Neural network algorithm fusing physical information for lithium battery SOH estimation

Abstract

The invention discloses a neural network algorithm fused with physical information for lithium battery SOH estimation, and particularly relates to the technical field of an electrochemical energy storage system battery management system, a data driving method is widely applied to the field of lithium ion battery (LIB) state of health (SOH) estimation, but the methods are generally lack of interpretability. In order to solve the problem, a physical information neural network method is provided to enhance the interpretability and accuracy of the neural network in the aspect of SOH prediction. The extracted features are used as the input of the physical information network, testing is carried out on the energy storage battery cabinet, and the verification result shows that the prediction error RMSE of the method is within 0.4%.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management systems for electrochemical energy storage systems, and more specifically, to a neural network algorithm integrating physical information for lithium battery SOH estimation. Background Art

[0002] Neural network algorithms that incorporate physics information offer an effective solution to the lack of interpretability inherent in purely data-driven models. These algorithms combine physical principles with data-driven approaches, leveraging prior physical knowledge to guide and enhance the performance of machine learning models. Existing deep learning methods have achieved promising results in dealing with complex nonlinear systems, but they lack interpretability. Explanation methods in machine learning, particularly deep learning, have attracted widespread attention in recent years. These techniques are primarily categorized as ex ante and ex post interpretation. Ex ante interpretation methods design models to be structurally interpretable. Deep unfolding constructs neural networks as a mirror image of iterative optimization algorithms or physical processes, providing clear explanations for the model's decision-making process. Ex post interpretation methods use visualization techniques to interpret trained models, revealing which input features are most critical for model predictions. While these methods can improve interpretability, they often require additional analytical tools. In contrast, physics-integrated neural networks inherently enhance model interpretability by embedding physical constraints during training. These methods directly incorporate real-world physical laws into the model architecture, making predictions more physically consistent. Summary of the Invention

[0003] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a neural network algorithm integrating physical information for lithium battery SOH estimation to solve the problems raised in the above-mentioned background technology.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A neural network algorithm integrating physical information is used for lithium battery SOH estimation, including the following steps:

[0006] Step S1: Collecting discharge cycle data and preprocessing it, integrating the current and time data using the trapezoidal method to calculate the capacity of the discharge cycle; and calculating the IC curve using the finite difference method;

[0007] Step S2: performing Kalman filtering on the IC curve, dynamically adjusting the noise covariance matrix using an adaptive method, and extracting peak height, peak position, and peak area from the IC curve;

[0008] Step S3: Build an LSTM neural network based on the Anaconda environment, modify the loss function to incorporate physical information, optimize the loss function by adjusting the ratio of physical loss to original loss, and finally combine the two to minimize the loss; use a random search algorithm to search and select the optimal model;

[0009] Step S4: The trained neural network model is adjusted to a prediction mode. When new discharge cycle characteristic parameters are extracted, the extracted parameters are automatically put into the neural network for prediction; and the prediction results are output in real time.

[0010] In a preferred embodiment, in step S1, voltage, current, and time data are automatically extracted after the discharge cycle is completed;

[0011] The capacity increment curve is calculated according to the following formula: Q = It; V = f(Q), Q = f -1 (V); where I is the charging current, t is the charging time, and V is the battery voltage;

[0012] The IC curve can be expressed as follows: Use the finite difference method to approximate the IC curve: Where Qk and Qk-1 represent the charge capacity of the battery at time steps k and k-1, respectively, and Vk and Vk-1 represent the terminal voltage of the battery at time steps k and k-1, respectively.

[0013] In a preferred embodiment, in step S2, a Kalman filter process is applied to the calculated IC curve; the following is a dynamic adjustment formula that integrates these factors: Where Q0 is the initial process noise covariance; is the predicted state estimate; x k is the actual system state; is the residual, z k is the actual measurement value, H is the measurement matrix; γ1 and γ2 are adjustment coefficients, which control the weight of the influence of error and residual on Q; R0 is the initial measurement noise covariance; is the predicted state estimate; x k is the actual system state (which can be obtained through measurement or some reference value); is the residual; α1 and α2 are adjustment coefficients, which control the impact of error and residual on R.

[0014] In a preferred embodiment, in step S2, feature extraction is performed on the IC curve to find the highest value y of the curve. h , the horizontal coordinate V corresponding to the highest value h , take a fixed voltage width ΔV = 0.1V, and get V according to the following formulal and V r The peak area Q is obtained by integrating the peak interval. G , so far we get three characteristic value peak heights (y h ), peak position (V h ) and peak area (Q G );

[0015] In a preferred embodiment, in step S3, the construction of the physical information neural network is integrated, and an LSTM neural network is built based on the Anaconda environment. The loss function is modified so that the neural network structure incorporates physical information. The specific modification follows the following formula: in Represents the neural network model for x i The predicted value of represents the first eigenvalue, represents the second eigenvalue, represents the third eigenvalue, d(x i ) represents the physical constraints learned based on the three eigenvalues; L d is the original loss function in the neural network. After adding three physical constraint formulas, the optimal solution can be expressed as: Finally, the physical constraint loss function can be expressed as follows: minL train =min(L d +ω1L p ).

[0016] In a preferred embodiment, after embedding the physical constraints into the neural network in step S3, the weight distribution between the loss functions is also involved. The physical constraint loss and the original model loss are reasonably distributed during the test phase, and the two are adjusted by adjusting the penalty factor to optimize the total loss function; L' train =L train +L test ;L train =L d +ω1L p ;L test =ω2L' p ; where L p represents the physical loss of the neural network's prediction results on the training set, and L' p Represents the physical loss of the neural network's prediction results on the test set.

[0017] In a preferred embodiment, a random search algorithm is used to search in step S3, and the model with the smallest loss function is saved as the optimal model; specifically, the number of searches of the search algorithm is calculated according to the dimension of the search space and the roughness of the objective function.

[0018] In a preferred embodiment, the dimension of the search space is determined according to the number of decision variables in the problem; the roughness of the objective function can be expressed by the standard deviation as follows: in is the mean of the objective function value, κ is the roughness;

[0019] Determine the dimension of the search space and the roughness of the objective function, and calculate the number of searches using the following formula: T = C*n χ *κ δ ; T is the number of searches; C is a constant that depends on the characteristics of the specific problem and the algorithm; n is the dimension of the problem; χ and δ are constants that represent the degree of influence of dimension and roughness on the number of searches, respectively; κ is the roughness of the objective function.

[0020] In a preferred embodiment, in step S4, the mode of the neural network is adjusted to a prediction mode. When new discharge cycle characteristic parameters are extracted, the extracted parameters are automatically put into the neural network for prediction, and the prediction results are finally displayed on the edge interface.

[0021] In a preferred embodiment, the system comprises the following modules: a physical information extraction and processing module, a physical information neural network prediction module, a confirmation module, a result visualization module, and a storage module;

[0022] The physical information extraction and processing module is used to collect and preprocess discharge cycle data, integrate current and time data using the trapezoidal method, and calculate the discharge cycle capacity; use the finite difference method to calculate the IC curve; then perform Kalman filtering on the IC curve, dynamically adjust the noise covariance matrix using an adaptive method, and extract the peak height, peak position, and peak area from the IC curve; the peak height, peak position, and peak area are transmitted to the physical information neural network prediction module and confirmation module;

[0023] The physical information neural network prediction module builds an LSTM neural network based on the Anaconda environment, modifies the loss function to incorporate physical information, optimizes the loss function by adjusting the ratio of physical loss to original loss, and ultimately combines the two to minimize loss. It uses a random search algorithm to search and select the optimal model, and transmits the optimal model to the confirmation module.

[0024] The confirmation module uses the trained optimal model to put the three collected IC curve features into the neural network to predict the health status of the lithium battery; and transmits the prediction results to the result visualization module;

[0025] The result visualization module is used to present the prediction of lithium battery health status;

[0026] The storage module is used to save the data generated during the system operation process.

[0027] The technical effects and advantages of the present invention are as follows:

[0028] 1. The neural network integrated with the physical model enhances the interpretability of the model, proving that this method has more advantages in predicting the health status of lithium-ion batteries.

[0029] 2. Extract features through electrochemical knowledge and find the connection between them and SOH, describe this relationship using mathematical expressions and embed it into the neural network.

[0030] 3. The neural network established by this method achieved more accurate results in predicting the SOH of lithium batteries.

[0031] 4. This neural network-integrated algorithm can be applied to predict the health status of batteries in energy storage cabinets, accurately tracking the health status changes of lithium-ion batteries throughout their life cycle in real time, thereby improving battery management efficiency and extending battery life. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0033] Figure 1 This is a flow chart of a neural network algorithm integrating physical information for lithium battery SOH estimation according to the present invention;

[0034] Figure 2 This is a schematic diagram of the structure of a neural network algorithm that integrates physical information and is used for lithium battery SOH estimation in the present invention;

[0035] Figure 3 This is a flowchart of a neural network algorithm that integrates physical information and is used for estimating the SOH of lithium batteries.

[0036] Figure 4 Schematic diagram of IC curve and characteristic parameter extraction. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0038] The present invention collects and preprocesses discharge cycle data, integrates current and time data using the trapezoidal method, and calculates the capacity of the discharge cycle; uses the finite difference method to calculate the IC curve; then performs Kalman filtering on the IC curve, uses an adaptive method to dynamically adjust the noise covariance matrix, and extracts peak height, peak position, and peak area from the IC curve; builds an LSTM neural network based on the Anaconda environment, modifies the loss function to incorporate physical information, optimizes the loss function by adjusting the ratio of physical loss to original loss, and finally combines the two to minimize loss; uses a random search algorithm to search and select the optimal model; adjusts the trained neural network model to a prediction mode, and when new discharge cycle characteristic parameters are extracted, automatically puts the extracted parameters into the neural network for prediction; and outputs the prediction results in real time.

[0039] Example 1

[0040] The present invention provides Figure 1 and Figure 3 A neural network algorithm integrating physical information is shown for lithium battery SOH estimation. The specific steps are as follows:

[0041] Step S1: Collecting discharge cycle data and preprocessing it, integrating the current and time data using the trapezoidal method to calculate the capacity of the discharge cycle; and calculating the IC curve using the finite difference method;

[0042] Step S2: performing Kalman filtering on the IC curve, dynamically adjusting the noise covariance matrix using an adaptive method, and extracting peak height, peak position, and peak area from the IC curve;

[0043] Step S3: Build an LSTM neural network based on the Anaconda environment, modify the loss function to incorporate physical information, optimize the loss function by adjusting the ratio of physical loss to original loss, and finally combine the two to minimize the loss; use a random search algorithm to search and select the optimal model;

[0044] Step S4: The trained neural network model is adjusted to a prediction mode. When new discharge cycle characteristic parameters are extracted, the extracted parameters are automatically put into the neural network for prediction; and the prediction results are output in real time.

[0045] In step S1, voltage, current and time data are automatically extracted after the discharge cycle is completed.

[0046] The capacity increment curve is calculated according to the following formula: Q = It; V = f(Q), Q = f -1 (V); where I is the charging current, t is the charging time, and V is the battery voltage. The IC curve can be expressed as follows: In order to reduce the impact of noise on the IC curve calculation, from a mathematical point of view, the finite difference method is used to approximate the IC curve:

[0047] After analysis, the error is smallest when the differential value is 10mv, so the IC curve when the differential value is 10mv is used as the final result; where Qk and Qk-1 represent the charging capacity of the battery at time steps k and k-1 respectively, and Vk and Vk-1 represent the terminal voltage of the battery at time steps k and k-1 respectively.

[0048] In step S2, to further reduce the impact of noise, a Kalman filter is applied to the calculated IC curve. The process noise covariance and measurement noise covariance are dynamically adjusted based on the system state error and residual. Generally, larger prediction errors indicate greater system uncertainty, requiring an increase in the process noise covariance. Conversely, larger measurement errors indicate unreliable measurement signals, requiring an increase in the measurement noise covariance.

[0049] Against this backdrop, here’s a dynamic adjustment formula that incorporates these factors: Where Q0 is the initial process noise covariance; is the predicted state estimate; x k is the actual system state (which can be obtained through measurement or some reference value); is the residual, z k is the actual measurement value, H is the measurement matrix; γ1 and γ2 are adjustment coefficients, which control the weight of the influence of error and residual on Q. R0 is the initial measurement noise covariance; is the predicted state estimate; x k is the actual system state (which can be obtained through measurement or some reference value); is the residual; α1 and α2 are adjustment coefficients, which control the impact of error and residual on R.

[0050] In summary, the system state error It reflects the deviation between the predicted system state and the actual state. A larger state error means that the system dynamics change is larger or the model is inaccurate, and it may be necessary to increase the process noise covariance Q K and the measurement noise covariance Rk ; Residual(||v k || 2 ) reflects the difference between the measured value and the predicted value. A larger residual usually indicates a larger measurement noise, so it is necessary to increase the measurement noise covariance R k At the same time, a larger residual may also indicate an increase in system uncertainty, requiring an appropriate increase in the process noise covariance Q K .

[0051] If the process noise covariance matrix is set to 0.01 and the measurement noise covariance matrix is set to 0.1, the final IC curve is as follows Figure 4 shown.

[0052] Extract features from the IC curve and find the highest value y of the curve h , the horizontal coordinate V corresponding to the highest value h , take a fixed voltage width ΔV = 0.1V, and get V according to the following formula l and V r The peak area Q is obtained by integrating the peak interval. G , so far we get three characteristic value peak heights (y h ), peak position (V h ) and peak area (Q G ); The three features of each discharge cycle calculated are saved in a fixed directory in the form of CSV files. The discharge cycle characteristic parameters are peak height (y h ), peak position (V h ) and peak area (Q G ).

[0053] In step S3, the construction of the neural network integrating physical information is carried out. An LSTM neural network is built based on the Anaconda environment, and the loss function is modified so that the neural network structure incorporates physical information. The specific modification follows the following formula: in Represents the neural network model for x i The predicted value of represents the first eigenvalue, represents the second eigenvalue, represents the third eigenvalue, d(x i ) represents the physical constraints learned based on the three eigenvalues. L d is the original loss function in the neural network. After adding three physical constraint formulas, the optimal solution can be expressed as: Finally, the physical constraint loss function can be expressed as follows: minL train=min(L d +ω1L p );

[0054] Further embedding physical constraints into the neural network also involves weight distribution between loss functions. During the test phase, the physical constraint loss and the original model loss are reasonably distributed, and the two are adjusted by adjusting the penalty factor to optimize the total loss function. train =L train +L test ;L train =L d +ω1L p ;L test =ω2L' p ; where L p Represents the physical loss of the neural network's prediction results on the training set, which is used to measure the degree to which the model violates physical constraints during training, while L' p Represents the physical loss of the neural network's prediction results on the test set.

[0055] This method can ensure that the prediction results of the model in the test phase are more consistent with the physical laws, while preventing the model from losing the knowledge learned during the training process. p and L' p , fine-tune the model. This method can effectively improve the prediction accuracy and interpretability of the model, making it more consistent with physical laws and more reliable in practical application scenarios. p When optimized to the minimum, the physical loss and the original loss function L d Combine them as the final loss function. Set the hyperparameter ω1 and adjust the ratio of the two losses during training to achieve the optimal model.

[0056] Furthermore, a random search algorithm is used for searching, and each search is trained for 1500 rounds. The model with the smallest loss function is saved as the optimal model. Specifically, the number of searches of the search algorithm is calculated according to the dimension of the search space and the "roughness" of the objective function.

[0057] Furthermore, the dimension of the search space is determined by the number of decision variables in the problem; for example, if the objective function has only one variable, the dimension of the search space is 1; if the objective function has n decision variables, the dimension of the search space is n.

[0058] The roughness of the objective function is usually used to measure the volatility or non-smoothness of the objective function in the search space. Specifically, the higher the roughness, the more drastic the local minima, maxima or changes of the objective function, and the more difficult the optimization problem is to solve. There is no single standardized method for calculating the roughness of the objective function, which usually depends on the optimization algorithm used and the characteristics of the problem; a simple roughness measurement method is to calculate the standard deviation of the objective function value, especially in a certain area. Suppose we select several points x1,x2,......,x in the search space. n ; and calculate the objective function value of each point f(x1),f(x2),......,f(x n ); the roughness of the objective function can be expressed by the standard deviation as follows: in is the mean of the objective function value, and κ is the roughness. If the standard deviation of the objective function in a certain area is large, it means that the fluctuation in this area is large, that is, the objective function is relatively rough.

[0059] Determine the dimension of the search space and the roughness of the objective function, and calculate the number of searches using the following formula: T = C*n χ *κ δ ; T is the number of searches; C is a constant that depends on the characteristics of the specific problem and the algorithm; n is the dimension of the problem; χ and δ are constants that represent the degree of influence of dimension and roughness on the number of searches, respectively; κ is the roughness of the objective function.

[0060] In step S4, the mode of the neural network is adjusted to the prediction mode. When new discharge cycle characteristic parameters are extracted, the extracted parameters are automatically put into the neural network for prediction, and the prediction results are finally displayed on the edge interface.

[0061] The present invention provides Figure 2 A neural network algorithm integrating physical information is shown for lithium battery SOH estimation, which specifically includes the following modules: physical information extraction and processing module, physical information neural network prediction module, confirmation module, result visualization module, and storage module;

[0062] The physical information extraction and processing module is used to collect and preprocess discharge cycle data, integrate current and time data using the trapezoidal method, and calculate the discharge cycle capacity; use the finite difference method to calculate the IC curve; then perform Kalman filtering on the IC curve, dynamically adjust the noise covariance matrix using an adaptive method, and extract the peak height, peak position, and peak area from the IC curve; the peak height, peak position, and peak area are transmitted to the physical information neural network prediction module and confirmation module;

[0063] The physical information neural network prediction module builds an LSTM neural network based on the Anaconda environment, modifies the loss function to incorporate physical information, optimizes the loss function by adjusting the ratio of physical loss to original loss, and ultimately combines the two to minimize loss. It uses a random search algorithm to search and select the optimal model, and transmits the optimal model to the confirmation module.

[0064] The confirmation module uses the trained optimal model to put the three collected IC curve features into the neural network to predict the health status of the lithium battery; and transmits the prediction results to the result visualization module;

[0065] The result visualization module is used to present the prediction of lithium battery health status;

[0066] The storage module is used to save the data generated during the system operation process.

[0067] Example 2

[0068] In order to verify the prediction effect of the present invention, the prediction model is deployed on the edge device to automatically predict the health status of lithium batteries. The iESS-CAB-W215H smart energy storage air-cooled integrated cabinet, which is composed of 240 lithium iron phosphate batteries with a rated capacity of 280Ah produced by Ruipu Lanjun Energy Co., Ltd., is used as the research object. 100 actual charge and discharge cycles of energy storage are simulated as an experiment. The average MSE error of these 100 experiments is within 0.4% for the 240 battery cells.

[0069] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0070] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0071] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0072] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0073] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A neural network algorithm integrating physical information is used for lithium battery SOH estimation, characterized in that: The following steps are involved: Step S1: collecting discharge cycle data and preprocessing it, integrating the current and time data using the trapezoidal method, and calculating the capacity of the discharge cycle; IC curves were calculated using the finite difference method; Step S2: performing Kalman filtering on the IC curve, dynamically adjusting the noise covariance matrix using an adaptive method, and extracting peak height, peak position, and peak area from the IC curve; Step S3: Build an LSTM neural network based on the Anaconda environment, modify the loss function to incorporate physical information, optimize the loss function by adjusting the ratio of physical loss to original loss, and finally combine the two to minimize the loss; use a random search algorithm to search and select the optimal model; Step S4: The trained neural network model is adjusted to a prediction mode. When new discharge cycle characteristic parameters are extracted, the extracted parameters are automatically put into the neural network for prediction; and the prediction results are output in real time.

2. The neural network algorithm integrating physical information for lithium battery SOH estimation according to claim 1 is characterized in that: In step S1, voltage, current, and time data are automatically extracted when the discharge cycle is completed; The capacity increment curve is calculated according to the following formula: Q = It; V = f(Q), Q = f -1 (V); where I is the charging current, t is the charging time, and V is the battery voltage; The IC curve can be expressed as follows: Use the finite difference method to approximate the IC curve: Where Qk and Qk-1 represent the battery charge capacity at time steps k and k-1, respectively, and Vk and Vk-1 represent the battery terminal voltage at time steps k and k-1, respectively.

3. The neural network algorithm integrating physical information for lithium battery SOH estimation according to claim 1 is characterized in that: In step S2, Kalman filtering is applied to the calculated IC curve; Here is a dynamic adjustment formula that combines these factors: Where Q0 is the initial process noise covariance; is the predicted state estimate; x k is the actual system state; is the residual, z k is the actual measurement value, H is the measurement matrix; γ1 and γ2 are adjustment coefficients, which control the weight of the influence of error and residual on Q; R0 is the initial measurement noise covariance; is the predicted state estimate; x k is the actual system state (which can be obtained through measurement or some reference value); is the residual; α1 and α2 are adjustment coefficients that control the impact of error and residual on R.

4. The neural network algorithm integrating physical information for lithium battery SOH estimation according to claim 1 is characterized in that: In step S2, the IC curve is feature extracted to find the highest value y of the curve h , the horizontal coordinate V corresponding to the highest value h , take a fixed voltage width ΔV = 0.1V, and get V according to the following formula l and V r The peak area Q is obtained by integrating the peak interval. G , so far we get three characteristic value peak heights (y h ), peak position (V h ) and peak area (Q G ); 5. The neural network algorithm integrating physical information for lithium battery SOH estimation according to claim 1 is characterized in that: In step S3, the construction of the neural network integrating physical information is carried out. An LSTM neural network is built based on the Anaconda environment, and the loss function is modified so that the neural network structure incorporates physical information. The specific modification follows the following formula: in Represents the neural network model for x i The predicted value of represents the first eigenvalue, represents the second eigenvalue, represents the third eigenvalue, d(x i ) represents the physical constraints learned based on the three eigenvalues; L d is the original loss function in the neural network. After adding three physical constraint formulas, the optimal solution can be expressed as: Finally, the physical constraint loss function can be expressed as follows: minL train =min(L d +ω1L p ).

6. The neural network algorithm integrating physical information for lithium battery SOH estimation according to claim 1 is characterized in that: After embedding the physical constraints into the neural network in step S3, the weight distribution between the loss functions is also involved. The physical constraint loss and the original model loss are reasonably distributed during the test phase, and the two are adjusted by adjusting the penalty factor to optimize the total loss function. train =L train +L test ;L train =L d +ω1L p ;L test =ω2L' p ; where L p represents the physical loss of the neural network's prediction results on the training set, and L' p Represents the physical loss of the neural network's prediction results on the test set.

7. The neural network algorithm integrating physical information for lithium battery SOH estimation according to claim 1 is characterized in that: In step S3, a random search algorithm is used to search, and the model with the smallest loss function is saved as the optimal model. Specifically, the number of searches of the search algorithm is calculated according to the dimension of the search space and the roughness of the objective function.

8. The neural network algorithm integrating physical information for lithium battery SOH estimation according to claim 7 is characterized in that: The dimension of the search space is determined by the number of decision variables in the problem; the roughness of the objective function can be expressed in terms of the standard deviation as follows: in is the mean of the objective function value, κ is the roughness; Determine the dimension of the search space and the roughness of the objective function, and calculate the number of searches using the following formula: T = C*n χ *κ δ ; T is the number of searches; C is a constant that depends on the characteristics of the specific problem and the algorithm; n is the dimension of the problem; χ and δ are constants that represent the degree of influence of dimension and roughness on the number of searches, respectively; κ is the roughness of the objective function.

9. The neural network algorithm integrating physical information for lithium battery SOH estimation according to claim 1 is characterized in that: In step S4, the mode of the neural network is adjusted to the prediction mode. When new discharge cycle characteristic parameters are extracted, the extracted parameters are automatically put into the neural network for prediction, and the prediction results are finally displayed on the edge interface.

10. The neural network algorithm integrating physical information for lithium battery SOH estimation according to claim 1 is characterized in that: It includes the following modules: physical information extraction and processing module, physical information neural network prediction module, confirmation module, result visualization module, and storage module; The physical information extraction and processing module is used to collect and pre-process the discharge cycle data, integrate the current and time data using the trapezoidal method, and calculate the capacity of the discharge cycle; The IC curve is calculated using the finite difference method; the IC curve is then Kalman filtered, and the noise covariance matrix is dynamically adjusted using an adaptive method to extract the peak height, peak position, and peak area from the IC curve; the peak height, peak position, and peak area are transmitted to the physical information neural network prediction module and confirmation module; The physical information neural network prediction module builds an LSTM neural network based on the Anaconda environment, modifies the loss function to incorporate physical information, optimizes the loss function by adjusting the ratio of physical loss to original loss, and ultimately combines the two to minimize loss. It uses a random search algorithm to search and select the optimal model, and transmits the optimal model to the confirmation module. The confirmation module uses the trained optimal model to put the three collected IC curve features into the neural network to predict the health status of the lithium battery; and transmits the prediction results to the result visualization module; The result visualization module is used to present the prediction of lithium battery health status; The storage module is used to save the data generated during the system operation process.

Images