithms, such as principal component analysis, partial principal component analysis, singular value decomposition, or autoencoder algorithms. Correlation analysis is then performed on the dimensionality-reduced features based on physical knowledge, followed by classification using data-driven operating condition feature classification methods, such as K-means clustering, DBSCAN density clustering, OPTICS density clustering, random forest, and support vector machine algorithms. The probability graph of the power battery transitioning between different operating conditions is calculated, and operating condition generation algorithms are used to establish the power battery's characteristic operating conditions. These algorithms may include Monte Carlo simulation, decision tree algorithms, and decision forest algorithms. The established power battery operating condition characteristic conditions include current-time operating condition curves, with curve lengths of less than 1800 seconds or longer segments.

[0041] In this embodiment, a Support Vector Machine (SVM) is used as the classifier. After the SVM training is complete, most of the training samples do not need to be retained, and the final model is only related to the support vectors. The input training set is T = {(x1,y1),(x2,y2),...,(x...} N ,y N )}, where x i ={U i ,I i ,T i ,K i G i}, i=1,2,...,N, with voltage, current, time, and temperature as continuous features, y i∈{+1,-1}, i=1,2,...N, and by using the operating conditions as classification features, the operating conditions of the data under the aforementioned voltage, current, temperature, and time conditions can be obtained relatively quickly, establishing a probabilistic process with operating conditions as random variables. Then, using a Monte Carlo stochastic process, distribution sampling is performed from the above probabilistic process to establish the battery operating conditions within the required time period. The Monte Carlo process for establishing operating conditions eliminates the need for discretization processing for the problem of continuous battery energy availability. Furthermore, for a large number of cloud data samples, the Monte Carlo operating condition simulation results will approximate the optimal solution. The Monte Carlo stochastic process will establish the battery's operating condition characteristics, i.e., extract operating condition features such as the current-time operating condition curve, such as... Figure 3 As shown.

[0042] S2. Establish a power battery model (or battery knowledge model). The model predicts available energy, which can also be called the model predicts available energy calculation step.

[0043] Based on the operating principles of power batteries and using real-world application data as initial input values, a mechanistic model of the power battery is established on the cloud platform. The input conditions for this model are charging / discharging current / power, temperature, and initial state of charge (SOC). The output conditions are the positive and negative electrode ion concentrations and the electrolyte ion concentration. A battery characteristic prediction model, integrating a metaheuristic algorithm and a physical information neural network, is then established based on this model. This model simulates the characteristic operating conditions of the input power battery to predict the positive and negative electrode ion concentrations and the electrolyte ion concentration under given input conditions. The predicted positive and negative electrode ion concentrations are then compared with calibrated values ​​on the cloud platform. The SOC of the positive and negative electrodes is calculated based on the maximum ion concentration of the positive and negative electrodes. Then, the positive electrode potential and negative electrode potential are calculated based on the calculated SOC and the calibrated SOC-OCV curves of the positive and negative electrodes. The electrolyte potential is calculated using the predicted electrolyte ion concentration. Then, the negative electrode potential, electrolyte potential, and ohmic voltage drop obtained based on the characteristic operating conditions of the power battery are subtracted from the positive electrode potential to obtain the power battery terminal voltage. The power battery power is obtained by using the power battery terminal voltage and the current obtained from the characteristic operating conditions of the power battery. When the power battery is continuously virtually discharged to the cutoff voltage, the available energy of the power battery is calculated by power-time integration, which is used as the model to predict the available energy.

[0044] S2.1 Establishing a mechanistic model of power batteries

[0045] Based on the operating principles of power batteries, a mechanistic model of the power battery is established on the cloud platform, including equivalent circuit models, electrochemical models, or heterogeneous models. Among them, the equivalent circuit model can be a first-order equivalent circuit model, a second-order equivalent circuit model, a fractional-order equivalent circuit model, or other electrochemically reduced equivalent circuit models, and the electrochemical model can be a pseudo-two-dimensional electrochemical model, a single-particle model, or a three-dimensional electrochemical model.

[0046] In this embodiment, a pseudo-two-dimensional electrochemical model of the battery is established as the mechanistic model of the power battery. This pseudo-two-dimensional electrochemical model integrates the calculation of solid-phase lithium-ion concentration, liquid-phase lithium-ion concentration, solid-phase potential, liquid-phase potential, and lithium insertion / extraction. Specifically, the formula of the pseudo-two-dimensional electrochemical model is as follows:

[0047] 1) Calculation of solid-phase lithium-ion concentration

[0048] A spherical coordinate system is established with the center of the spherical particles of the positive and negative electrode active materials as the origin. According to Fick's law, the expression for the solid-phase lithium ion concentration at coordinate r is shown in equation (1):

[0049]

[0050] The subscript s indicates a solid-phase region, and y = n or p corresponds to negative and positive pole particles, respectively; C s,y (r,t) represents the solid-phase lithium-ion concentration; D s,y t represents the lithium-ion diffusion coefficient of the positive and negative electrodes within the solid phase region; t represents time.

[0051] 2) Calculation of liquid-phase lithium ion concentration

[0052] The movement of lithium ions in the electrolyte is influenced by two factors: one is the lithium ion concentration gradient in the liquid phase, which leads to the diffusion process of lithium ions; the other is the migration of lithium ions under an electric field. The lithium ion concentration in the liquid phase satisfies equation (2):

[0053]

[0054] Where x is the integral quantity along the electrode thickness direction; the subscript e indicates the liquid phase region; c e (x,t) represents the concentration of lithium ions in the liquid phase; ε e D represents the volume fraction of the liquid electrolyte. eff t is the effective diffusion coefficient of lithium ions in the liquid phase; F is the Faraday constant; t + j is the transport number of the lithium ion; tot (t) represents the total current density.

[0055] 3) Calculation of solid-state potential

[0056] The solid-phase potential distribution in the positive and negative electrode materials is shown in equation (3):

[0057]

[0058] Where, φ s,y (x,t) represents the solid-state potential distribution, with subscripts y = n or y = p, where n corresponds to the negative electrode and p corresponds to the positive electrode; σ eff,y j is the equivalent conductivity in a solid material. tot (t) represents the total current density.

[0059] 4) Calculation of liquid phase potential

[0060] In the P2D model, the liquid phase potential distribution of the lithium-ion battery satisfies Ohm's law and the theory of concentrated solutions, and its expression is shown in equation (4):

[0061]

[0062] Where, φ e (x,t) represents the liquid phase potential; c e (x,t) represents the liquid phase ion concentration; R is the ideal gas constant; T is the temperature of the single cell; κ eff denoted as , where is the effective conductivity of the electrolyte; f is the ion activity coefficient of the electrolyte. tot (t) represents the total current density within the electrode material.

[0063] 5) Delithiation and Lithium Intercalation

[0064] Lithium insertion / extraction occurs at the electrolyte / electrode interface. The current density generated by the lithium insertion / extraction reaction on the particle surface can be calculated using the Butler-Volmer equation, as shown in equation (5):

[0065]

[0066] Where i 0,int η represents the electrode reaction exchange current density during the lithium insertion / extraction process. int,y This represents the reaction overpotential during the lithium insertion / extraction process. α a,int α is the anode transfer coefficient. c,int is the cathode transfer coefficient.

[0067] i 0,int The expression for is shown in equation (6):

[0068]

[0069] Where, k int c is the reaction rate parameter for the delithiation or lithium insertion process. e c represents the liquid phase ion concentration. e,ref c is the reference concentration of lithium ions in the electrolyte.s,y,max c is the maximum lithium intercalation concentration of the positive or negative electrode material. s,y,surf This represents the concentration of lithium ions on the surface of spherical particles of the positive or negative electrode material.

[0070] S2.2 Establish a battery characteristic prediction model

[0071] A battery characteristic prediction model based on the S2.1 power battery mechanistic model is established, integrating a metaheuristic algorithm and a physical information neural network. This model simulates the characteristic operating conditions of the input power battery to predict the concentrations of positive and negative electrodes and electrolyte ions under given input conditions. Specifically, the prediction bias of the power battery mechanistic model is introduced into the penalty function of the physical information neural network algorithm and assigned a weight greater than a weight threshold. Then, during training, a metaheuristic algorithm is used for parameter optimization. The metaheuristic algorithms used include particle swarm optimization-gray wolf optimization coupled algorithms, chaotic search algorithms, and particle swarm optimization-simulated annealing coupled algorithms. The physical information neural network algorithm includes long short-term memory neural network algorithms, graph neural network algorithms, or bidirectional long short-term memory neural network algorithms with attention mechanisms.

[0072] In this embodiment, as Figure 4 As shown, a battery characteristic prediction model based on a power battery mechanistic model is established using particle swarm optimization, simulated annealing, and a gated recurrent neural network with attention mechanism. In the penalty function of the bidirectional recurrent neural network, the prediction bias of the power battery mechanistic model (the error between the model prediction and the actual sampled voltage) is introduced and assigned a weight of 75%. Then, particle swarm optimization and simulated annealing are used to optimize the training process. Finally, the trained model is used to predict the positive and negative electrode and electrolyte ion concentrations of the power battery under a given input condition by simulating the characteristic operating conditions of the power battery.

[0073] The training parameters include: number of neural network layers, learning rate, gradient threshold, learning rate, slice size, and number of iterations.

[0074] The penalty function, which includes the prediction bias of the power battery mechanism model and the training residual of the neural network, is expressed in the following formula:

[0075] Γ total =ω F Γ F +ω D Γ D (7)

[0076] In the formula: Γ total This refers to the total penalty function, ω F It is the weight of the training residual of the neural network, Γ F It is the training residual of the neural network, ωD It is the weight of the prediction bias in the mechanistic model of power batteries, Γ D It is a prediction bias in the mechanistic model of power batteries.

[0077] S2.3 Model Prediction Available Energy Calculation Processing

[0078] like Figure 5 As shown, the SOC of the positive and negative electrodes is calculated on the cloud platform based on the predicted positive and negative electrode ion concentrations and the calibrated maximum ion concentrations of the positive and negative electrodes. Then, the positive electrode potential and negative electrode potential are calculated based on the calculated SOC and the calibrated SOC-OCV curves of the positive and negative electrodes. The electrolyte potential is calculated using the predicted electrolyte ion concentration. Then, the negative electrode potential, electrolyte potential, and ohmic voltage drop obtained based on the characteristic operating conditions of the power battery are subtracted from the positive electrode potential to obtain the power battery terminal voltage. The power battery power is obtained by using the power battery terminal voltage and the current obtained from the characteristic operating conditions of the power battery. When the power battery is continuously virtually discharged to the cutoff voltage, the available energy of the power battery is calculated by power-time integration, which is used as the model to predict the available energy.

[0079] S3. Conduct offline testing experiments.

[0080] Offline testing experiments may include capacity testing and dynamic charge-discharge testing, which measure the actual usable energy (also known as real usable energy) of the power battery under the same initial conditions of charge / discharge current / power, temperature, and initial SOC parameters. This may include the actual usable energy of the battery system under given test conditions (such as NEDC, CLTC, etc.), or the actual usable energy of the battery system under characteristic operating conditions, or the actual usable energy of a single battery cell under given test conditions (such as DST, FUDS, etc.).

[0081] In this embodiment, the battery operating condition established in step S1 is used as the test condition. The available energy over a period of time is obtained through battery testing equipment to measure the battery's actual available energy under the same test conditions. Offline testing uses the NEDC operating condition, and the power battery operating condition is as follows: Figure 6 As shown.

[0082] S4: Establish a data-driven model, or in other words, establish a corrected model for estimating the available energy of the power battery.

[0083] A data-driven approach is employed on the cloud platform. Using the model-predicted available energy obtained in step S2 as input and the actual tested available energy obtained in step S3 as output, training and testing sets are jointly established to construct a data-driven model. This model establishes a mapping relationship between the model-predicted available energy and the actual tested available energy through a data-driven method, thereby correcting the model-predicted available energy and achieving accurate available energy prediction. The data-driven model is essentially a correction model for estimating the available energy of the power battery. Data-driven methods can employ algorithms such as Extreme Learning Machine (ELM), Backpropagation (BP) neural networks, recurrent neural networks, Elman neural networks, or LSTM neural networks.

[0084] In this embodiment, an ELMAN neural network is used to establish a corrected model of the available energy of the power battery. The ELMAN neural network is highly sensitive to historical data, and its internal feedback network can increase the processing capability of dynamic information and has strong generalization ability. The nonlinear state-space expression of the ELMAN neural network is:

[0085]

[0086] Where y is the m-dimensional output node vector; x is the n-dimensional intermediate layer node unit vector; u is the r-dimensional input vector; x c Let w be an n-dimensional feedback state vector. 3 The connection weights from the intermediate layer to the output layer; w 2 Weights for the link from the input layer to the intermediate layer; w 1 The connection weights from the receiving layer to the intermediate layer are defined as follows: g(*) is the transfer function of the output neuron, and f(*) is the transfer function of the intermediate layer neuron. The battery available energy data from steps S2 and S3 are used as feature data, where the input feature is the predicted available energy and the output feature is the actual available energy. A sample matrix is ​​constructed using the two available energy values ​​as column vectors, dividing the data into a training set and a sample set. The training set is X%, and the sample set is 1-X%. In this embodiment, the training set is 70%, and the sample set is 30%. The ELMAN neural network is trained using the training set, with the sum of squared errors of the output as the index function. The weights are corrected using the backpropagation algorithm. A corrected model for estimating the available energy of the power battery (i.e., a trained machine learning corrected model – a data-driven model) is obtained using the above training data.

[0087]

[0088] Where y k (w) is the target observation vector. The target input vector is denoted as . In this embodiment, the Elman neural network is implemented in the algorithm engineering.

[0089] The results show that the corrected model has a significant corrective effect. The model trained under cloud data extraction conditions remains effective under real-world conditions. Before neural network correction, the maximum error was 2.7%. After neural network correction, the average error was <1%, and the maximum error was 1.6%. The calculation results are as follows: Figure 7 As shown.

[0090] This invention also relates to a knowledge-data fusion-driven power battery available energy prediction system, which corresponds to the aforementioned knowledge-data fusion-driven power battery available energy prediction method. It can be understood as a system that implements the aforementioned knowledge-data fusion-driven power battery available energy prediction method. The structure of this system is as follows: Figure 8 As shown, the system includes a power battery usage characteristic condition establishment module, a model prediction available energy calculation module, an actual test available energy testing module, and a data-driven model building module, all located on a cloud platform. The power battery usage characteristic condition establishment module is connected to the model prediction available energy calculation module, and both the model prediction available energy calculation module and the actual test available energy testing module are connected to the data-driven model building module.

[0091] The power battery usage characteristic condition establishment module acquires cloud data uploaded by the vehicle-side power battery system to the cloud platform, extracts the power battery usage condition characteristics by performing data analysis on the acquired cloud data, uses a data dimensionality reduction algorithm to reduce the dimensionality of the high-dimensional features of the power battery usage condition, performs correlation analysis on the dimensionality-reduced features based on physical knowledge, classifies them using a data-driven condition feature classification method, calculates the probability map of the power battery switching between different conditions, and uses a condition generation algorithm to establish the power battery usage characteristic conditions.

[0092] The model-predicted available energy calculation module includes a power battery mechanistic model construction module, a battery characteristic prediction model construction module, and a model-predicted available energy calculation processing module connected in sequence. The power battery mechanistic model construction module establishes a power battery mechanistic model based on the power battery's operating principles. The input conditions of the power battery mechanistic model are charging / discharging current / power, temperature, and initial SOC parameters. The output conditions of the power battery mechanistic model are the positive and negative electrode ion concentrations and the electrolyte ion concentration. The battery characteristic prediction model construction module establishes a battery characteristic prediction model based on the power battery mechanistic model, fusing a metaheuristic algorithm with a physical information neural network. This model simulates the characteristic operating conditions of the input power battery to predict the positive and negative electrode ion concentrations under given input conditions. The model predicts available energy calculation and processing module calculates the SOC of the positive and negative electrodes based on the predicted positive and negative electrode ion concentrations and the calibrated maximum ion concentrations of the positive and negative electrodes of the power battery. Then, it calculates the positive electrode potential and negative electrode potential based on the calculated SOC of the positive and negative electrodes and the calibrated SOC-OCV curves of the positive and negative electrodes. It calculates the electrolyte potential using the predicted electrolyte ion concentration. Then, it subtracts the negative electrode potential, electrolyte potential, and ohmic voltage drop obtained based on the characteristic operating conditions of the power battery from the positive electrode potential to obtain the power battery terminal voltage. The power battery power is obtained through the power battery terminal voltage and the current obtained from the characteristic operating conditions of the power battery. When the power battery is continuously virtually discharged to the cutoff voltage, the available energy of the power battery is calculated by power-time integration as the model predicted available energy.

[0093] The actual test usable energy test module conducts offline testing experiments, including capacity testing and dynamic charge-discharge testing, to measure the actual test usable energy of the power battery under the same initial conditions of charge-discharge current / power, temperature, and initial SOC parameters.

[0094] The data-driven model building module adopts a data-driven approach, using the model's predicted available energy as input and the actual tested available energy as output to jointly establish a training set and a test set, thereby constructing a data-driven model. The data-driven model establishes a mapping relationship between the model's predicted available energy and the actual tested available energy through the data-driven approach, thereby correcting the model's predicted available energy.

[0095] Furthermore, the power battery usage characteristic condition establishment module acquires cloud data uploaded by the vehicle-side power battery system to the cloud platform, including voltage, current, time, temperature, SOC, charge / discharge status, battery output power, battery fault status, and vehicle speed information. It then performs data analysis on the acquired cloud data to extract power battery usage characteristic features, including charging frequency, charge / discharge depth, acceleration time, deceleration time, average current, average voltage, average power, maximum power, minimum power, and power distribution curve features. The data dimensionality reduction algorithms employed include principal component analysis, partial principal component analysis, singular value decomposition, or autoencoder algorithms. The data-driven operating condition feature classification methods employed include K-means clustering, random forest, and support vector machine algorithms. Finally, the module establishes power battery usage characteristic conditions using operating condition generation algorithms, including Monte Carlo simulation, decision tree, and decision forest algorithms. These battery usage characteristic conditions include current-time operating condition curves, with curve lengths selected to be less than 1800 seconds.

[0096] Furthermore, the battery characteristic prediction model construction module in the model prediction available energy calculation module establishes a battery characteristic prediction model based on the power battery mechanistic model, which integrates a metaheuristic algorithm and a physical information neural network. This involves introducing the prediction bias of the power battery mechanistic model into the penalty function of the physical information neural network algorithm and assigning it a weight greater than a weight threshold. Then, during training, a metaheuristic algorithm is used for parameter optimization. The metaheuristic algorithm used is a particle swarm optimization-gray wolf optimization coupled algorithm, a chaotic search algorithm, or a particle swarm optimization-simulated annealing coupled algorithm. The physical information neural network algorithm is a long short-term memory neural network algorithm, a graph neural network algorithm, or a bidirectional long short-term memory neural network algorithm with an attention mechanism.

[0097] The present invention relates to a knowledge-data fusion-driven method and system for predicting the available energy of power batteries. It successively establishes a battery mechanistic model, a battery characteristic prediction model, and a data-driven model based on reliable data, all conforming to the basic principles of power batteries. The mechanistic model and battery characteristic prediction model achieve battery available energy prediction in open-loop mode, while the data-driven model achieves available energy correction under reliable data conditions. This achieves a fusion of knowledge-driven and data-driven methods for predicting the available energy of power batteries. Testing has shown that the method and system of this invention can improve the accuracy and scenario coverage of battery available energy prediction, reduce the uninterpretability of the data-driven model, and improve the problem of poor open-loop prediction accuracy of the mechanistic model. It has significant application value in improving the accuracy of range estimation for new energy vehicles.

[0098] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail with reference to the accompanying drawings and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention. In short, all technical solutions and improvements that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention patent.

Claims

1. A knowledge-data fusion driven power battery available energy prediction method, characterized in that, Comprise the following steps; S1, obtain the cloud data uploaded to the cloud platform by the vehicle end power battery system, extract the use condition characteristics of the power battery by data analysis on the obtained cloud data in the cloud platform, reduce the high-dimensional characteristics of the use condition of the power battery by using data dimension reduction algorithm, analyze the correlation of the reduced characteristics according to physical knowledge, and classify by using data driven condition characteristic classification method, calculate the probability graph of the power battery jumping between different conditions, and establish the use condition characteristics of the power battery by using condition generation algorithm; S2, in the cloud platform, according to the use principle of power battery, the input condition of the power battery mechanism model is established, the input condition of the power battery mechanism model is charge / discharge current / power, temperature, SOC initial value parameter, the output condition of the power battery mechanism model is power battery positive and negative ion concentration and electrolyte ion concentration, based on the power battery mechanism model, the battery characteristic prediction model of meta heuristic algorithm and physical information neural network fusion is established, the use condition characteristics of the power battery are simulated by the battery characteristic prediction model, the positive and negative ion concentration and electrolyte ion concentration of the power battery under the given input condition are predicted; In the cloud platform, according to the predicted positive and negative ion concentration and the calibrated maximum ion concentration of the positive and negative electrodes of the power battery, the SOC of the positive and negative electrodes is calculated, and then the positive and negative electrode potentials are calculated according to the calculated SOC of the positive and negative electrodes and the calibrated positive and negative electrode SOC-OCV curve. The electrolyte potential is calculated by using the predicted electrolyte ion concentration. Then, the negative electrode potential, the electrolyte potential and the ohmic voltage drop obtained according to the use characteristics of the power battery are subtracted from the positive electrode potential respectively, so as to obtain the terminal voltage of the power battery, and the power of the power battery is obtained by the terminal voltage of the power battery and the current obtained from the use characteristics of the power battery. When the power battery is continuously virtually discharged to the cut-off voltage, the available energy of the power battery is calculated by power-time integration, which is the model predicted available energy; S3, carry out offline detection experiment, including capacity test, dynamic charge and discharge test, measure the actual test available energy of the power battery under the same initial conditions of charge / discharge current / power, temperature, SOC initial value parameter; S4, in the cloud platform, using data driven method, taking the model predicted available energy obtained in step S2 as input, and taking the actual test available energy obtained in step S3 as output, the training set and test set are established together, the data driven model is constructed, the mapping relationship from the model predicted available energy to the actual test available energy is established by the data driven method, and the correction of the model predicted available energy is realized. 2.The knowledge-data fusion driven power battery usability prediction method according to claim 1, wherein, S1 step acquires the cloud data uploaded to the cloud platform by the vehicle end power battery system, including voltage, current, time, temperature, SOC, charge / discharge state, battery output power, battery fault state, vehicle speed information; the use condition characteristics of the power battery are extracted by data analysis on the acquired cloud data in the cloud platform, including charging frequency, charging and discharging depth, acceleration time, deceleration time, average current, average voltage, average power, maximum power, minimum power, power distribution curve characteristics; the use condition generation algorithm is adopted to establish the power battery use condition characteristics, and the battery use condition characteristics include a current-time condition curve, and the curve length selects a characteristic condition below 1800s.

3. The knowledge-data fusion driven power battery usability-ability prediction method according to claim 2, characterized in that, The data dimension reduction algorithm adopted in S1 step includes principal component analysis or partial principal component analysis or singular value decomposition or autoencoder algorithm; the data-driven condition characteristic classification method adopted includes K-means clustering algorithm, random forest algorithm, support vector machine algorithm; the condition generation algorithm adopted includes Monte Carlo simulation method, decision tree algorithm, decision forest algorithm.

4. The knowledge-data fusion driven power battery usability prediction method according to any one of claims 1 to 3, characterized in that, S2 step establishes a power battery mechanism model including an equivalent circuit model or an electrochemical model or a heterogeneous model in the cloud platform according to the power battery use principle, the equivalent circuit model is a first-order equivalent circuit model, a second-order equivalent circuit model, a fractional-order equivalent circuit model or other electrochemical reduced-order equivalent circuit model, the electrochemical model is a pseudo two-dimensional electrochemical model, a single particle model or a three-dimensional electrochemical model, the pseudo two-dimensional electrochemical model fuses solid-phase lithium ion concentration calculation, liquid-phase lithium ion concentration calculation, solid-phase potential calculation, liquid-phase potential calculation and lithium deintercalation.

5. The knowledge-data fusion driven power battery usability-ability prediction method according to claim 4, characterized in that, S2 step is to establish a battery characteristic prediction model based on the meta-heuristic algorithm and the physical information neural network fusion based on the power battery mechanism model, that is, the prediction deviation of the power battery mechanism model is introduced into the penalty function of the physical information neural network algorithm, and a weight greater than a weight threshold is given, and then the parameter optimization is carried out in the training process by using the meta-heuristic algorithm; the meta-heuristic algorithm adopted is a particle swarm optimization-gray wolf optimization coupling algorithm, a chaos search algorithm, a particle swarm optimization-simulated annealing coupling algorithm; the physical information neural network algorithm is a long short-term memory neural network algorithm, a graph neural network algorithm or a bidirectional long short-term memory neural network algorithm with attention mechanism.

6. The knowledge-data fusion driven power battery usability-ability prediction method according to claim 5, characterized in that, S3 step carries out offline detection experiments to measure the actual test available energy of the power battery under the same initial conditions of charging and discharging current / power, temperature and SOC initial value parameters, including the actual test available energy of the power battery system under the NEDC or CLTC test condition, or the actual test available energy of the power battery system under the use condition characteristics, or the actual test available energy of the power battery monomer under the DST, FUDS test condition.

7. The knowledge-data fusion driven power battery usability prediction method according to any one of claims 1 to 3, characterized in that, The data-driven model constructed in S4 step establishes a mapping relationship between the model predicted available energy and the actual test available energy by using a data-driven method, and the data-driven method adopts an extreme learning machine algorithm, a BP neural network, a recurrent neural network, an ELMAN neural network or an LSTM neural network method.

8. A knowledge-data fusion driven power battery available energy prediction system, characterized in that, The power battery use characteristic working condition establishment module, the model predicted available energy calculation module, the actual test available energy test module and the data driven model construction module are all located in the cloud platform. The power battery use characteristic working condition establishment module obtains cloud data uploaded by a vehicle end power battery system to the cloud platform, extracts power battery use working condition characteristics through data analysis on the obtained cloud data, reduces dimensions of the high-dimensional characteristics of the power battery use working condition by using a data dimension reduction algorithm, analyzes the correlation of the reduced characteristics according to physical knowledge, classifies the characteristics by using a data driven working condition characteristic classification method, calculates a probability graph of the power battery jumping between different working conditions, and establishes power battery use characteristic working conditions by using a working condition generation algorithm. The model predicted available energy calculation module includes a power battery mechanism model construction module, a battery characteristic prediction model construction module and a model predicted available energy calculation processing module connected in sequence. The power battery mechanism model construction module establishes a power battery mechanism model according to a power battery use principle. The input conditions of the power battery mechanism model are charging and discharging current / power, temperature and SOC initial value parameters. The output conditions of the power battery mechanism model are positive and negative electrode ion concentrations and electrolyte ion concentration. The battery characteristic prediction model construction module establishes a battery characteristic prediction model based on the power battery mechanism model, which is a fusion of a meta-heuristic algorithm and a physical information neural network. The battery characteristic prediction model simulates input power battery use characteristic working conditions to predict positive and negative electrode ion concentrations and electrolyte ion concentration of the power battery under given input working conditions. The model predicted available energy calculation processing module calculates the SOC of the positive and negative electrodes according to the predicted positive and negative electrode ion concentrations and the calibrated maximum positive and negative electrode ion concentrations. Then, the model predicted available energy calculation processing module calculates the positive electrode potential and the negative electrode potential according to the calculated SOC of the positive and negative electrodes and the calibrated positive and negative electrode SOC-OCV curves. The model predicted available energy calculation processing module calculates the electrolyte potential using the predicted electrolyte ion concentration. Then, the model predicted available energy calculation processing module obtains the power battery terminal voltage by subtracting the negative electrode potential, the electrolyte potential and the ohmic voltage drop obtained according to the power battery use characteristic working condition from the positive electrode potential. The model predicted available energy calculation processing module obtains the power battery power by using the power battery terminal voltage and the current obtained according to the power battery use characteristic working condition. The model predicted available energy calculation processing module calculates the available energy of the power battery by power-time integration when the power battery is continuously virtually discharged to the cut-off voltage, as the model predicted available energy. The actual test available energy test module carries out offline detection experiments, including capacity test and dynamic charging and discharging test, to measure the actual test available energy of the power battery under the same initial conditions of charging and discharging current / power, temperature and SOC initial value parameters. The data-driven model construction module adopts a data-driven method to jointly establish a training set and a test set with model predicted available energy as input and actual test available energy as output, and to construct a data-driven model.

9. The knowledge-data fusion driven power battery usability-ability prediction system according to claim 8, wherein, The power battery use feature working condition establishment module acquires cloud end data uploaded by a vehicle end power battery system to a cloud platform, including voltage, current, time, temperature, SOC, charge / discharge state, battery output power, battery fault state, and vehicle speed information; uses data analysis on the acquired cloud end data to extract power battery use working condition features, including charging frequency, charge / discharge depth, acceleration time, deceleration time, average current, average voltage, average power, maximum power, minimum power, and power distribution curve features; uses a data dimension reduction algorithm including principal component analysis or partial principal component analysis or singular value decomposition or autoencoder algorithm; uses a data-driven working condition feature classification method including K-means clustering algorithm, random forest algorithm, and support vector machine algorithm; uses a working condition generation algorithm to establish power battery use feature working conditions, and the working condition generation algorithm includes Monte Carlo simulation method, decision tree algorithm, and decision forest algorithm; and the battery use feature working conditions include current-time working condition curves, and the curve length selects a feature working condition below 1800s.

10. The knowledge-data fusion driven, power battery usability-ability-to- perform predictability system according to claim 8 or 9, characterized in that, The battery characteristic prediction model construction module in the model predicted available energy calculation module establishes a battery characteristic prediction model based on a meta-heuristic algorithm and a physical information neural network fused power battery mechanism model, introduces a power battery mechanism model prediction deviation into a penalty function of a physical information neural network algorithm, and gives a weight greater than a weight threshold value, and then uses a meta-heuristic algorithm for parameter optimization in a training process; the meta-heuristic algorithm is a particle swarm optimization-gray wolf optimization coupling algorithm, a chaos search algorithm, or a particle swarm optimization-simulated annealing coupling algorithm; and the physical information neural network algorithm is a long short-term memory neural network algorithm, a graph neural network algorithm, or a bidirectional long short-term memory neural network algorithm with an attention mechanism.

Images