Lithium ion battery SOC multi-model fusion estimation method based on neural network

Abstract

The invention provides a lithium ion battery SOC multi-model fusion estimation method based on a neural network, and belongs to the technical field of lithium ion battery management, and the method comprises the steps: carrying out the segmentation preprocessing of the current, voltage and temperature data of a battery according to the charging and discharging direction; fitting an SOC-OCV relation curve based on a second-order RC equivalent circuit model, and optimizing dynamic response parameters; synchronously predicting OCV and SOC by adopting a multi-task neural network; a high-precision SOC initial value is obtained by looking up a table through an OCV predicted value, and dynamic updating is carried out in combination with an ampere-hour integral method; comparing the ampere-hour integral result with the SOC directly predicted by the neural network in real time, and dynamically fusing according to an error; for a long-term standing scene, a life model is constructed based on historical data to predict SOH, the nominal capacity is updated, and the integral reference is reset. According to the method, a physical model and a data driving method are fused, and the SOC estimation precision and reliability under complex working conditions and long-term standing are improved through dual-path collaborative estimation and SOH closed-loop calibration.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery management technology, and in particular to a multi-model fusion estimation method for lithium-ion battery SOC based on neural networks. Background Technology

[0002] To meet the demands of the rapid development of new energy vehicles and improve the performance and safety of lithium-ion power batteries, state-of-charge (SOC) estimation has become a core function of battery management systems. Lithium-ion power batteries, as the primary energy storage device in new energy vehicles, are widely used due to their high energy density, long cycle life, and wide operating temperature range. However, the internal electrochemical reactions of lithium-ion batteries are complex, especially under conditions such as long-term vehicle idling or complex operating conditions. Traditional SOC estimation methods suffer from problems such as strong dependence on initial values ​​and lack of capacity decay calibration, directly affecting the accuracy of electric vehicle range and battery safety.

[0003] Current methods for estimating State of Charge (SOC) are based on battery models with multi-order equivalent circuits. Filters or observers are used to identify model parameters online to obtain the open-circuit voltage. An OCV-SOC model is established based on the results of offline and online parameter identification. A neural network is then used to correct errors in the initial estimation results, overcoming the limitations of single methods. This approach retains the simplicity of the ampere-hour integration method while improving accuracy and reducing accumulated errors using a DNN. However, the ampere-hour integration method is highly dependent on the initial SOC. The accurate initial SOC of the battery must be known at the start of the calculation; if the initial SOC is incorrectly set, the entire estimation process will be inaccurate from the beginning. Because the SOC of a lithium-ion battery has a non-linear relationship with OCV, it is difficult to accurately obtain the initial SOC by simply measuring the voltage. Furthermore, existing methods do not consider the degradation of the battery's State of Health (SOH) due to long-term quiescent or aging, and the nominal capacity is not dynamically updated. After quiescent, the SOC estimation error can reach over 15%. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-model fusion estimation method for SOC of lithium-ion batteries based on neural networks. This method simultaneously predicts open-circuit voltage (OCV) and state of charge (SOC) through neural networks. The OCV prediction value is used to obtain a high-precision initial SOC through a lookup table method, providing a reliable benchmark for the ampere-hour integration method. At the same time, the ampere-hour integration result is compared with the SOC prediction value directly output by the neural network in real time. The two outputs are dynamically fused based on an error threshold to form a closed-loop calibration mechanism, which significantly improves the long-term accuracy and operating condition adaptability of SOC estimation. This solves the problem of accumulated error caused by the high sensitivity of the ampere-hour integration method to the initial SOC value in the prior art, as well as the capacity calibration inaccuracy caused by the degradation of the battery state of health (SOH) after long-term vehicle idling.

[0005] To achieve the above objectives, this invention proposes a multi-model fusion estimation method for the state of charge (SOC) of lithium-ion batteries based on neural networks, comprising the following steps: Step S1: Extract the measured data of the hybrid power pulse characteristic (HPPC) of lithium-ion batteries, segment the data according to the current direction, divide each data segment into data fragments according to the time window or the change amplitude of battery charge state (SOC), remove irrelevant information and data noise, use linear interpolation to complete the missing data and normalize it to obtain the preprocessed data. Step S2: Construct an equivalent circuit model of a lithium-ion battery based on the preprocessed data, conduct a constant current charge-discharge experiment, fit the SOC-OCV curve using a polynomial function based on the obtained experimental raw data, and identify the experimental data parameters using the least squares method. Step S3: Build a multi-task neural network model. Input real-time current, voltage, temperature and historical average features. The output includes two independent paths: Path 1: Predict the battery open-circuit voltage OCV; Path 2: Directly predict the battery state of charge SOC. Step S4: Based on the predicted OCV value output by path one, obtain the initial SOC value by looking up a table. Using the initial SOC value as a reference, dynamically update the SOC value by combining the ampere-hour integration method with the real-time current. Step S5: Compare the results predicted by the ampere-hour integration method with the SOC value directly predicted by the neural network in real time, dynamically weight and fuse them according to the error threshold, and output the final optimal estimate of SOC. Step S6: Set a vehicle idle time threshold. When the vehicle idle time exceeds the vehicle idle time threshold, estimate the battery health status (SOH) based on historical operating data and calibrate the SOC benchmark.

[0006] Preferably, step S1 includes the following steps: Step S11: Divide the HPPC measured data into charging and discharging segments according to the current direction, and divide each segment into data fragments based on the time window or the SOC change amplitude. Step S12: Remove irrelevant information by using a low-pass filter to remove data noise. The cutoff frequency of the filter is set to be less than twice the main frequency of the signal. Step S13: Use linear interpolation to fill in the missing data, normalize the filled data, and eliminate the dimensional differences between different physical quantities. Preferably, in step S12, irrelevant information includes, but is not limited to, test steps and relative timestamps.

[0007] Preferably, step S2 includes the following steps: Step S21: Establish a second-order RC circuit model, where the fast polarization part is composed of... R 1 andC 1. Description: The slow polarization component is composed of... R 2 and C 2. Description, the formula is: ; in, For real-time voltage, Open circuit voltage, For real-time current, For ohmic internal resistance, For rapid polarization voltage, This is a slow polarization voltage; Step S22: Based on the constant current charge-discharge experiment, record the OCV values ​​under different SOCs. Use a cubic polynomial to fit the relationship between SOC and OCV, and determine the fitting coefficients using the least squares method. The formula is: ; in, The coefficients are the fitting coefficients for the cubic term. The battery is in its state of charge. The fitting coefficients are quadratic terms. The coefficients for the first-order term are the fitting coefficients. These are the fitting coefficients for the constant term; Step S23: Define the sum of squared errors between the model's predicted values ​​and the actual measured values. Identify the experimental data parameters using the least squares method. The objective function is to minimize the sum of squared errors between the model's predicted voltage and the actual measured voltage. By minimizing the error, optimize the model parameters. R 1. R 2. C 1 and C 2. The objective function formula is: ; in, To minimize the sum of squared errors between the model-predicted voltage and the actual measured voltage, i.e., the objective function, min(·) is the minimization operation function. For the selected window length, k For the data point sequence, For actual voltage measurement, Predict voltage for the model.

[0008] Preferably, step S3 includes the following steps: Step S31: Construct the initial model of the neural network. The input data of the input layer includes, but is not limited to, real-time current, real-time voltage and temperature. The hidden layer contains two LSTM layers and one fully connected layer. The output layer contains two independent paths and outputs the predicted values ​​of open-circuit voltage (OCV) and charge state (SOC). Define the hyperparameters of the network, including the number of layers, the number of neurons and the activation function. Step S32: Divide the preprocessed data into training set, validation set and test set, optimize the weights using Adam optimizer, adjust the network weights through forward propagation and back propagation algorithms to minimize prediction error, and introduce regularization to prevent overfitting. Step S33: Monitor the root mean square error (RMSE) and mean absolute error (MAE) in the test set. When the model's generalization ability meets the application requirements, output the OCV prediction value and the SOC prediction value simultaneously.

[0009] Preferably, step S4 includes the following steps: Step S41: Based on the predicted OCV value output by the neural network of path one, query the OCV-SOC mapping table to obtain the initial SOC value. The mapping table is established through the second-order RC model parameter identification experiment. Step S42: Based on the initial SOC value, update the SOC value in real time using the ampere-hour integration method, with the following formula: ; in, To obtain the predicted value through ampere-hour integration, For the first k The time corresponding to each time sampling point This is the initial value of SOC. Nominal capacity For Coulomb efficiency, for t The current at any given time is integrated numerically using the trapezoidal rule.

[0010] Preferably, in step S4, the SOC value obtained by looking up the table is compared with the SOC estimated by the ampere-hour integration method at regular intervals, and a preset error range threshold is set. When the error is within the preset error range threshold, the neural network table lookup correction is paused.

[0011] Preferably, step S5 includes the following steps: Step S51: Calculate the absolute error between the ampere-hour integral prediction and the neural network prediction in real time. The formula is: ; in, The absolute error between the ampere-hour integral prediction and the neural network prediction. The value is a neural network prediction. Step S52: Dynamically adjust the fusion weights based on the error threshold, using the following formula: ; in, This is the final value of the SOC result.

[0012] Preferably, step S6 includes the following steps: Step S61: Extract historical operating data, including the number of complete charge-discharge cycles, average depth of discharge, and historical average ambient temperature. Step S62: Input the lifetime decay model to calculate SOH. The calculation formula is as follows: ; in, SOH For battery health status, For a complete charge-discharge cycle, The average depth of discharge. This represents the historical average ambient temperature. , , , These are the model calibration coefficients; Step S63: Update the nominal capacity using the following formula: ; in, For the new capacity parameter of the ampere-hour integration method, This is the rated capacity.

[0013] Preferably, in step S62, , , , Through experimental calibration, the calibration process needs to cover the number of charge-discharge cycles, depth of discharge, and ambient temperature range throughout the battery's entire life cycle.

[0014] Therefore, this invention proposes a multi-model fusion estimation method for the state of charge (SOC) of lithium-ion batteries based on neural networks, which has the following advantages: (1) This invention removes noise and irrelevant information by grouping and filtering the working condition data, making the data input clearer and more efficient, which greatly improves the accuracy of SOC estimation and the robustness of the model.

[0015] (2) The present invention adopts a second-order RC circuit model and combines polynomial fitting and least squares method to identify parameters, accurately captures the fast and slow polarization characteristics of the battery, optimizes the OCV-SOC relationship curve, and improves the applicability and accuracy of the SOC estimation model.

[0016] (3) The present invention outputs OCV prediction value and SOC direct prediction value simultaneously, providing high-precision initial SOC, avoiding the problem of dependence on the initial value of ampere-hour integral, making SOC estimation more accurate and stable, and improving the generalization and adaptability of the model through automatic feature extraction.

[0017] (4) Based on the adaptive weighting of the error threshold of the ampere-hour integral result and the neural network prediction, the present invention strengthens the continuity of the integral with small errors and prioritizes the anti-interference ability of the neural network with large errors. It realizes the dynamic calibration and error correction of SOC and solves the problem of error accumulation, ensuring the long-term stability and adaptability of SOC estimation.

[0018] (5) This invention combines battery model, ampere-hour integration method and neural network correction to improve SOC estimation accuracy; it dynamically updates the SOC estimation benchmark by predicting SOH after resting through historical data, making up for the impact of resting state on battery performance and adapting to different operating conditions; it integrates data-driven and physical models, and has both real-time performance and robustness, making it suitable for complex application scenarios. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the technical route of a multi-model fusion estimation method for lithium-ion battery SOC based on neural networks; Figure 2 This is a schematic diagram of the neural network structure for a multi-model fusion estimation method for lithium-ion battery SOC based on neural networks. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] Example 1 like Figure 1 As shown, this invention provides a multi-model fusion estimation method for the state of charge (SOC) of lithium-ion batteries based on neural networks, comprising the following steps: Step S1: Acquire operating condition test data, group the acquired operating condition test data, divide each group of data into segments, extract the measured data of the hybrid power pulse characteristic (HPPC) of lithium-ion batteries, segment the HPPC measured data according to the current direction, divide each segment of data into data fragments according to the time window or the change amplitude of battery state of charge (SOC), remove irrelevant information and data noise, use linear interpolation to complete missing data and normalize it to obtain preprocessed data, specifically including the following steps: Step S11: Divide the HPPC measured data into charging and discharging segments according to the current direction. The charging segment current... Discharge segment current Each data segment is divided into data fragments based on a time window or the magnitude of SOC change. The time window... Or the magnitude of SOC change ; Step S12: Filter out irrelevant information such as test steps and relative timestamps, retain the core parameters of current, voltage and temperature, and use a low-pass filter with a cutoff frequency less than twice the main signal frequency to remove data noise and suppress noise interference. Step S13: Use linear interpolation to fill in missing data, eliminate discontinuities between data, and perform Z-score standardization on the filled data to eliminate dimensional differences between different physical quantities.

[0023] Step S2: Construct an equivalent circuit model of a lithium-ion battery based on the preprocessed data, conduct a constant current charge-discharge experiment, fit the SOC-OCV curve using a polynomial function based on the obtained raw experimental data, and identify the experimental data parameters using the least squares method. Specifically, this includes the following steps: Step S21: Considering the accuracy and complexity of the battery model, two sets of RC circuits are used to describe the internal polarization characteristics of the battery, establishing a second-order equivalent circuit model of the battery. The first RC network includes a polarization resistor R1 and a polarization capacitor C1, used to describe the battery's fast dynamic response; the second RC network includes a larger polarization resistor R2 and a polarization capacitor C2, used to describe the battery's slow dynamic response. The model formula is: ; in, For real-time voltage, Open circuit voltage, For real-time current, For ohmic internal resistance, For rapid polarization voltage, This is a slow polarization voltage; Step S22: Based on the constant current charge-discharge experiment, record the OCV values ​​under different SOCs. Use a cubic polynomial to fit the relationship between SOC and OCV, determine the fitting coefficients using the least squares method, and obtain the battery's open-circuit voltage-state-of-charge (OCV-SOC) curve. The formula is: ; in, The coefficients are the fitting coefficients for the cubic term. The battery is in its state of charge. The fitting coefficients are quadratic terms. The coefficients for the first-order term are the fitting coefficients. These are the fitting coefficients for the constant term; Step S23: Define the sum of squared errors between the model's predicted values ​​and the actual measured values. Identify the experimental data parameters using the least squares method. The objective function is to minimize the sum of squared errors between the model's predicted voltage and the actual measured voltage. Minimize the error using the least squares method to optimize the model parameters. R 1. R 2.C 1 and C 2. The objective function formula is: ; in, To minimize the sum of squared errors between the model-predicted voltage and the actual measured voltage, i.e., the objective function, min(·) is the minimization operation function. For the selected window length, k For the data point sequence, For actual voltage measurement, Predict voltage for the model.

[0024] Step S3: Build a multi-task neural network model, with the following structure: Figure 2 As shown, the input consists of real-time current, voltage, temperature, and historical average characteristics. The output includes two independent paths: Path 1: Predicting the battery open-circuit voltage (OCV); Path 2: Directly predicting the battery state of charge (SOC), specifically including the following steps: Step S31: Construct the initial model of the neural network. The input data of the input layer are real-time current, real-time voltage and temperature. The hidden layer is used to extract data features and contains two LSTM layers and one fully connected layer. The output layer contains two independent paths and outputs the predicted values ​​of open-circuit voltage (OCV) and charge state (SOC). Define the hyperparameters of the network, including the number of layers, the number of neurons and the activation function. Step S32: Before model training, perform data partitioning and serialization. Divide the preprocessed data into training set, validation set and test set. Set the time step to construct the input sequence and perform normalization. Use Adam optimizer to optimize weights. Adjust network weights through forward propagation and back propagation algorithms to minimize prediction error. Introduce regularization to prevent overfitting. Step S33: After training, verify the performance of the neural network model on an independent test set, monitor the root mean square error (RMSE) and mean absolute error (MAE), and output the OCV prediction value and SOC prediction value simultaneously when the model's generalization ability meets the application requirements.

[0025] Step S4: Based on the predicted OCV value output by path one, obtain the initial SOC value using a lookup table method. Using the initial SOC value as a reference, dynamically update the SOC value using the ampere-hour integration method combined with real-time current. This specifically includes the following steps: Step S41: Based on the predicted OCV value output by the neural network of path one, query the OCV-SOC mapping table to obtain the initial SOC value. The mapping table is established through the second-order RC model parameter identification experiment. Step S42: Using the initial SOC value as a reference, the dynamic update process of the battery charge state is initiated. The SOC value is updated in real time using the ampere-hour integration method, with the following formula: ; in, To obtain the predicted value through ampere-hour integration, For the first k The time corresponding to each time sampling point This is the initial value of SOC. Nominal capacity For Coulomb efficiency, for t The current at each time step is integrated using the trapezoidal rule; at each time step, the SOC value is gradually adjusted using the ampere-hour integration method based on the real-time current data to reflect the changes in the battery during the charging and discharging process. To reduce errors in long-term estimation, the SOC value obtained from the lookup table is periodically compared with the SOC estimated by the ampere-hour integration method. When the error between the two is within a preset range, the neural network lookup table correction is paused, and the open-circuit voltage measurement value is used directly for estimation to avoid overcorrection.

[0026] The trained neural network is used to perform forward calculations on the normalized current, voltage, and temperature data, and directly outputs the predicted charge state value, forming a SOC estimation path independent of the ampere-hour integral.

[0027] Step S5: Compare the results predicted by the ampere-hour integration method with the SOC value directly predicted by the neural network in real time, dynamically weight and fuse them according to the error threshold, and output the final optimal estimate of SOC. This includes the following steps: Step S51: Calculate the absolute error between the ampere-hour integral prediction and the neural network prediction in real time. The formula is: ; in, The absolute error between the ampere-hour integral prediction and the neural network prediction. The value is a neural network prediction. Step S52: Dynamically adjust the fusion weights based on the error threshold, using the following formula: ; in, This is the final value of the SOC result.

[0028] Step S6: Set the vehicle inactivity time threshold to 72 hours. If the vehicle inactivity time exceeds 72 hours, estimate the battery health status (SOH) based on historical operating data and calibrate the SOC benchmark. This includes the following steps: Step S61: Extract historical operating data, including the number of complete charge-discharge cycles, average depth of discharge, and historical average ambient temperature. Step S62: Input the lifetime decay model to calculate SOH. The calculation formula is as follows: ; in, SOH For battery health status, For a complete charge-discharge cycle, The average depth of discharge. This represents the historical average ambient temperature. , , , These are the model calibration coefficients; Step S63: Update the nominal capacity using the following formula: ; in, For the new capacity parameter of the ampere-hour integration method, This is the rated capacity.

[0029] Specific Implementation Case 1 Real-time data is acquired from the vehicle's battery management system, including battery charging and discharging current, terminal voltage, ambient or internal temperature during battery operation, and the time of data acquisition. This data is stored on dedicated recording equipment or a server to ensure availability for subsequent processing and modeling.

[0030] Step S1: Extract the measured data of the hybrid power pulse characteristic (HPPC) of lithium-ion batteries, segment the data according to the current direction, divide each segment into data fragments based on the time window or the change amplitude of the battery state of charge (SOC), remove irrelevant information and data noise, use linear interpolation to complete the missing data and normalize it to obtain the preprocessed data. The specific steps include: Step S11: Divide the data according to the current direction, when At this time, it is the charging phase. At any given time, the data is divided into segments based on the discharge phase. For each data segment, the data is further divided into segments based on the time window (5 seconds) or the change in SOC (a 5% decrease in SOC) to ensure the locality and clarity of subsequent analysis. Step S12: Remove irrelevant data such as test steps, relative time, marker signals and debugging information, process the noise of voltage and current signals, remove high-frequency noise with a low-pass filter, and set the cutoff frequency of the filter to less than twice the main frequency of the signal. Step S13: Use linear interpolation to fill in the missing data, normalize the filled data, and eliminate the dimensional differences between different physical quantities.

[0031] Step S2: Construct an equivalent circuit model of a lithium-ion battery based on the preprocessed data, conduct a constant current charge-discharge experiment, fit the SOC-OCV curve using a polynomial function based on the obtained raw experimental data, and identify the experimental data parameters using the least squares method. Specifically, this includes the following steps: Step S21: Establish an equivalent circuit model for the lithium iron phosphate battery. Use a second-order RC circuit model to describe the battery's dynamic characteristics. The circuit structure includes ohmic internal resistance, a fast polarization circuit, and a slow polarization circuit. Establish a mathematical model for the terminal voltage. The mathematical expression for the battery terminal voltage is as follows: ; in, For real-time voltage, This is the open-circuit voltage, reflecting the battery's state of charge (SOC). For real-time current, For ohmic internal resistance, For rapid polarization voltage, This is a slow polarization voltage; Step S22: Based on the constant current charge-discharge experiment, record the OCV value at different SOCs. Each time the SOC decreases by 5% or increases by 5%, allow the battery to stand for 30 minutes to 1 hour to ensure electrochemical equilibrium. Use a cubic polynomial to fit the relationship between SOC and OCV, and determine the fitting coefficient using the least squares method. The formula is: ; in, The coefficients are the fitting coefficients for the cubic term. The battery is in its state of charge. The fitting coefficients are quadratic terms. The coefficients for the first-order term are the fitting coefficients. The constant term fitting coefficients are, where, , , , The fitting coefficients are obtained by the least squares method. The SOC range is [0,1], and the OCV varies between [2.5V,4.2V].

[0032] Goodness of fit is evaluated using mean squared error (MSE), as shown in the formula: ; in, Mean square error, N For the sample size, The sample number. For the first The actual open-circuit voltage value corresponding to each sample. For the first The fitted open-circuit voltage value corresponding to each sample.

[0033] Step S23: Define the sum of squared errors between the model's predicted values ​​and the actual measured values. Identify the experimental data parameters using the least squares method. The objective function is to minimize the sum of squared errors between the model's predicted voltage and the actual measured voltage. By minimizing the error, optimize the model parameters.R 1. R 2. C 1 and C 2. The objective function formula is: ; in, To minimize the sum of squared errors between the model-predicted voltage and the actual measured voltage, i.e., the objective function, min(·) is the minimization operation function. For the selected window length, k For the data point sequence, For actual voltage measurement, Predict voltage for the model; The parameters are updated using gradient descent, as shown in the formula: ; in, For the first The parameter values ​​after the next iteration. for Parameter values ​​at time, The learning rate for the gradient descent algorithm; Step S3: Build a multi-task neural network model. Input real-time current, voltage, temperature, and historical average features. The output includes two independent paths: Path 1: Predict the battery open-circuit voltage (OCV); Path 2: Directly predict the battery state of charge (SOC), including the following steps: Step S31: Based on the original data and the data after parameter identification, predict the OCV using a neural network. Real-time current, voltage, and temperature data are used as input features, while also considering the historical average value and rate of change of voltage and current, and the historical SOC value and its rate of change. Feature normalization is performed using the following formula: ; in, These are the normalized eigenvalues. For data to be normalized, The maximum value in the dataset. The minimum value in the dataset; Choosing the right neural network, such as using a Long Short-Term Memory (LSTM) network to capture the time dependence of battery state, a Convolutional Neural Network (CNN) to extract local features, and combining multi-scale convolution operations to handle short-term dynamic changes; and an attention mechanism network to weight the contribution of historical data to improve adaptability to complex working conditions. Step S32: The neural network predicts the current open-circuit voltage through forward propagation. The model is trained using the preprocessed training set, and the Adam optimizer is used to optimize the weight parameters. The validation set is used to adjust hyperparameters such as the learning rate and the number of hidden layers. The loss functions are as follows: ; ; in, The loss function for OCV predictions. The loss function for the SOC predicted value. for The truth of every moment value, for Prediction of time value, for The predicted open-circuit voltage value at any given time; The data is divided into a training set (80%), a validation set (10%), and a test set (10%). Feature normalization is performed to ensure consistent scale across different physical quantities. The training set is used to optimize the neural network weights, and Dropout regularization or weight decay is added to prevent overfitting. Step S33: Monitor the root mean square error (RMSE) and mean absolute error (MAE) on the test set. When the model's generalization ability meets the application requirements, simultaneously output the OCV and SOC predicted values. Use the root mean square error (RMSE) to evaluate the model performance. The formula is: Using a neural network to predict OCV, in actual operation, data such as current, voltage, and temperature are acquired and normalized in real time, input into the neural network, and the neural network outputs the current predicted open-circuit voltage and battery SOC. The predicted value is compared with the measured value, and the error is calculated. If the error is less than a threshold, the predicted value is considered valid.

[0034] ; in, This represents the deviation between the predicted and measured values ​​of the open-circuit voltage (OCV). This is the predicted value of the open-circuit voltage (OCV). This is the measured value of the open-circuit voltage (OCV).

[0035] Step S4: Based on the predicted OCV value output by path one, obtain the initial SOC value using a lookup table method. Using the initial SOC value as a benchmark, dynamically update the SOC value using the ampere-hour integration method combined with real-time current. Periodically, compare the SOC value obtained from the lookup table with the SOC estimated by the ampere-hour integration method. A preset error range threshold is used. When the error is within the preset error range threshold, pause the neural network lookup correction. This includes the following steps: Step S41: Based on the predicted OCV value output by the neural network of path one, query the OCV-SOC mapping table to obtain the initial SOC value. The mapping table is established through the second-order RC model parameter identification experiment. Step S42: Using the initial SOC value as a reference, update the SOC value in real time using the ampere-hour integration method, and calculate the SOC change based on the real-time current. The formula is as follows: ; in, It is the nominal capacity, and the integration process is achieved through numerical integration using the trapezoidal integration method.

[0036] Step S5: Compare the results predicted by the ampere-hour integration method with the SOC value directly predicted by the neural network in real time, dynamically weight and fuse them according to the error threshold, and output the final optimal estimate of SOC. By combining SOC estimation and error correction, and integrating neural network prediction, table lookup method, and ampere-hour integration method, the final SOC is obtained, and the weights of each method are dynamically adjusted: ; in, For the final estimate Estimated value For the model Estimated value These are empirical weighting coefficients; Step S6: Set a vehicle inactivity time threshold. When the vehicle inactivity time exceeds the threshold, estimate the battery health status (SOH) using historical operating data and calibrate the SOC benchmark. Specifically: Dynamically update the State of Health (SOH) of the vehicle after it has been idle for an extended period, extracting the number of charge-discharge cycles before the extended period of inactivity. and average depth of discharge Based on battery operating data, lifespan models or data-driven methods are used to predict SOH (State of Health) and update the nominal capacity according to the SOH value. ; ; in, For the updated nominal capacity, This is the initial nominal capacity; After the vehicle is restarted, the nominal capacity is recalibrated based on the updated battery state of health (SOH), and the battery state of charge (SOC) is estimated based on this, thereby significantly improving the accuracy and reliability of SOC estimation.

[0037] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0038] Therefore, this invention proposes a multi-model fusion estimation method for SOC of lithium-ion batteries based on neural networks. By fusing battery models, ampere-hour integration, neural networks, and lookup table methods, the method dynamically adjusts and corrects the SOC estimation results, which can significantly improve the accuracy and adaptability of SOC estimation. By using neural networks to predict the nonlinear relationship of open-circuit voltage (OCV) in real time, combined with the ampere-hour integration method to dynamically update SOC, this method can not only cope with rapid dynamic changes under complex operating conditions, but also effectively suppress error accumulation in long-term operation. At the same time, this invention predicts the state of health (SOH) of the battery after long-term idling using historical data, providing a more reliable benchmark for SOC estimation, ensuring that the SOC estimation results remain highly accurate even when the vehicle is idle or aging. It is applicable to various scenarios and has high robustness and applicability.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-model fusion estimation method for the state of charge (SOC) of lithium-ion batteries based on neural networks, characterized in that, Includes the following steps: Step S1: Extract the measured data of the hybrid power pulse characteristic (HPPC) of lithium-ion batteries, segment the data according to the current direction, divide each data segment into data fragments according to the time window or the change amplitude of battery charge state (SOC), remove irrelevant information and data noise, use linear interpolation to complete the missing data and normalize it to obtain the preprocessed data. Step S2: Construct an equivalent circuit model of a lithium-ion battery based on the preprocessed data, conduct a constant current charge-discharge experiment, fit the SOC-OCV curve using a polynomial function based on the obtained experimental raw data, and identify the experimental data parameters using the least squares method. Step S3: Build a multi-task neural network model. Input real-time current, voltage, temperature and historical average features. The output includes two independent paths: Path 1: Predict the battery open-circuit voltage OCV; Path 2: Directly predict the battery state of charge SOC. Step S4: Based on the predicted OCV value output by path one, obtain the initial SOC value by looking up a table. Using the initial SOC value as a reference, dynamically update the SOC value by combining the ampere-hour integration method with the real-time current. Step S5: Compare the results predicted by the ampere-hour integration method with the SOC value directly predicted by the neural network in real time, dynamically weight and fuse them according to the error threshold, and output the final optimal estimate of SOC. Step S6: Set a vehicle idle time threshold. When the vehicle idle time exceeds the vehicle idle time threshold, estimate the battery health status (SOH) based on historical operating data and calibrate the SOC benchmark.

2. The method for multi-model fusion estimation of lithium-ion battery SOC based on neural networks according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Divide the HPPC measured data into charging and discharging segments according to the current direction, and divide each segment into data fragments based on the time window or the SOC change amplitude. Step S12: Remove irrelevant information by using a low-pass filter to remove data noise. The cutoff frequency of the filter is set to be less than twice the main frequency of the signal. Step S13: Use linear interpolation to fill in the missing data, normalize the filled data, and eliminate the dimensional differences between different physical quantities.

3. The method for estimating the state of charge (SOC) of a lithium-ion battery based on a neural network according to claim 2, characterized in that: In step S12, irrelevant information includes, but is not limited to, test steps and relative timestamps.

4. The method for multi-model fusion estimation of lithium-ion battery SOC based on neural networks according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: Establish a second-order RC circuit model, where the fast polarization part is composed of... R 1 and C 1. Description: The slow polarization component is composed of... R 2 and C 2. Description, the formula is: ； in, For real-time voltage, Open circuit voltage, For real-time current, For ohmic internal resistance, For rapid polarization voltage, This is a slow polarization voltage; Step S22: Based on the constant current charge-discharge experiment, record the OCV values ​​under different SOCs. Use a cubic polynomial to fit the relationship between SOC and OCV, and determine the fitting coefficients using the least squares method. The formula is: ； in, The coefficients are the fitting coefficients for the cubic term. The battery is in its state of charge. The fitting coefficients are quadratic terms. The coefficients for the first-order term are the fitting coefficients. The fitting coefficients are constant terms; Step S23: Define the sum of squared errors between the model's predicted values ​​and the actual measured values. Identify the experimental data parameters using the least squares method. The objective function is to minimize the sum of squared errors between the model's predicted voltage and the actual measured voltage. By minimizing the error, optimize the model parameters. R 1. R 2. C 1 and C 2. The objective function formula is: ； in, To minimize the sum of squared errors between the model-predicted voltage and the actual measured voltage, i.e., the objective function, min(·) is the minimization operation function. For the selected window length, k For the data point sequence, For actual voltage measurement, Predict voltage for the model.

5. The method for multi-model fusion estimation of lithium-ion battery SOC based on neural networks according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: Construct the initial model of the neural network. The input data of the input layer includes, but is not limited to, real-time current, real-time voltage and temperature. The hidden layer contains two LSTM layers and one fully connected layer. The output layer contains two independent paths and outputs the predicted values ​​of open-circuit voltage (OCV) and charge state (SOC). Define the hyperparameters of the network, including the number of layers, the number of neurons and the activation function. Step S32: Divide the preprocessed data into training set, validation set and test set, optimize the weights using Adam optimizer, adjust the network weights through forward propagation and back propagation algorithms to minimize prediction error, and introduce regularization to prevent overfitting. Step S33: Monitor the root mean square error (RMSE) and mean absolute error (MAE) in the test set. When the model's generalization ability meets the application requirements, output the OCV prediction value and the SOC prediction value simultaneously.

6. The method for multi-model fusion estimation of lithium-ion battery SOC based on neural networks according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: Based on the predicted OCV value output by the neural network of path one, query the OCV-SOC mapping table to obtain the initial SOC value. The mapping table is established through the second-order RC model parameter identification experiment. Step S42: Based on the initial SOC value, update the SOC value in real time using the ampere-hour integration method, with the following formula: ； in, To obtain the predicted value through ampere-hour integration, For the first k The time corresponding to each time sampling point This is the initial value of SOC. Nominal capacity For Coulomb efficiency, for t The current at any given time is integrated numerically using the trapezoidal rule.

7. The method for multi-model fusion estimation of lithium-ion battery SOC based on neural networks according to claim 1, characterized in that: In step S4, the SOC value obtained by looking up the table is compared with the SOC estimated by the ampere-hour integration method at regular intervals. A preset error range threshold is set. When the error is within the preset error range threshold, the neural network table lookup correction is paused.

8. The method for multi-model fusion estimation of lithium-ion battery SOC based on neural networks according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: Calculate the absolute error between the ampere-hour integral prediction and the neural network prediction in real time. The formula is: ； in, The absolute error between the ampere-hour integral prediction and the neural network prediction. The value is a neural network prediction. Step S52: Dynamically adjust the fusion weights based on the error threshold, using the following formula: ； in, This is the final value of the SOC result.

9. The method for multi-model fusion estimation of lithium-ion battery SOC based on neural networks according to claim 1, characterized in that: Step S6 includes the following steps: Step S61: Extract historical operating data, including the number of complete charge-discharge cycles, average depth of discharge, and historical average ambient temperature. Step S62: Input the lifetime decay model to calculate SOH. The calculation formula is as follows: ； in, SOH For battery health status, For a complete charge-discharge cycle, The average depth of discharge. This represents the historical average ambient temperature. , , , These are the model calibration coefficients; Step S63: Update the nominal capacity using the following formula: ； in, For the new capacity parameter of the ampere-hour integration method, This is the rated capacity.

10. The method for multi-model fusion estimation of lithium-ion battery SOC based on neural networks according to claim 9, characterized in that: In step S62, , , , Through experimental calibration, the calibration process needs to cover the number of charge-discharge cycles, depth of discharge, and ambient temperature range throughout the battery's entire life cycle.

Images