A battery state of charge prediction method fusing physical information and neural network

Abstract

The application relates to the technical field of battery state of charge prediction, in particular to a battery state of charge prediction method fusing physical information and a neural network, which comprises the following steps: constructing an equivalent circuit model of a battery, obtaining a state of charge calculation formula of the battery according to ampere-hour integration, and obtaining a recursive formula of terminal voltage and a recursive formula of the state of charge; constructing a neural network model, predicting terminal voltage and the state of charge of the battery at the next moment according to input parameters at the current moment, and calculating a loss function based on the error between a predicted value and an actual value and the error between the predicted value and a recursive value obtained through the recursive formula of the terminal voltage and the recursive formula of the state of charge. The application has the beneficial effect that machine learning and physical information based on the equivalent circuit model are fused, the physical information of the equivalent circuit model is integrated into the neural network loss function, the model training considers both physical laws and data characteristics, and the prediction accuracy of the state of charge is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of battery state of charge prediction technology, and specifically to a method for predicting battery state of charge. Background Technology

[0002] Battery state of charge (SOC) prediction technology is a key technology for ensuring the rational and safe operation of energy storage systems, and it is also the foundation for monitoring and maintaining these systems. It directly relates to the assessment of equipment's endurance, operational stability, and battery lifespan management. Batteries include, but are not limited to, lithium-ion batteries, sodium-ion batteries, quasi-solid-state batteries, and all-solid-state batteries. With the continuous development of battery technology and the increasing diversity of application scenarios, the requirements for the accuracy of battery SOC prediction are also becoming increasingly stringent.

[0003] In practical applications, the battery SOC changes nonlinearly with changes in environmental factors. Traditional ampere-hour integration and open-circuit voltage prediction methods have low accuracy, and SOC prediction methods based solely on electrochemical models or equivalent circuit models are greatly affected by battery operating conditions and have weak generalization ability. Summary of the Invention

[0004] The purpose of this invention is to provide a battery state of charge prediction method that integrates physical information and neural networks, thereby solving the above-mentioned technical problems; A method for predicting the state of charge of a battery that integrates physical information and neural networks includes, Step S1: Construct an equivalent circuit model of the battery, obtain the dynamic equation of the battery based on the equivalent circuit model, and obtain the state of charge calculation formula of the battery according to the ampere-hour integral. Discretize the dynamic equation and the state of charge calculation formula to obtain the recursive formula of the terminal voltage and the recursive formula of the state of charge. Step S2: Construct a neural network model to predict the battery's terminal voltage and state of charge at the next moment based on the input parameters at the current moment. Calculate the loss function based on the error between the predicted value and the true value, as well as the error between the predicted value and the recursive value obtained through the recursive formulas for the terminal voltage and the state of charge. Step S3: By minimizing the loss function and iterating the model parameters, a trained neural network model is obtained, and the trained neural network model is used to predict the state of charge of the battery.

[0005] Preferably, in step S1, the equivalent circuit model includes, Equivalent voltage source, used to simulate the open-circuit voltage of a battery; An equivalent ohmic resistor, the first end of which is connected to the positive terminal of the equivalent voltage source, is used to simulate the internal resistance of a battery in ohms. An RC network, connected to the second terminal of the equivalent ohmic resistor, is used to simulate the polarization resistance and polarization capacitance of the battery.

[0006] Preferably, in step S1, the RC network is a first-order RC network, including: Equivalent polarization resistor, the first end of which is connected to the equivalent ohmic resistor, is used to simulate the polarization internal resistance of the battery; An equivalent polarization capacitor is provided, wherein the first end of the equivalent polarization capacitor is connected to the first end of the equivalent polarization resistor, and the second end of the equivalent polarization capacitor is connected to the second end of the equivalent polarization resistor, for simulating the polarization capacitance of a battery.

[0007] Preferably, in step S1, the expression for the dynamic equation obtained based on the equivalent circuit model is: ; ; in, This indicates the resistance value of the equivalent ohmic resistor; This indicates the capacitance value of the equivalent polarization capacitor; This indicates the resistance value of the equivalent polarization resistor; express The rate of change of voltage across the equivalent polarized capacitor at time t; express The terminal voltage of the equivalent polarization capacitor at the given time; express The terminal voltage at that moment; express Open-circuit voltage at any given moment; express Current at any given moment.

[0008] Preferably, in step S1, the state of charge of the battery is obtained based on the ampere-hour integration. The calculation formula is as follows: ; in, Indicates charging capacity; Indicates charge / discharge efficiency; express Current at any given moment.

[0009] Preferably, in step S1, the discretized terminal voltage The recursive formula is, ; in, Representing discrete time Open-circuit voltage at that time; Representing discrete time The current at that time.

[0010] Preferably, in step S1, the recursive formula for the discretized state of charge is: ; in, Representing discrete time The state of charge at that time; Representing discrete time The state of charge at that time; Representing discrete time The current at that time.

[0011] Preferably, in step S2, the error term of the loss function of the neural network model includes, The error between the state of charge predicted by the neural network model and the actual state of charge. ; The error between the terminal voltage predicted by the neural network model and the actual terminal voltage ; The error between the terminal voltage predicted by the neural network model and the recursive terminal voltage obtained through the recursive formula of the terminal voltage. ; The error between the state of charge predicted by the neural network model and the recursive state of charge obtained through the recursive formula of the state of charge. ; The predicted value includes the state of charge predicted by the neural network model and the predicted terminal voltage; the true value includes the true state of charge and the true terminal voltage; the recursive value includes the recursive terminal voltage obtained by the recursive formula of the terminal voltage and the recursive state of charge obtained by the recursive formula of the state of charge.

[0012] Preferably, the input parameters include the true value and the recursive value, and the true value further includes the true current and the true temperature.

[0013] Preferably, in step S2, the loss function The calculation formula is as follows: ; in The weight matrix represents the error term of the loss function.

[0014] The beneficial effects of the present invention are as follows: By adopting the above technical solutions, the battery state of charge prediction method of the present invention integrates machine learning and physical information based on equivalent circuit models. By incorporating the physical information of the equivalent circuit model into the neural network loss function, the model training takes into account both physical laws and data characteristics, effectively improving the prediction accuracy of the state of charge. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the steps of the battery state of charge prediction method that integrates physical information and neural networks in an embodiment of the present invention; Figure 2 This is a schematic diagram of the first-order equivalent circuit model in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0019] A method for predicting the state of charge of batteries that integrates physical information and neural networks, such as Figure 1 , Figure 2 As shown, including, Step S1: Construct an equivalent circuit model of the battery, obtain the dynamic equation of the battery based on the equivalent circuit model, and obtain the state of charge calculation formula of the battery according to the ampere-hour integral. Discretize the dynamic equation and the state of charge calculation formula to obtain the recursive formula of the terminal voltage and the recursive formula of the state of charge. Step S2: Construct a neural network model to predict the battery's terminal voltage and state of charge at the next moment based on the input parameters at the current moment. Calculate the loss function based on the error between the predicted value and the true value, as well as the error between the predicted value and the recursive value obtained through the recursive formulas for terminal voltage and state of charge. Step S3: Iterate the model parameters by minimizing the loss function to obtain the trained neural network model, and use the trained neural network model to predict the state of charge of the battery.

[0020] Specifically, the battery state of charge prediction method of the present invention integrates machine learning and physical information based on equivalent circuit models. By incorporating the physical information of the equivalent circuit model into the neural network loss function, the model training takes into account both physical laws and data characteristics, effectively improving the prediction accuracy of the state of charge.

[0021] In a preferred embodiment, in step S1, as follows: Figure 2 As shown, the equivalent circuit model includes, The equivalent voltage source VOC is used to simulate the open-circuit voltage of a battery. The equivalent ohmic resistance R0 is connected at its first end to the positive terminal of the equivalent voltage source VOC, and is used to simulate the internal ohmic resistance of the battery. An RC network, connected to the second terminal of the equivalent ohmic resistance R0, is used to simulate the polarization resistance and polarization capacitance of a battery.

[0022] Specifically, the present invention first establishes an equivalent circuit model through an equivalent voltage source VOC, a first-order (or multi-order) RC network and an equivalent ohmic resistor R0. The introduced RC network is used to simulate the effects of battery polarization internal resistance and polarization capacitance, and the equivalent ohmic resistor R0 is used to simulate battery ohmic internal resistance.

[0023] In a preferred embodiment, in step S1, the RC network is a first-order RC network, including: Equivalent polarization resistance R1, with its first end connected to equivalent ohmic resistance R0, is used to simulate the polarization internal resistance of the battery. The equivalent polarization capacitor C1 is connected to the first terminal of the equivalent polarization resistor R1, and the second terminal of the equivalent polarization capacitor C1 is connected to the second terminal of the equivalent polarization resistor R1. This is used to simulate the polarization capacitance of a battery.

[0024] Specifically, the first-order RC equivalent circuit model established in this invention is as follows: Figure 1 As shown, it includes an equivalent ohmic resistor R0, an equivalent polarization resistor R1, an equivalent polarization capacitor C1, and an equivalent voltage source VOC.

[0025] More specifically, taking a first-order RC network as an example, the equivalent polarization resistance R1 and capacitance simulate polarization characteristics, and the equivalent ohmic resistance R0 simulates ohmic internal resistance, providing a basis for subsequent accurate calculations. The formula for calculating the state of charge is obtained through ampere-hour integration, and the relevant equations are discretized to obtain recursive formulas, making the continuous physical process calculable at discrete time points.

[0026] The parameters of the equivalent circuit model are obtained by using parameter identification methods, such as the least squares method. By minimizing the error between the estimated voltage and the actual voltage, the parameters are updated, allowing the model to better reflect the actual situation of the battery and laying the physical model foundation for accurate prediction of the state of charge.

[0027] In a preferred embodiment, in step S1, the expression for the dynamic equation obtained based on the equivalent circuit model is: ; ; in, This represents the resistance value of the equivalent ohmic resistance R0; This represents the capacitance value of the equivalent polarization capacitance C1; This represents the resistance value of the equivalent polarization resistance R1; express The rate of change of voltage across the equivalent polarization capacitor C1 at any given time; express The terminal voltage of the equivalent polarization capacitor C1 at any given time; express The terminal voltage at that moment; express Open-circuit voltage at any given moment; express Current at any given moment.

[0028] In a preferred embodiment, in step S1, the state of charge of the battery is obtained based on the ampere-hour integration. The calculation formula is as follows: ; in, Indicates charging capacity; Indicates charge / discharge efficiency; express Current at any given moment.

[0029] In a preferred embodiment, in step S1, the discretized terminal voltage The recursive formula is, ; in, Representing discrete time Open-circuit voltage at that time; Representing discrete time The current at that time.

[0030] In a preferred embodiment, in step S1, the recursive formula for the discretized state of charge is: ; in, Representing discrete time The state of charge at that time; Representing discrete time The state of charge at that time; Representing discrete time The current at that time.

[0031] Specifically, the parameters (including R0, R1 and C1) in the equivalent circuit model of the present invention need to be obtained through parameter identification methods, such as least squares method, recursive least squares method and particle swarm algorithm.

[0032] The equivalent circuit model estimates the battery terminal voltage based on the discretized battery dynamic equations, and updates and obtains the model parameters by minimizing the error between the estimated voltage and the actual voltage.

[0033] In a preferred embodiment, in step S2, the error term of the loss function of the neural network model includes, The error between the state of charge predicted by the neural network model and the actual state of charge. ; The error between the terminal voltage predicted by the neural network model and the actual terminal voltage. ; The error between the terminal voltage predicted by the neural network model and the recursive terminal voltage obtained through the recursive formula for the terminal voltage. ; The error between the state of charge predicted by the neural network model and the recursive state of charge obtained through the recursive formula for the state of charge. ; The predicted values ​​include the state of charge and the predicted terminal voltage predicted by the neural network model; the true values ​​include the true state of charge and the true terminal voltage; the recursive values ​​include the recursive terminal voltage obtained by the recursive formula for the terminal voltage and the recursive state of charge obtained by the recursive formula for the state of charge.

[0034] Specifically, this invention constructs a deep neural network model for predicting the state of charge (SOC) of a battery. The constructed neural network uses battery parameters from n historical time points, including battery voltage, current, temperature, and SOC, to predict the battery voltage and SOC at the (n+1)th time point.

[0035] The battery current and temperature at historical moments can be regarded as the input of the battery system, and the corresponding battery voltage and SOC are the system response.

[0036] The deep neural network of this invention can be a multilayer perceptron (MLP), a convolutional neural network (CNN), a recurrent neural network (RNN), or other neural networks.

[0037] More specifically, the loss function Loss of the constructed deep neural network consists of the following four parts: Neural network predicted state of charge vs. actual state of charge Error, error term Represented as, ; in This represents the state of charge predicted by the neural network.

[0038] Neural network predicted terminal voltage vs. actual voltage Error, error term Represented as, ; in This represents the terminal voltage predicted by the neural network.

[0039] In constructing a deep neural network model to predict battery SOC, this invention incorporates physical information related to output voltage and SOC into the loss function. The aim is to allow the neural network model to learn patterns that conform to the actual physical characteristics of the battery, thereby improving the model's accuracy. More specifically, physical information related to the output voltage is incorporated into the loss function, and the recursive terminal voltage obtained based on the recursive formula for the terminal voltage is denoted as... Calculate the terminal voltage predicted by the neural network and the recursive terminal voltage. Error, error term Represented as, ; Physical information related to the state of charge is incorporated into the loss function, and the recursive state of charge obtained based on the recursive formula of the state of charge is denoted as follows. Calculate the state of charge predicted by the neural network and the recursive state of charge. The error between them, the error term Represented as, ; In a preferred embodiment, the input parameters include true values ​​and recursive values, with the true values ​​further including true current and true temperature.

[0040] In a preferred embodiment, in step S2, the loss function The calculation formula is as follows: ; in The weight matrix represents the error term of the loss function.

[0041] More specifically, determine the weight matrices for the four parts. Then the loss function The calculation method is as follows: ; The deep neural network model is trained by dividing the training set. During forward propagation, the model accepts the input parameters at time t, including the actual battery voltage, the voltage predicted by the equivalent circuit model, the actual current, the actual temperature, and the actual state of charge, and outputs the predicted values ​​of the battery voltage and state of charge at time t+1. During backpropagation, four loss functions are obtained based on the predicted voltage and state of charge, as well as the input parameters. to The model's internal parameters are updated and iterated by minimizing the loss and weighted summation.

[0042] Leveraging the powerful nonlinear fitting capabilities of deep neural network models, the battery voltage and state of charge (SOC) at the next moment are predicted using battery parameters from multiple historical moments. Physical information related to battery voltage and SOC is incorporated into the loss function. By calculating the errors between the predicted values ​​and the true and recursive values, the neural network is trained to consider both physical laws and data characteristics.

[0043] For example, error term , It reflects the difference between predicted and actual values, prompting the model to learn data features; , This reflects the difference between the predicted values ​​and the recursive values ​​based on the physics model, ensuring that the model follows physical laws. Weight matrix The loss function is obtained by weighted summation of different error terms. By adjusting the weights, the impact of each error term on model training can be balanced, guiding the model to train in a direction that is more in line with reality.

[0044] In one specific embodiment, the present invention constructs a first-order RC equivalent circuit model and a deep neural network model. In this example, the deep neural network is a two-layer long short-term memory (LSTM) neural network.

[0045] The model is validated based on laboratory operating condition tests, including constant current charging and constant current discharging of the battery. The dataset is divided into training set, validation set and test set. The training set should contain data under different operating conditions to ensure the generalization ability of the model.

[0046] A first-order RC equivalent circuit model is constructed, and the physical constraints of the equivalent circuit model are fused with the loss function of LSTM, including constraints on battery voltage and state of charge, and the four error terms of the loss function. to The weights of the model are set to 0.2, 0.2, 0.3 and 0.3 respectively, and the Adam optimizer is used to minimize the total loss function Loss to obtain the model parameters.

[0047] During the model training phase, forward propagation outputs predicted values ​​based on the input parameters, while backpropagation updates the model parameters by minimizing the loss function. The input parameters include true values ​​and recursive values. The true values, such as real current and temperature, reflect the real-time state of the battery, while the recursive values ​​embody information from the physical model. This rich input allows the model to learn more comprehensive battery characteristics.

[0048] In practice, a two-layer Long Short-Term Memory (LSTM) neural network is employed, which has excellent processing capabilities for time series data. The dataset is divided into training, validation, and test sets. The training set contains data under different operating conditions to enhance the model's generalization ability, enabling it to adapt to various battery operating scenarios. The Adam optimizer is used to minimize the total loss function to obtain model parameters, ensuring continuous optimization during training and improving prediction accuracy.

[0049] The above description is merely a preferred embodiment of the present invention and does not limit the implementation and protection scope of the present invention. Those skilled in the art should realize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the state of charge of a battery by integrating physical information and neural networks, characterized in that, include, Step S1: Construct an equivalent circuit model of the battery, obtain the dynamic equation of the battery based on the equivalent circuit model, and obtain the state of charge calculation formula of the battery according to the ampere-hour integral. Discretize the dynamic equation and the state of charge calculation formula to obtain the recursive formula of the terminal voltage and the recursive formula of the state of charge. Step S2: Construct a neural network model to predict the battery's terminal voltage and state of charge at the next moment based on the input parameters at the current moment. Calculate the loss function based on the error between the predicted value and the true value, as well as the error between the predicted value and the recursive value obtained through the recursive formulas for the terminal voltage and the state of charge. Step S3: By minimizing the loss function and iterating the model parameters, a trained neural network model is obtained, and the trained neural network model is used to predict the state of charge of the battery.

2. The battery state of charge prediction method integrating physical information and neural networks according to claim 1, characterized in that, In step S1, the equivalent circuit model includes, Equivalent voltage source (VOC) is used to simulate the open-circuit voltage of a battery; An equivalent ohmic resistor (R0) is provided, with its first terminal connected to the positive terminal of the equivalent voltage source (VOC) to simulate the internal ohmic resistance of a battery. An RC network, connected to the second terminal of the equivalent ohmic resistance (R0), is used to simulate the polarization resistance and polarization capacitance of a battery.

3. The battery state of charge prediction method integrating physical information and neural networks according to claim 2, characterized in that, In step S1, the RC network is a first-order RC network, including: Equivalent polarization resistance (R1), the first end of which is connected to the equivalent ohmic resistance (R0), is used to simulate the polarization internal resistance of the battery. An equivalent polarization capacitor (C1) is provided, with its first terminal connected to the first terminal of the equivalent polarization resistor (R1) and its second terminal connected to the second terminal of the equivalent polarization resistor (R1), to simulate the polarization capacitance of a battery.

4. The battery state-of-charge prediction method integrating physical information and neural networks according to claim 3, characterized in that, In step S1, the expression for the dynamic equation obtained based on the equivalent circuit model is: ； ； in, This indicates the resistance value of the equivalent ohmic resistance (R0); This indicates the capacitance value of the equivalent polarization capacitor (C1); This indicates the resistance value of the equivalent polarization resistance (R1); express The rate of change of voltage at the equivalent polarization capacitor (C1) at time t; express The terminal voltage of the equivalent polarization capacitor (C1) at that moment; express The terminal voltage at that moment; express Open-circuit voltage at any given moment; express Current at any given moment.

5. The battery state of charge prediction method integrating physical information and neural networks according to claim 1, characterized in that, In step S1, the state of charge of the battery is obtained based on the ampere-hour integration. The calculation formula is as follows: ； in, Indicates charging capacity; Indicates charge / discharge efficiency; express Current at any given moment.

6. The battery state of charge prediction method integrating physical information and neural networks according to claim 4, characterized in that, In step S1, the discretized terminal voltage The recursive formula is, ； in, Representing discrete time Open-circuit voltage at that time; Representing discrete time The current at that time.

7. The battery state of charge prediction method integrating physical information and neural networks according to claim 5, characterized in that, In step S1, the recursive formula for the discretized state of charge is: ； in, Representing discrete time The state of charge at that time; Representing discrete time The state of charge at that time; Representing discrete time The current at that time.

8. The battery state of charge prediction method integrating physical information and neural networks according to claim 1, characterized in that, In step S2, the error term of the loss function of the neural network model includes, The error between the state of charge predicted by the neural network model and the actual state of charge. ; The error between the terminal voltage predicted by the neural network model and the actual terminal voltage ; The error between the terminal voltage predicted by the neural network model and the recursive terminal voltage obtained through the recursive formula of the terminal voltage. ; The error between the state of charge predicted by the neural network model and the recursive state of charge obtained through the recursive formula of the state of charge. ; The predicted value includes the state of charge predicted by the neural network model and the predicted terminal voltage; the true value includes the true state of charge and the true terminal voltage; the recursive value includes the recursive terminal voltage obtained by the recursive formula of the terminal voltage and the recursive state of charge obtained by the recursive formula of the state of charge.

9. The battery state-of-charge prediction method integrating physical information and neural networks according to claim 8, characterized in that, The input parameters include the true value and the recursive value, and the true value also includes the true current and the true temperature.

10. The battery state-of-charge prediction method integrating physical information and neural networks according to claim 8, characterized in that, In step S2, the loss function The calculation formula is as follows: ； in The weight matrix represents the error term of the loss function.

Images