Battery charge state estimation method and system based on physical gating neural network

Abstract

The invention provides a battery state-of-charge estimation method and system based on a physical gating neural network, and the method comprises the steps: obtaining battery operation data, and inputting the data to a data drive estimation module to obtain a preliminary SOC estimation value; meanwhile, the physical correction module performs integration on the battery operation data to obtain a physical estimation value, and generates a correction estimation value under physical constraint; calculating confidence through a confidence gating module to perform weighted fusion on the initial estimation value and the corrected estimation value, and outputting a final state of charge estimation result; and determining an optimal hyper-parameter through K-fold cross validation and completing network training to obtain a trained physical gating neural network, and predicting a state-of-charge estimation value of the battery based on the trained physical gating neural network. According to the method, high-precision and low-error SOC estimation can be realized under a complex multi-cycle charging and discharging working condition, and the stability and the reliability of an estimation result are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management systems, and particularly relates to a battery state of charge estimation method and system based on a physical gated neural network. BACKGROUND

[0002] With the wide application of electric vehicles, renewable energy storage systems and portable electronic devices, lithium batteries have become the core energy storage device. The state of charge (SOC) as a key indicator to measure the remaining capacity and health level of the battery is crucial for the health management (PHM) and safe operation. Accurate estimation of SOC not only helps to improve the efficiency of the battery, but also reduces the risk of system operation and prolongs the life of the equipment. Traditional SOC estimation methods gradually reveal their limitations, so new technologies and methods need to be introduced to improve the estimation accuracy and system reliability.

[0003] The current mainstream SOC estimation methods include coulomb counting method, open circuit voltage method, electrochemical model, and equivalent circuit model based method. Although widely used, these methods generally rely on idealized conditions or assumed parameters, making it difficult to accurately estimate in complex environmental factors such as temperature fluctuations, load changes and other actual scenarios. Especially in extreme weather conditions, these traditional methods often produce large errors and consume a lot of computing resources.

[0004] In recent years, data-driven methods, especially deep learning technology, have been widely used in SOC estimation problems. For example, long short-term memory (LSTM) network has become a research hotspot due to its advantages in processing time series data. These methods do not require complex mechanism analysis, and directly use measured voltage and current for SOC prediction. For example, the invention patent with publication number CN120352779A discloses a storage lithium battery state of charge evaluation method and system combining physical model and neural network, which includes: using a lumped parameter equivalent circuit model battery system representation method to extract power lithium ion battery charge characteristic information; using the power lithium ion battery charge characteristic information as the input feature of the neural network, a new type of lumped parameter equivalent circuit model long short-term memory neural network model is constructed; using the new type of lumped parameter equivalent circuit model long short-term memory neural network model, the state of charge of the measured power lithium ion battery is estimated. However, this method needs to construct an equivalent circuit model and combine LSTM regression prediction to achieve high estimation accuracy. For example, researchers have proposed other improved methods, such as LSTM network combined with attention mechanism and Kalman filter, bidirectional LSTM network integrated with attention mechanism, stress-introduced LSTM estimation network, model combining convolutional neural network and LSTM network, and hybrid neural network CNNSAMLSTM model. Although this type of method has significant advantages in modeling complex nonlinear relationships, due to the lack of physical constraints, the model has weak generalization ability and poor interpretability.

[0005] Therefore, it is urgent to research and develop a new SOC estimation method, which can improve the explainability and robustness by establishing a neural network architecture that integrates physical knowledge, especially for precise estimation of battery state of charge in complex and variable working environments, significantly reducing estimation error and improving prediction stability and reliability. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a battery state of charge estimation method and system based on a physical gating neural network, which combines data-driven preliminary estimation and physical calibration results of coulomb counting, and performs weighted fusion through constraint gating and confidence gating mechanisms, thereby improving estimation accuracy and considering physical consistency in multi-cycle charging and discharging scenarios.

[0007] The present application estimates the state of charge of a lithium battery by the following technical solutions: The present application provides a battery state of charge estimation method based on a physical gating neural network, which comprises the following steps: Step 1: Obtain the operating data of the battery, wherein the operating data of the battery includes battery discharge end voltage, current, temperature, time interval and historical state of charge estimation value; Step 2: Input the operating data of the battery, and perform preliminary estimation through a data-driven estimation module to obtain a preliminary estimation value; integrate the operating data of the battery through a physical correction module to obtain a physical estimation value, generate a corrected estimation value under physical constraints based on the preliminary estimation value and the physical estimation value; calculate the confidence through a confidence gating module, which is used to weight and fuse the preliminary estimation value and the corrected estimation value under physical constraints, output the state of charge estimation value of the battery, and complete the construction of the physical gating neural network; Step 3: Based on the operating data of the battery as input, calculate the loss function of the physical gating neural network through K-fold cross-validation, screen out the optimal hyperparameters, train and optimize the physical gating neural network based on the optimal hyperparameters, and obtain the trained physical gating neural network; Step 4: Input the operating data of the battery, and output the state of charge estimation value of the battery through the trained physical gating neural network.

[0008] In one embodiment, the operating data of the battery is divided in step 1 to obtain a training set, a validation set and a test set; wherein the training set is used to construct the physical gating neural network, the training set and the validation set are mixed to establish a new training set for training and optimization of the physical gating neural network, and the test set is used for the physical gating neural network to predict the state of charge estimation of the battery.

[0009] In one embodiment, in step 2, the preliminary estimation value obtained by the data-driven estimation module includes: Based on the input battery operation data, the battery operation data is time-series modeled using a convolutional neural network layer to generate convolutional features, the convolutional features are processed through a pooling layer and a flattening layer to obtain a one-dimensional feature vector; the current time interval and the historical state of charge estimation value are spliced into an enhanced feature vector; the one-dimensional feature vector and the enhanced feature vector are spliced to obtain a preliminary estimation feature vector; The preliminary estimation feature vector is nonlinearly processed by a first multi-layer perceptron, and a data-driven preliminary estimation value is generated by a Sigmoid function.

[0010] In one embodiment, the first multi-layer perceptron includes: an input layer, a fully connected hidden layer with a RELU activation function, and a linear output layer; wherein the output layer of the fully connected hidden layer is represented as , is the input of the first multi-layer perceptron, i.e., the preliminary estimation feature vector, and are the weight matrix and the bias of the fully connected hidden layer, respectively; the output of the linear output layer is represented as , and are the weight matrix and the bias of the linear output layer, respectively.

[0011] In one embodiment, in step 2, the integration of the battery operation data to obtain a physical estimation value includes: based on the battery operation data, the current current value , the time interval , and the state of charge estimation value of the previous time are input into a coulomb counter module, and integration is performed according to the principle of charge conservation to obtain a physical estimation value , the formula of which is: , wherein, is the coulomb count.

[0012] In one embodiment, in step 2, the generation of a corrected estimation value under physical constraints based on the preliminary estimation value and the physical estimation value includes: the preliminary estimation value and the physical estimation value are input into a limiting gate module, and a trainable threshold is used to regulate the generation of a corrected estimation value under physical constraints , the formula of which is: .

[0013] In one embodiment, in step 2, the process of calculating the confidence is: inputting the one-dimensional feature vector and the historical state of charge estimate into a second multi-layer perception in the confidence gating module to calculate the confidence of the data-driven estimate , the second multi-layer perception comprises: an input layer, a fully connected hidden layer with a RELU activation function, and a linear output layer; wherein the output layer of the fully connected hidden layer is represented as , is represented as the input of the second multi-layer perception, i.e., the one-dimensional feature vector and the historical state of charge estimate, and are the weight matrix and the bias of the fully connected hidden layer, respectively; the output of the linear output layer is represented as , and are the weight matrix and the bias of the linear output layer, respectively.

[0014] In one embodiment, in step 2, the calculation formula of the weighted fusion of the preliminary estimate value and the revised estimate value under the physical constraint is as follows: , wherein, represents the state of charge estimate value of the battery.

[0015] In one embodiment, the optimal hyperparameters screened in step 3 include the learning rate, the batch size, the physical gating neural network lookback length, the convolution kernel size, and the number of hidden layer neurons.

[0016] On the other hand, the present application also provides a battery state of charge estimation system based on a physical gating neural network, which applies the battery state of charge estimation method based on the physical gating neural network, comprising: a data acquisition unit for acquiring the operating data of the battery, wherein the operating data of the battery comprises: the discharge end voltage, the current, the temperature, the time interval, and the historical state of charge estimate value of the battery; a physical gating neural network construction unit for inputting the operating data of the battery, performing preliminary estimation by a data-driven estimation module to obtain a preliminary estimate value; performing integration on the operating data of the battery by a physical correction module to obtain a physical estimate value, generating a revised estimate value under the physical constraint based on the preliminary estimate value and the physical estimate value, calculating a confidence by a confidence gating module, weighting the preliminary estimate value and the revised estimate value under the physical constraint, and outputting the state of charge estimate value of the battery to complete the construction of the physical gating neural network; The physical gated neural network training unit takes battery-based operating data as input, calculates the loss function of the physical gated neural network through K-fold cross-validation, selects the optimal hyperparameters, trains and optimizes the physical gated neural network based on the optimal hyperparameters, and obtains the trained physical gated neural network. The state of charge estimation prediction unit is used to take the battery's operating data as input and output the battery's state of charge estimate through a trained physical gated neural network.

[0017] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) The physical gated neural network constructed in this invention not only has the advantages of deep models in feature extraction and expression, but also integrates battery physics knowledge such as coulomb counting, realizing the unity of data-driven and knowledge-driven approaches.

[0018] (2) This invention, through the collaborative design of estimation gating, constraint gating, and confidence gating, can effectively constrain the model output within the physically feasible range, thereby improving the stability and reliability of the estimation. It can still achieve low-error, highly robust SOC estimation even in complex multi-cycle, multi-temperature environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the battery state-of-charge estimation method based on a physical gated neural network provided by the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of a physical gated neural network provided in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the training process of a physical gated neural network provided in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the structure of a battery state-of-charge estimation system based on a physical gated neural network, provided in an embodiment of the present invention. Detailed Implementation

[0024] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0026] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0027] It should also be further understood that the term "and or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0028] The present invention provides a battery state of charge estimation method and system based on a physical gated neural network. It combines data-driven preliminary estimation with physical calibration results of coulomb counting, and performs weighted fusion through constraint gating and confidence gating mechanisms. While ensuring the expressive power of the model, it introduces physical constraints to improve interpretability and robustness. It is particularly suitable for accurate estimation of battery state of charge in complex and variable working environments, significantly reducing estimation errors and improving the predictive stability and reliability of battery state of charge estimates.

[0029] The solution of the present invention is as follows: Figure 1 As shown, a battery state-of-charge estimation method based on a physical gated neural network includes the following steps: S1. Acquire battery operating data, wherein the battery operating data includes: battery discharge terminal voltage, current, temperature, time interval, and historical state of charge estimation values. Specifically: S11. In this embodiment, battery operating data is acquired, including the voltage at the battery discharge terminal. Current ,temperature Time interval Data and historical state of charge estimates These data are derived from the actual operating environment of lithium batteries and can fully reflect the battery's charging and discharging behavior and environmental characteristics at a certain time period: Voltage at the battery discharge terminal Monitoring changes in the battery's terminal voltage during discharge can reflect the battery's health and discharge efficiency. Voltage variations help analyze battery performance under different loads and identify potential faults. Current Recording current data helps understand the battery's discharge intensity and rate. The relationship between current and battery capacity is the basis for estimating SOC (State of Charge), and by analyzing current changes, the remaining energy of the battery can be better predicted.

[0030] Temperature data Temperature is a crucial factor affecting battery performance. Batteries may exhibit different operating characteristics at different temperatures. Monitoring temperature changes can help optimize battery usage conditions and provide appropriate protection and management in high or low temperature environments.

[0031] Battery capacity data Obtaining battery capacity data is crucial for understanding battery status, as capacity reflects the total amount of energy a battery can store. Capacity data allows us to assess the battery's remaining charge and lifespan, thereby optimizing battery usage strategies.

[0032] corresponding time Recording the timestamps for each set of data is crucial, as it establishes a correlation between battery operating data and time. This is significant for analyzing battery performance trends, making time-series predictions, and optimizing charging / discharging strategies.

[0033] By integrating various operational data, the battery's State of Charge (SOC) can be estimated more accurately, reducing the risk of information loss and improving the reliability of the estimation results. Visualizing the relationship between various battery operating parameters and SOC makes the model output more transparent, facilitating user understanding and analysis. Monitoring operational data under different environmental conditions allows the battery management system to adjust its operating strategies in real time to adapt to temperature changes and load demands, ensuring the battery operates at its optimal state. Real-time analysis of battery operating data enables the rapid identification of potential faults or anomalies, allowing for preventative measures and extending battery life.

[0034] The accumulated operational data can be used for subsequent algorithm optimization and model training, improving the intelligence level of the battery management system and making it more adaptable to the needs of different application scenarios.

[0035] In summary, obtaining comprehensive battery operation data not only improves the intelligence level of battery management, but also significantly enhances battery performance and safety, providing users with a better user experience.

[0036] S2. Using battery operating data as input, a preliminary estimate is obtained through a data-driven estimation module. A physical correction module integrates the battery operating data to obtain a physical estimate. Based on the preliminary and physical estimates, a corrected estimate under physical constraints is generated. A confidence gating module calculates the confidence level, which is used to weight and fuse the preliminary and corrected estimates under physical constraints, outputting the battery's state of charge estimate, thus completing the construction of the physical-gated neural network. Details are as follows: The above data is input into a physical gated neural network for state of charge estimation. The physical gated neural network (SocGate) structure used in this embodiment is as follows: Figure 2 As shown, the whole consists of three main parts: The first part is the data-driven estimation module. This module first uses convolutional neural network layers (Conv1, Conv2) to perform time-series modeling on historical sequence data, extracting the spatial and temporal relationships between voltage, current, temperature, time intervals, and historical estimates, generating a dimension-wise model. The intermediate feature map; the spatial and temporal relationships between voltage, current, temperature and historical SOC refer to the interaction and dependence of these variables at different points in time.

[0037] Spatial relationships: This involves the interactions between voltage, current, temperature, and state of charge (SOC) at any given moment. For example, a change in voltage at a given point in time may affect the state of current and SOC. Convolutional neural networks (CNNs) can effectively capture the correlations and patterns between these variables through their convolutional layers.

[0038] Temporal relationships: This refers to the dynamic changes of these variables over time. For example, changes in current may affect changes in SOC after a certain delay, or changes in temperature may show a certain pattern with changes in voltage. By processing historical data, convolutional neural networks can identify these temporal dependencies and trends.

[0039] In summary, convolutional neural networks can learn the complex spatial and temporal relationships between these variables through feature extraction, thereby more accurately estimating the state and performance of the battery.

[0040] Specifically, the convolutional neural network layer structure is used as the starting point for feature extraction to extract spatial relationships between different inputs and temporal relationships between different historical data. The input sequence of the neural network, i.e., the historical data length, is... The voltage, current, temperature, SOC, and corresponding time series are stacked into a 5×L matrix. The matrix passes through two convolutional layers, named Conv1 and Conv2, and the kernel size can be adjusted as a hyperparameter.

[0041] Specifically, the convolution process can be represented as: Historical Matrix The first set of intermediate feature maps is obtained after passing through the first convolutional layer (Conv1), denoted as... Then, the set of feature maps The second set of intermediate feature maps is obtained after the second convolutional layer (Conv2), denoted as... Both sets of feature maps have a dimension of 1. For convolution output Perform pooling and flatten into a one-dimensional feature vector. ; the current value at the current moment and time interval value and the historical estimate of the previous moment. Concatenate into a feature enhancement vector This is concatenated with the flattened feature vector to form the enhanced feature vector. ; Enhanced feature vector The data is input to the first multilayer perceptron module, undergoes nonlinear processing, and is then used to generate an output data-driven preliminary estimate via the Sigmoid function. .

[0042] In this embodiment, the first multilayer perceptron module, named MLP1, is used at the end of feature extraction. The input to MLP1 is the enhanced feature vector. The output is a preliminary data-driven estimate. .

[0043] Specifically, the MLP1 module includes an input layer, a fully connected hidden layer with ReLU activation, and a linear output layer. The input of the MLP1 module is denoted as... The hidden layer weight matrix and bias vector are respectively and .

[0044] The output layer of a fully connected hidden layer is represented as: ; The ReLU activation function is defined as follows: ; The weight matrix and bias vector of the output layer are respectively and MLP1 output Represented as: ; Substituting the output of the hidden layer into the expression, the complete MLP1 model is represented as follows: .

[0045] This estimate reflects the battery's current state of charge prediction obtained by the model based on data learning, and serves as an important input for subsequent module fusion. This module, combined with physical estimation and confidence gating, effectively improves the accuracy and stability of the overall estimation results.

[0046] The second part is the physical correction module, which includes two sub-modules: a coulomb counting module and a limit gating module. The coulomb counting module uses the state-of-charge estimate from the previous time step. Current current and time interval Integrating, we obtain the physical estimate. The restriction gating module, on the other hand, is based on a preliminary, data-driven estimate. The deviation between the physical estimate and the actual value is determined using a trainable threshold. Adjust the fusion method of the two to generate corrected estimates under physical constraints. .

[0047] Specifically, it includes the following steps: Current current value Time interval The estimated state of charge at the previous time step Input the data into the coulomb counter module, and perform integration based on the principle of charge conservation to obtain the physical estimate. Its formula is: ,in, It is the coulomb number; Data-driven preliminary estimation Compared with physical estimates A common input restriction gating module limits the gating module to pass through a trainable threshold. Controlling the influence of the deviation between the two, a corrected estimate under physical constraints is generated. This module design enhances the physical plausibility of the estimation results and avoids prediction biases that do not conform to the battery charging and discharging patterns. The specific formula is: .

[0048] The third part is the confidence gating module, which learns the confidence level of data-driven estimation in the current environment through a neural network (second multilayer perceptron) and outputs confidence parameters. Finally, the data-driven preliminary estimate and the physically constrained corrected estimate are weighted and fused according to the confidence level to obtain the final estimate. .

[0049] Specifically, the following steps are included: One-dimensional feature vector and historical estimates The second multilayer perceptron in the input confidence gating module outputs the confidence level of the data-driven estimate. ; The second multilayer perceptron module includes an input layer, a fully connected hidden layer with ReLU activation function, and a linear output layer connected in sequence.

[0050] The second multilayer perceptron module, MLP2, is used in the confidence gating module. The input to MLP2 is a one-dimensional feature vector. Combined with historical estimates, the output is the confidence level of the data-driven estimation results. .

[0051] The input to the MLP2 module is denoted as... The hidden layer weight matrix and bias vector are respectively and .

[0052] The output of the hidden layer is represented as: ; The ReLU activation function is defined as follows: ; The weight matrix and bias vector of the output layer are respectively and MLP2 output Represented as: ; Substituting the output of the hidden layer into the expression, the complete MLP2 model is represented as follows: .

[0053] Based on this confidence level, the preliminary data-driven estimate and the corrected estimate under physical constraints are weighted and averaged to obtain the final state of charge estimate. The weighting formula is as follows: , in, This represents the estimated state of charge of the battery.

[0054] Through the synergy of the three modules mentioned above, the physical gating neural network model can effectively integrate the advantages of both data learning and physical rules, providing stable, accurate, and interpretable SOC estimation outputs under different environments and cycle stages.

[0055] By extracting key features from battery operating data (such as voltage, current, temperature, and SOC), physical gated neural network models can more accurately reflect the true state of the battery, thereby improving the ability to predict battery performance.

[0056] Convolutional neural networks effectively learn the spatial and temporal relationships in battery data, which helps to identify dependencies and dynamic changes between variables, thus providing a more comprehensive understanding of battery performance.

[0057] By storing SOC values ​​from a past period, the physical gated neural network model can leverage historical information to enhance its judgment of the current state and improve the reliability of its predictions.

[0058] The adoption of a step-by-step, modular approach (including data-driven preliminary estimation and correction of physical constraints) results in a clear system structure that is easy to maintain and improve. Different modules can be optimized independently, flexibly adapting to different application scenarios.

[0059] Using a multilayer perceptron to process features can effectively reduce computational complexity, while the ReLU activation function can accelerate the training process.

[0060] In summary, this method, through the application of deep learning technology, not only extracts complex patterns from battery operating data but also effectively constrains the rationality of neural network outputs through physical computation mechanisms. Compared to traditional deep neural network models, this method exhibits stronger stability and interpretability under sparse or extreme data conditions, effectively improving the accuracy of battery state estimation and the predictive ability of the model, which is of great significance to the development of battery management systems.

[0061] S3. Using battery-based operating data as input, the loss function of the physical gated neural network is calculated through K-fold cross-validation. The optimal hyperparameters are then selected, and the physical gated neural network is trained and optimized based on these optimal hyperparameters to obtain the trained physical gated neural network. Details are as follows: In this embodiment, as Figure 3 As shown, the aforementioned physical gated neural network model is trained by inputting historical battery operation data, corresponding time values, and historical SOC as a sample set into the SocGate neural network model, including: Define the current SOC of the battery; specifically, ;in, express Battery capacity at any time, This indicates the initial capacity of the battery; if capacity loss occurs, the capacity value from the previous moment is assigned to the moment of loss. Acquire historical battery operating data, corresponding time values, and historical SOC; The historical battery operating data, corresponding time values, and historical SOC are normalized and combined with the current SOC of the battery to form a sample set; The sample set is divided into a training set, a validation set, and a test set; specifically, the ratio of the training set, validation set, and test set is approximately 6:1:3.

[0062] 10-fold cross-validation is performed using the training and validation sets. The average network loss is calculated. When the average network loss is less than a set threshold, the validation ends, and the optimal hyperparameters of the SocGate neural network model are recorded. These include the learning rate, batch size, network backtracking length, convolutional kernel size, and number of hidden layer neurons.

[0063] The training set and validation set are mixed into a new training set, the SocGate neural network model is trained using the optimal network hyperparameter settings, and the SocGate neural network model is saved. The saved model is tested using the test set, and the estimated battery capacity value is recorded to obtain the physical gated neural network.

[0064] By using voltage Current ,temperature Corresponding time interval and historical SOC estimates As input, the SocGate neural network model can learn battery behavior more comprehensively, thereby improving the accuracy of SOC estimation. Handling capacity loss allows the model to still provide reasonable SOC estimates when facing battery performance degradation, enhancing its robustness. Normalizing the input data eliminates the influence between different units, making model training more stable and convergent faster. Dividing the training, validation, and test sets according to temperature ensures the model's generalization ability under different environmental conditions, making it more reliable in practical applications. Using 10-fold cross-validation effectively evaluates model performance, reduces the risk of overfitting, and improves the model's predictive ability on new data. Recording the optimal network hyperparameters improves the model's learning efficiency and accuracy, making the training process more efficient. Saving the trained model facilitates rapid deployment and use in subsequent practical applications, enhancing the model's practicality.

[0065] This method enhances the model's applicability under various environmental conditions, ensuring good performance across different application scenarios. Evaluating the capabilities of the saved model using a test set provides reliable battery capacity predictions for practical applications, helping users better manage and utilize batteries.

[0066] In summary, this method not only improves the accuracy and reliability of SOC estimation, but also enhances the model's adaptability under various environmental conditions, providing strong support for the optimization of battery management systems.

[0067] S4. Using the battery's operating data as input, the trained physical gated neural network outputs an estimated value of the battery's state of charge and returns it to the terminal or battery management system for subsequent energy management and safety control.

[0068] The method in this embodiment combines coulomb counting with deep learning through physical gating functions, enabling more accurate estimation of the battery's state of charge (SOC) in complex and variable operating environments. Particularly under extreme temperature conditions, this method significantly reduces errors and improves the accuracy of SOC estimation. The robustness of the physically gated neural network to temperature variations is enhanced; the convolutional feature extraction module and the constraint gating mechanism work together to ensure that the physically gated neural network model maintains stable performance under different temperature conditions.

[0069] The constraint gating mechanism utilizes the deviation between the data-driven estimate and the physical estimate to dynamically adjust and physically correct the SOC result, thereby improving the continuity and accuracy of the estimation result. This is particularly suitable for scenarios with long-term operation or significant changes in ambient temperature. Geographical location differences, day-night cycles, and seasonal changes can all affect battery performance, and this method can effectively adapt to these changes, ensuring the stability and reliability of the SOC estimate.

[0070] Existing deep learning models, due to their "black box" nature, often lack interpretability, leading users to question their effectiveness. However, the physically-gated neural network model used in this embodiment improves model transparency by introducing explicit physical computation paths and confidence control strategies, enabling users to better understand and trust the prediction results. This structured design not only enhances the model's interpretability but also increases user confidence in the estimation results.

[0071] By combining data-driven and knowledge-driven approaches, and fully utilizing existing operational data and physical laws, the overall performance of the model is improved, resulting in more accurate, stable, and easily understandable SOC estimation. This method is applicable to various lithium battery applications, including electric vehicles, renewable energy storage systems, and portable electronic devices. The required data includes only the battery's voltage, current, temperature, time intervals, and historical estimates.

[0072] The aforementioned battery state-of-charge (SOC) estimation method based on a physical gated neural network acquires battery discharge terminal voltage, current, temperature, and time data to form a complete battery operation dataset. This data is then input into the physical gated neural network model for SOC estimation. SocGate, trained using historical samples, includes a data-driven estimation module, a physical correction module, and a confidence gating module. It effectively identifies complex patterns and, combined with physical consistency constraints, improves the model's accuracy and stability under varying environments. Ultimately, the output estimation results achieve high-precision and high-reliability SOC estimation.

[0073] Corresponding to the above-described battery state-of-charge estimation method based on a physically gated neural network, this invention also provides a battery state-of-charge estimation system based on a physically gated neural network. For example... Figure 4 As shown, the system can be deployed on a server to execute the aforementioned estimation method and achieve online battery SOC estimation.

[0074] Specifically, the battery state of charge estimation system based on physical gated neural network includes a data acquisition unit, a physical gated neural network construction unit, a physical gated neural network training unit, and a state of charge estimation and prediction unit.

[0075] The data acquisition unit is used to acquire battery operating data, which includes voltage, current, temperature, time interval data, and historical state of charge estimates. The physical gated neural network construction unit is used to take the battery's operating data as input, perform a preliminary estimation through the data-driven estimation module to obtain a preliminary estimate, integrate the battery's operating data through the physical correction module to obtain a physical estimate, generate a corrected estimate under physical constraints based on the preliminary estimate and the physical estimate, and calculate the confidence level through the confidence gating module to weight and fuse the preliminary estimate and the corrected estimate under physical constraints, outputting the battery's state of charge estimate, thus completing the construction of the physical gated neural network. The physical gated neural network training unit takes battery-based operating data as input, calculates the loss function of the physical gated neural network through K-fold cross-validation, selects the optimal hyperparameters, trains and optimizes the physical gated neural network based on the optimal hyperparameters, and obtains the trained physical gated neural network. The state of charge estimation prediction unit is used to take the battery's operating data as input and output the battery's state of charge estimate through a trained physical gated neural network.

[0076] In this embodiment, the physical gated neural network construction unit includes: an input subunit for inputting the battery operation data into the physical gated neural network; and a feature extraction subunit for using a convolutional feature extraction module to obtain a multidimensional feature representation of the operation data. The estimation gate sub-unit is used to combine the current and time data at the current moment, process the feature vector, and output a data-driven preliminary estimate. The physical estimation subunit is used to generate corrected estimates under physical constraints using the coulomb counting module and the constraint gating module. The confidence fusion subunit is used to weight and fuse the initial and physical correction estimates based on the confidence level of the current state, and output the final estimate.

[0077] In this embodiment, the feature extraction subunit includes: The relation extraction module is used to extract voltage, current, temperature, time interval and history layer using convolutional neural network layers, and to pool and flatten the convolution output into a one-dimensional feature vector. The connection layer is used to concatenate the current, time interval, and historical state of charge estimates into an enhanced feature vector; the one-dimensional feature vector and the enhanced feature vector are concatenated to obtain the preliminary estimated feature vector. The vector processing layer is used to input the concatenated preliminary estimated feature vector into the first multilayer perceptron for processing, generating a data-driven preliminary estimate.

[0078] In this embodiment, the physical estimation subunit includes: The historical estimation and storage module is used to save the estimated state of charge value at the previous moment; The coulomb integration module is used to calculate physical estimates based on the current current, time interval, and historical estimates. The constraint gating module is used to adjust the deviation between the data-driven preliminary estimate and the physical estimate, and output the corrected estimate under physical constraints.

[0079] In this embodiment, the confidence fusion subunit includes: The confidence calculation module is used to evaluate the confidence of data-driven estimates based on current state characteristics. The fusion output module is used to fuse the data-driven estimate and the physical correction value according to the confidence level, and output the final estimation result.

[0080] It should be noted that those skilled in the art will understand that the specific implementation of the battery state of charge estimation system based on physical gated neural networks and its various units can be found in the corresponding descriptions in the aforementioned embodiments of the battery state of charge estimation method in this specification, and will not be repeated here.

[0081] The aforementioned battery state-of-charge estimation system based on a physical gated neural network can be implemented as a computer program that runs on a computer device. This computer device can be a server, which can be a single unit or a server cluster consisting of multiple servers.

[0082] The computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory includes non-volatile storage media and internal memory.

[0083] Non-volatile storage media can store an operating system and a computer program. The computer program includes program instructions that, when executed, cause the processor to perform a battery state-of-charge estimation method based on a physically-gated neural network.

[0084] The processor provides computing and control capabilities for the entire computer device and supports the operation of the method of the present invention.

[0085] Internal memory provides the execution environment for computer programs; during execution, the program can: estimate the battery state of charge based on battery operating data through a physically gated neural network model.

[0086] Network interfaces are used to communicate and interact with external terminals, databases, sensor systems, or remote servers.

[0087] It should be understood that the above system structure is only a schematic structure for implementing the solution of this application and does not constitute a limitation on the specific computer device structure; the computing device may include more or fewer components, or be implemented in different arrangements.

[0088] In this embodiment, the processor runs a computer program to perform the following steps: Step 1: Obtain battery operating data, which includes: battery discharge terminal voltage, current, temperature, time interval, and historical state of charge estimation values; Step 2: Using the battery's operating data as input, a preliminary estimate is obtained through the data-driven estimation module; the physical correction module integrates the battery's operating data to obtain a physical estimate; based on the preliminary estimate and the physical estimate, a corrected estimate under physical constraints is generated; the confidence gating module calculates the confidence level, which is used to weight and fuse the preliminary estimate and the corrected estimate under physical constraints, outputting the battery's state of charge estimate, thus completing the construction of the physical gated neural network; Step 3: Using the battery's operating data as input, calculate the loss function of the physical gated neural network through K-fold cross-validation, select the optimal hyperparameters, train and optimize the physical gated neural network based on the optimal hyperparameters, and obtain the trained physical gated neural network; Step 4: Using the battery's operating data as input, the trained physical gated neural network outputs an estimated value of the battery's state of charge.

[0089] During model training, the processor is further used for: Define the estimated current state of charge of the battery; acquire historical operating data and corresponding time values, and perform normalization processing; divide the sample set into training set, validation set and test set; Construct a physically gated neural network structure and determine network hyperparameters such as convolutional kernel size and number of perceptron layers; select the optimal network structure parameters through cross-validation; retrain the model based on the optimal parameters and save it. Finally, the performance of the physical gated neural network model was verified using a test set, and the estimation results were recorded.

[0090] During the inference phase, the processor further performs the following refinement steps: The running data is input into the neural network model, and multidimensional feature relationships are extracted through the convolutional feature extraction module; the current at the current moment, the time interval concatenation, and the historical state of charge estimates are spliced ​​together to form an enhanced feature vector; The preliminary estimated feature vector is processed nonlinearly using the first multilayer perceptron, and a data-driven preliminary estimate is generated using the Sigmoid function. The physical estimate is calculated using the coulomb counting module; The constraint gating module generates a corrected estimate under physical constraints based on the deviation between the two. The confidence gating module outputs confidence scores, which are then weighted and fused. The final output is the SOC estimation result at the current time.

[0091] This method explicitly combines the principle of battery power conservation with a time-series awareness mechanism within the neural network model, thereby enhancing the physical rationality and interpretability of the estimation results while ensuring predictive performance.

[0092] It should be understood that, in the embodiments of this application, the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or other conventional processor.

[0093] Those skilled in the art will understand that all or part of the methods described in the embodiments of the present invention can be implemented by a computer program instructing corresponding hardware. This computer program includes program instructions, can be stored in a computer-readable storage medium, and can be executed by a processor to implement the method steps of the present invention.

[0094] Therefore, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the program to perform the following operations: Step 1: Obtain battery operating data, which includes: battery discharge terminal voltage, current, temperature, time interval, and historical state of charge estimation values; Step 2: Using the battery's operating data as input, a preliminary estimate is obtained through the data-driven estimation module; the physical correction module integrates the battery's operating data to obtain a physical estimate; based on the preliminary estimate and the physical estimate, a corrected estimate under physical constraints is generated; the confidence gating module calculates the confidence level, which is used to weight and fuse the preliminary estimate and the corrected estimate under physical constraints, outputting the battery's state of charge estimate, thus completing the construction of the physical gated neural network; Step 3: Using the battery's operating data as input, calculate the loss function of the physical gated neural network through K-fold cross-validation, select the optimal hyperparameters, train and optimize the physical gated neural network based on the optimal hyperparameters, and obtain the trained physical gated neural network; Step 4: Using the battery's operating data as input, the trained physical gated neural network outputs an estimated value of the battery's state of charge.

[0095] In this embodiment, when the processor executes the above program to implement the training steps of the physical gated neural network model, it specifically includes: defining the current state of charge; acquiring and normalizing historical battery operation data, time values, and historical estimates to construct a sample set; dividing the sample set into a training set, a validation set, and a test set; constructing a physical gated neural network structure; performing cross-validation using the training set and the validation set, and recording the optimal network hyperparameters when the mean of the loss function is less than a preset threshold; performing final training on the model using the optimal hyperparameters and saving the training results; and validating the saved model using the test set, recording the prediction performance to form a usable estimation system.

[0096] In this embodiment, when the processor performs the state of charge estimation step, it further implements the following process: Battery operating data is input into a neural network; a convolutional feature extraction module is used to extract time-series features such as voltage, current, and temperature; the current current and time interval are concatenated and input into a first multilayer perceptron to generate a data-driven preliminary estimate; a coulomb counting module calculates a physical estimate based on the previous estimate, current, and time; the data-driven preliminary estimate and the physical estimate are input into a constraint gating module to obtain a corrected estimate under physical constraints; a confidence gating module calculates the confidence level, and the two estimates are fused to output the final SOC estimation result.

[0097] This design closely integrates the expressive power of neural networks with the physical laws of batteries, achieving accurate predictions while improving the physical consistency and interpretability of the model, making it suitable for real-time assessment of battery state of charge in various environments.

[0098] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0099] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0100] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0101] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0102] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0103] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A battery state of charge estimation method based on a physical gated neural network, characterized by, The battery state of charge estimation method comprises the following steps: Step 1: obtaining the operation data of the battery, wherein the operation data of the battery comprises the discharge end voltage, current, temperature, time interval and historical state of charge estimation value of the battery; Step 2: taking the operation data of the battery as input, a preliminary estimation value is obtained through a data-driven estimation module; a physical estimation value is obtained by integrating the operation data of the battery through a physical correction module, a corrected estimation value under physical constraints is generated based on the preliminary estimation value and the physical estimation value; a confidence degree is calculated through a confidence gating module, which is used to weight and fuse the preliminary estimation value and the corrected estimation value under physical constraints, and the state of charge estimation value of the battery is output, and the construction of the physical gating neural network is completed; Step 3: taking the operation data of the battery as input, the physical gating neural network loss function is calculated through K-fold cross-validation, the optimal hyperparameters are screened out, and the physical gating neural network is trained and optimized based on the optimal hyperparameters, and the trained physical gating neural network is obtained; Step 4: taking the operation data of the battery as input, the state of charge estimation value of the battery is output through the trained physical gating neural network.

2. The battery state-of-charge estimation method based on a physical-gated neural network according to claim 1, wherein, In step 1, the operation data of the battery is divided to obtain a training set, a validation set and a test set; wherein the training set is used to construct the physical gating neural network, the training set and the validation set are mixed to establish a new training set for training and optimization of the physical gating neural network, and the test set is used for the physical gating neural network to predict the state of charge estimation of the battery.

3. The method of claim 1, wherein, In step 2, the preliminary estimation value obtained through the data-driven estimation module comprises: Based on the input battery operation data, a convolutional neural network layer is used to model the battery operation data in time sequence, generate convolutional features, and the convolutional features are processed through a pooling layer and a flattening layer to obtain a one-dimensional feature vector; the current time interval and the historical state of charge estimation value are spliced into an enhanced feature vector; the one-dimensional feature vector and the enhanced feature vector are spliced to obtain a preliminary estimation feature vector; The first multilayer perceptron is used to perform nonlinear processing on the preliminary estimation feature vector, and the Sigmoid function is used to generate the data-driven preliminary estimation value.

4. The battery state-of-charge estimation method based on a physical-gated neural network according to claim 3, characterized in that, The first multi-layer perceptron comprises: an input layer, a fully connected hidden layer with a RELU activation function, and a linear output layer; wherein the output of the fully connected hidden layer is represented as , is represented as the input of the first multi-layer perceptron, i.e., the preliminary estimated feature vector, and are respectively a weight matrix and a bias of the fully connected hidden layer; the output of the linear output layer is represented as , and are respectively a weight matrix and a bias of the linear output layer.

5. The battery state-of-charge estimation method based on a physical-gated neural network according to claim 3, wherein, Step 2, the integration of the battery's operating data to obtain a physical estimate, includes: based on the battery's operating data, integrating the current value... Time interval and the estimated state of charge at the previous moment Input the data into the coulomb counter module, and perform integration based on the principle of charge conservation to obtain the physical estimate. Its formula is: ， wherein, is the Coulomb number.

6. The method of claim 5, wherein the physical-gated neural network is a recurrent neural network. In step 2, the generating the revised estimation value under the physical constraint based on the preliminary estimation value and the physical estimation value comprises: inputting the preliminary estimation value and the physical estimation value into a limit gate module together, and generating the revised estimation value under the physical constraint through a trainable threshold . The limit gate module is configured to regulate the generation of the revised estimation value under the physical constraint , and the formula is: 。 7. The method of claim 5, wherein the physical-gated neural network is a recurrent neural network. In step 2, the process of calculating the confidence is: inputting the one-dimensional feature vector and the historical state-of-charge estimation into a second multi-layer perception in the confidence gating module to calculate the confidence of the data-driven estimation , the second multi-layer perception comprises: an input layer, a fully connected hidden layer with a RELU activation function, and a linear output layer; wherein the output layer of the fully connected hidden layer is represented as , , the input of the second multi-layer perception is the one-dimensional feature vector and the historical state-of-charge estimation, , and are the weight matrix and the bias of the fully connected hidden layer respectively; the output of the linear output layer is represented as , , and are the weight matrix and the bias of the linear output layer respectively.

8. The method of claim 6, wherein, The calculation formula of the weighted fusion of the preliminary estimation value and the corrected estimation value under physical constraints is as follows: ， wherein, represents an estimated value of the state of charge of the battery.

9. The method of claim 1, wherein, The optimal hyperparameters screened out in step 3 include learning rate, batch size, physical gating neural network lookback length, convolution kernel size and hidden layer neuron number.

10. A battery state of charge estimation system based on a physical gated neural network, the system comprising: The application of the battery state of charge estimation method based on the physical gating neural network according to any one of claims 1-9 comprises: a data acquisition unit for acquiring the operation data of the battery, wherein the operation data of the battery comprises the discharge end voltage, current, temperature, time interval and historical state of charge estimation value of the battery; The physical gating neural network construction unit is configured to take the operation data of the battery as input, perform preliminary estimation through a data-driven estimation module to obtain a preliminary estimation value, perform integration on the operation data of the battery through a physical correction module to obtain a physical estimation value, generate a corrected estimation value under physical constraints based on the preliminary estimation value and the physical estimation value, calculate a confidence degree through a confidence gating module, and output a state of charge estimation value of the battery by weighted fusion of the preliminary estimation value and the corrected estimation value under the physical constraints, thereby completing construction of the physical gating neural network. The physical gating neural network training unit is configured to take the operation data of the battery as input, calculate a physical gating neural network loss function through K-fold cross-validation, screen out optimal hyperparameters, train and optimize the physical gating neural network based on the optimal hyperparameters, and obtain a trained physical gating neural network. The state of charge estimation and prediction unit is configured to take the operation data of the battery as input, output a state of charge estimation value of the battery through the trained physical gating neural network.

Images