Lithium-ion battery charge and health state estimation method based on multi-model fusion

Abstract

The present invention is applicable to the field of lithium-ion battery management technology, and provides a lithium-ion battery charge and health state estimation method based on multi-model fusion, including the following steps: data acquisition and processing; battery SOH estimation based on equal voltage interval discharge time; battery SOC estimation model construction and training; ampere-hour integration method based on capacity correction: applying the SOH value calculated in step 2 to the ampere-hour integration method to adjust the estimated value of SOC in real time; estimation of battery SOC under current battery SOH: applying the SOC output by the neural network in step 3 and the SOC calculated by the ampere-hour integration method in step 4 to the Kalman filter for fusion; evaluation index setting. The present invention significantly improves the estimation accuracy of lithium-ion battery SOC and SOH, reduces errors, enhances BMS performance, improves battery safety and service life, and exhibits strong adaptability and good anti-noise performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium-ion battery management, and in particular relates to a lithium-ion battery charge and health state estimation method based on multi-model fusion. Background Art

[0002] In view of the increasingly serious problems of global warming and environmental pollution, it is crucial to promote the reduction of carbon dioxide emissions and improve energy efficiency. Electric vehicles have gradually become mainstream by replacing fossil fuels with batteries. Among them, lithium-ion batteries have become the core power of electric vehicles due to their excellent performance in energy and power density, self-discharge rate, charge and discharge rate, and cycle life. The introduction of battery management system (BMS) has greatly improved the safety and efficiency of battery use. It can effectively monitor the remaining power (i.e. state of charge, SOC) and health status (i.e. state of health, SOH) of the battery to ensure the safety and reliability of the vehicle under various driving conditions.

[0003] At present, the estimation methods of SOC and SOH are mainly divided into two categories: model-based methods and data-driven methods. Model-based methods rely on a deep understanding of the internal mechanism of the battery and simulate the behavior of the battery by establishing an electrochemical model or equivalent circuit model of the battery. These methods have different focuses in the estimation of SOC and SOH. SOC estimation: Model-based methods include open circuit voltage method, ampere-hour integration method and electrochemical model method. Although the open circuit voltage method is suitable for static estimation, it has poor real-time performance; the ampere-hour integration method is simple and intuitive, but the accumulated error and sensitivity to the initial SOC value limit its long-term accuracy; the electrochemical model law improves accuracy by simulating the internal reaction of the battery in detail, but the calculation is complex and requires high parameter identification. SOH estimation: In addition to the above methods, electrochemical impedance spectroscopy and particle filtering methods are also applied to SOH estimation. Electrochemical impedance spectroscopy identifies the health state by measuring the internal impedance characteristics of the battery, but it is costly and complex to operate; the particle filtering law estimates SOH by simulating the propagation of particles in the state space, which has a large amount of calculation and the accuracy is affected by the accuracy of the model. Data-driven methods use a large amount of data collected from battery sensors (such as voltage, current, temperature) to train models to achieve accurate estimation of SOC and SOH. Common points between SOC and SOH: Data-driven methods usually include machine learning (ML) such as artificial neural networks (ANN), support vector machines (SVM), gradient boosting decision trees (GBDT), structured neural networks (SNN), back propagation neural networks (BPNN), etc., as well as statistical techniques such as Gaussian process regression (GPR). These methods establish a nonlinear relationship between input data and SOC / SOH by extracting key features from battery data: indirect features (or internal features), such as the SOC, internal capacitance and resistance of the battery; processed features (usually extracted from charge and discharge data to generate differential curves, such as incremental capacity (IC) and differential voltage (DV) curves); direct features (or external features). Although data-driven methods perform well in both SOC and SOH estimation, they both require a large amount of data support and are at risk of overfitting and high computational cost.

[0004] In practical applications, the capacity loss caused by battery aging will interfere with the accurate estimation of SOC, so relying solely on SOC may not accurately reflect the actual working capacity and remaining service life of the battery. Therefore, a method for estimating the charge and health state of lithium-ion batteries based on multi-model fusion is proposed. Summary of the invention

[0005] The purpose of the present invention is to provide a lithium-ion battery charge and health state estimation method based on multi-model fusion, aiming to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The method for estimating the charge and health state of a lithium-ion battery based on multi-model fusion includes the following steps:

[0008] Step 1: Data collection and processing, including:

[0009] Data collection: In the SOC estimation task, the input is voltage, current and temperature; in the SOH estimation task, the input is the time span value of the 4.0V-3.2V discharge window;

[0010] Data processing: normalize, batch and window the data;

[0011] Step 2: Battery SOH estimation based on equal voltage interval discharge time;

[0012] Step 3: Building and training the battery SOC estimation model;

[0013] Step 4: Ampere-hour integration method based on capacity correction: Apply the SOH value calculated in step 2 to the Ampere-hour integration method to adjust the estimated SOC value in real time;

[0014] Step 5: Estimation of the battery SOC under the current battery SOH: The SOC output by the neural network in step 3 and the SOC calculated by the ampere-hour integration method in step 4 are fused by using a Kalman filter;

[0015] Step 6: Setting evaluation indicators.

[0016] Furthermore, in step 1, the specific process of the data processing step is as follows:

[0017] Normalize the acquired data: Use the maximum and minimum normalization method to scale the eigenvalues ​​of the index to the interval [0,1]; Suppose the input data is l , the normalization process is as follows:

[0018]

[0019] in, l min is the minimum value in the feature sequence, l max is the maximum value in the feature sequence, l * is the normalized value;

[0020] After normalization, the data is processed in batches and windows; after batch processing, the time series data is converted into a format suitable for neural network processing, and the window length is selected as 60. The actual prediction process is to use the input features of the first 60 time points to predict the 61st target value, and then gradually move forward through the time window to predict the entire time series, thereby realizing dynamic monitoring and prediction of battery SOC and SOH.

[0021] Furthermore, the specific process of step 2 is as follows:

[0022] At room temperature, the charging is performed at a constant current of 1.5A, the charging cut-off voltage is 4.2V, and the discharging is performed at a constant current of 2A, the discharging cut-off voltage is 2.4V, and the discharging window is defined as 3.8V-2.8V, and the discharging time t is intercepted. 100 ,t 300 and t 600 , the battery SOH is estimated based on the change of discharge time at equal voltage intervals as the number of cycles increases;

[0023] After obtaining the equal-voltage interval discharge time series of different cycles, the normalization, batching and windowing processing in step 1 are first performed, and then a SOH estimation model with LSTM network as the core is constructed, with the processed time series as input and SOH as output.

[0024] Furthermore, the specific process of step three is as follows:

[0025] The SOH window is selected as 100%-80%, and the SOC estimation model is built under 100% SOH, 95% SOH, 90% SOH, 85% SOH and 80% SOH respectively; the input variables of the model are voltage, current and temperature; the collected data is first normalized, batched and windowed in step one, and then the processed input data enters the CNN-LSTM network. The CNN-LSTM network includes a CNN layer, an LSTM layer, a fully connected layer and an output layer. The CNN layer performs preliminary feature extraction, and then the preliminary extracted features are passed to the LSTM layer. The LSTM layer further analyzes the dynamic changes of the features in the time series, captures the evolution of battery performance over time, and outputs the SOC after passing through the fully connected layer and the output layer.

[0026] Furthermore, in step 4, the ampere-hour integration method based on capacity correction is expressed by the following formula:

[0027]

[0028] in, SOC k+1 Represents the SOC at the next moment; SOC kRepresents the SOC at the current moment; I Indicates current; Δ t represents the time interval; C is the rated capacity of the battery; SOH is the estimated battery health status value.

[0029] Furthermore, the specific process of step five is as follows:

[0030] First, calculate the estimated SOC value of the CNN-LSTM network under the current SOH: determine the current SOH, assuming its value is x, and then calculate the difference between the current SOH value and the five fixed SOH nodes, namely 100% SOH, 95% SOH, 90% SOH, 85% SOH and 80% SOH, assuming the difference values ​​are m1, m2, m3, m4 and m5 respectively, and assuming that the estimated SOC under the five fixed SOH nodes are SOC 1, SOC 2, SOC 3. SOC 4 and SOC 5, then the estimated SOC value under the current SOH is:

[0031]

[0032] Secondly, the state space expression is established: For the SOC estimation problem, the state vector x is the SOC value calculated by the ampere-hour integration method, and the measurement vector z is the SOC value output by the CNN-LSTM network, that is, the SOC estimation value of the CNN-LSTM network under the current SOH calculated in Formula 4. Therefore, the established state space expression is as follows:

[0033]

[0034] in, SOC k+1 Represents the SOC value calculated by ampere-hour integration at time k+1; SOC k represents the SOC value at time k, I k is the current value at time k, represents a time interval; SOH is the estimated battery health status value; C is the rated capacity of the battery; ω is the Gaussian state noise, υ is the Gaussian measurement noise; LSTM k Represents the SOC value output by the CNN-LSTM network at time k.

[0035] Formula 2 introduces the proposed theory of the ampere-hour integration method based on capacity correction, and Formula 5 establishes the theory of Formula 2 in the form of a state space expression. In terms of meaning, the two are the same.

[0036] The standard form of state space expression is:

[0037]

[0038] in, x k is the state variable at time k; y k is the output variable at time k; A is the transfer matrix, B is the system input matrix, C is the system output matrix, D is the feedforward matrix, ω k is the process noise, υ k To measure noise;

[0039] Finally, the implementation of the Kalman filter is divided into two steps: time update and measurement update:

[0040] Time update:

[0041]

[0042] in, is the prior state estimate at time k; is the estimated value of the posterior state at time k; is the prior estimated covariance at time k; is the control input; P k-1 is the posterior estimated covariance at time k-1; is the transpose of the state transfer matrix; Q is the process excitation noise covariance;

[0043] Measurement Update:

[0044]

[0045] in, K k is the Kalman gain; is the prior estimated covariance at time k; H is the conversion matrix from state variables to measurements; for H The transpose of R is the measurement noise covariance; is the estimated value of the posterior state at time k; is the prior state estimate at time k; is the measured value; P k is the optimal estimate of the error covariance matrix at time point k; IIndicates current;

[0046] Applying Equation 8 and Equation 9 to Equation 5 and Equation 6, the output is ,Right now:

[0047]

[0048] Furthermore, the specific process of step six is ​​as follows:

[0049] set up RMSE The indicator is used to evaluate the performance of the model. The calculation formula is as follows:

[0050]

[0051] in, SOC k represents the SOC value at time k; SOC * k is the predicted SOC value at time k; N is the sample size.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. The present invention discloses a method for estimating the charge and health state of a lithium-ion battery based on multi-model fusion, which covers multiple key steps: battery SOH estimation based on equal voltage interval discharge time, estimating the current SOH by finding the relationship between the time of the same discharge voltage interval in different cycles and the SOH; building and training a battery SOC estimation model, estimating the SOC by establishing a neural network model under different SOHs; battery SOC estimation based on capacity correction, weighting the initial battery capacity by the estimated SOH and then applying the ampere-hour integration method to estimate the SOC under the current SOH; SOC estimation based on a Kalman filter, fusing the SOC output by the neural network with the SOC calculated by the ampere-hour integration method, further reducing the estimation error.

[0054] 2. The state estimation of the present invention in the actual vehicle application scenario has practical significance, which is mainly reflected in the following aspects:

[0055] 1) Improved the accuracy of battery state estimation: Through the multi-model fusion method, the battery SOH estimation based on equal voltage interval discharge time, SOC estimation by neural network, battery SOC estimation based on capacity correction and SOC estimation based on unscented Kalman filter fusion are combined. The fusion of these methods improves the accuracy of lithium-ion battery SOC and SOH estimation.

[0056] 2) Reduced estimation error: The use of Kalman filtering combines the SOC output by the neural network with the SOC calculated by the ampere-hour integration method. By integrating these two methods, the estimation error is further reduced and the reliability of the estimation is improved.

[0057] 3) Enhanced BMS performance: Accurate SOC and SOH estimation is crucial for battery management systems. The technology of the present invention can provide BMS with more accurate battery status information, thereby optimizing battery charge and discharge management and extending battery life.

[0058] 4) Improved battery safety and service life: Through real-time and accurate battery status monitoring, abnormal battery conditions can be detected and intervened in time, thereby eliminating safety hazards in advance and extending battery life.

[0059] 5) Strong adaptability: This method is applicable to different battery aging states and temperature conditions, and can provide accurate battery state estimation in all-climate and full-life cycle operation, enhancing the adaptability of the battery in various environments.

[0060] 6) Good anti-noise performance: The use of Kalman filtering can make real-time estimation and correction of the system process and measurement noise covariance under unknown noise conditions, thereby improving the stability of the filtering effect and maintaining a high estimation accuracy even under the influence of noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Flow chart of the method of the present invention.

[0062] Figure 2 Schematic diagram of time series construction; wherein, a is a schematic diagram of SOC estimation task time series construction, and b is a schematic diagram of SOH estimation task time series construction.

[0063] Figure 3 The graph shows the discharge voltage of a lithium-ion battery changing with time at the 100th, 300th, and 600th cycles.

[0064] Figure 4 This is the CNN-LSTM network structure diagram.

[0065] Figure 5 This is the Kalman filter fusion framework diagram.

[0066] Figure 6 The figure is a comparison chart of the results of different SOC estimation methods at 90% SOH under 25℃ conditions; among them, a is the SOC estimation result of the unified neural network SOC estimation method, b is the SOC estimation result of the SOC estimation method considering the influence of SOH, and c is the SOC estimation result of the SOC estimation method without considering the influence of SOH.

[0067] Figure 7 This is a comparison chart of the errors of different SOC estimation methods at 90% SOH at 25°C. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0069] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0070] An embodiment of the present invention provides a method for estimating the charge and health status of a lithium-ion battery based on multi-model fusion, such as Figure 1 As shown, the following steps are included:

[0071] Step 1: Data collection and processing, including:

[0072] Data collection: In the SOC estimation task, the input is voltage, current and temperature; in the SOH estimation task, the input is the time span value of the 4.0V-3.2V discharge window;

[0073] Data processing: normalize, batch and window the data; the specific process is as follows:

[0074] The maximum and minimum normalization method is used to scale the eigenvalues ​​of the index to the interval [0,1]. Suppose the input data is l , the normalization process is as follows:

[0075]

[0076] in, l min is the minimum value in the feature sequence, l max is the maximum value in the feature sequence, l * is the normalized value;

[0077] After normalization, the data input to the neural network will be processed in batches and windows. The reason why batch processing is performed first after obtaining the original data is that directly processing the entire data set will result in a large consumption of computing resources and may cause insufficient memory. Batch processing can effectively reduce memory usage, making the data processing process more efficient and feasible. Secondly, window division is performed to convert the time series data into a format suitable for neural network processing. Figure 2As shown in a and b, the time series construction process in the SOC estimation task and the SOH estimation task, respectively, is selected as 512 batches. The entire data set is divided into batches and then windowed. The window length is selected as 60. The actual prediction process is to use the input features of the first 60 time points to predict the 61st target value. In the SOC estimation task, the input feature X is voltage, current, and temperature. Batch 1 contains voltage, current, and temperature information from the 1st time point to the 512th time point, and batch 2 contains voltage, current, and temperature information from the 513th time point to the 1024th time point. And so on. The total length of the data set is 33310, and the last batch n contains voltage, current, and temperature information from the 32710th to the 33310th time point. The first window contains voltage, current, and temperature information from the 1st time point to the 60th time point, and the information in other windows slides backward with a step size of 1. In the SOH estimation task, the input feature X is the time of the equal voltage discharge interval. Batch 1 contains the time information of the equal voltage discharge interval from the 1st time point to the 512th time point, and batch 2 contains the time information of the equal voltage discharge interval from the 513th time point to the 1024th time point. The total length of the data set is 33310, and the last batch n contains the time information of the equal voltage discharge interval from the 32710th time point to the 33310th time point. The first window contains the time information of the equal voltage discharge interval from the 1st time point to the 60th time point, and the information in other windows slides backward with a step size of 1. This prediction method is based on the sliding window technology in time series analysis, which allows the model to use historical data to predict future trends. By moving the time window forward step by step, the entire time series is predicted to achieve dynamic monitoring and prediction of battery SOC and SOH.

[0078] Step 2: Battery SOH estimation based on equal voltage interval discharge time;

[0079] Battery SOH estimation based on equal voltage interval discharge time is a method to evaluate the current SOH of the battery by using the voltage change characteristics during the battery discharge process. This estimation method usually focuses on the time interval required for the battery voltage to drop to a specific threshold during discharge. These time intervals can reflect the changes in the internal chemical and physical state of the battery, which is associated with the aging degree and performance degradation of the battery. Figure 3 The figure shows some results of a lithium-ion battery charge and discharge experiment, specifically the discharge voltage over time of the 100th, 300th and 600th cycles. The experiment process is at room temperature, charging with a constant current of 1.5A, the charging cut-off voltage is 4.2V, discharging with a constant current of 2A, the discharge cut-off voltage is 2.4V, and the discharge window is defined as 3.8V-2.8V, and the discharge time t is intercepted. 100 ,t300 and t 600 It can be seen that with the increase of cycle number, the discharge time of equal voltage interval gradually decreases. The battery SOH can be estimated based on this rule.

[0080] After obtaining the equal-voltage interval discharge time series of different cycles, the normalization, batching and windowing processing in step 1 are first performed, and then a SOH estimation model with LSTM network as the core is constructed, with the processed time series as input and SOH as output.

[0081] Step 3: Building and training the battery SOC estimation model;

[0082] SOH is a measure of the current health of the battery, usually defined as the ratio of the current capacity of the battery to its rated capacity. It reflects the performance loss of the battery due to aging and degradation. The SOH value gradually decreases from 100% (new battery) to a certain threshold (such as 70% or 80%), at which point the battery is considered to have reached the end of its life. Therefore, the SOH window selected in the present invention is 100%-80%, and the SOC estimation model is built at 100% SOH, 95% SOH, 90% SOH, 85% SOH and 80% SOH respectively; the input variables of the model are voltage, current and temperature. The collected data is first normalized, batched and windowed in step one. Secondly, the processed input data enters the CNN-LSTM network. The CNN-LSTM network includes a CNN layer, an LSTM layer, a fully connected layer and an output layer ( Figure 4 ), the CNN layer performs preliminary feature extraction, and then the preliminary extracted features are passed to the LSTM layer. The LSTM layer further analyzes the dynamic changes of the features in the time series, captures the evolution of battery performance over time, and outputs the SOC after the fully connected layer and the output layer. The fully connected layer receives the output from the LSTM layer, which is usually a series of feature vectors of time steps. The fully connected layer integrates these feature vectors into a single feature vector to provide a more abstract representation for the output layer's decision. The output layer is usually one or more neurons, whose task is to make the final prediction decision. In the case of SOC prediction, there is only one neuron in the output layer, which is used to predict the remaining battery power. This deep learning architecture that combines CNN and LSTM not only extracts the local features and timing features of the battery state in parallel, but also the CNN network can simultaneously extract the coupling features between the three variables of voltage, current and temperature. The network enhances the model's ability to predict battery SOC changes through feature fusion.

[0083] Step 4: Ampere-hour integration method based on capacity correction: Apply the SOH value calculated in step 2 to the Ampere-hour integration method to adjust the estimated SOC value in real time;

[0084] The traditional ampere-hour integration method calculates the charge and discharge state of the battery based on the integration of the battery's current over time. It is mainly divided into the following steps: First, determine the initial SOC value of the battery. Secondly, measure the battery current in real time during the battery charging and discharging process. The current can be positive (charging) or negative (discharging). The measured current is then integrated over time. If the current is positive, it means that the battery is charging and the SOC will increase; if the current is negative, it means that the battery is discharging and the SOC will decrease. The integral formula can be expressed as:

[0085]

[0086] in, SOC k+1 Represents the SOC at the next moment; SOC k Represents the SOC at the current moment; I Indicates current; Δ t represents the time interval; C is the rated capacity of the battery.

[0087] Since the battery will be accompanied by capacity decay during use, it is necessary to correct the rated capacity of the battery. The ampere-hour integration method based on capacity correction is of great necessity to improve the accuracy of SOC estimation. This method dynamically corrects the rated capacity of the battery by considering the actual capacity changes caused by factors such as aging, temperature changes, and charge and discharge rates during the use of the battery. Specifically, the capacity correction model uses the charge and discharge test data of the battery under different working conditions to obtain a capacity correction factor related to the current state of the battery. This correction factor is the SOH value calculated in step two, which is applied to the ampere-hour integration method to adjust the estimated value of SOC in real time. It is expressed as the following formula:

[0088]

[0089] in, SOH is the estimated battery health status value.

[0090] Step 5: Estimation of the battery SOC under the current battery SOH: The SOC output by the neural network in step 3 and the SOC calculated by the ampere-hour integration method in step 4 are fused by using a Kalman filter;

[0091] This step solves the problem of overfitting and excessive burrs when the neural network estimates the SOC, as well as the problem of the ampere-hour integration method's dependence on the accuracy of the measured current. Figure 5 As shown, the specific process is as follows:

[0092] First, calculate the estimated SOC value of the CNN-LSTM network under the current SOH: determine the current SOH, assuming its value is x, and then calculate the difference between the current SOH value and the five fixed SOH nodes, namely 100% SOH, 95% SOH, 90% SOH, 85% SOH and 80% SOH, and set the difference values ​​to m1, m2, m3, m4 and m5 respectively. Since the smaller the difference, the closer the SOH, the output result should be given a greater weight, and the estimated SOC under the five fixed SOH nodes is SOC 1, SOC 2, SOC 3. SOC 4 and SOC 5, then the estimated SOC value under the current SOH is:

[0093]

[0094] Secondly, the state space expression is established: For the SOC estimation problem, the state vector x is the SOC value calculated by the ampere-hour integration method, and the measurement vector z is the SOC value output by the CNN-LSTM network, that is, the SOC estimation value of the CNN-LSTM network under the current SOH calculated in Formula 4. Therefore, the established state space expression is as follows:

[0095]

[0096] in, SOC k+1 Represents the SOC value calculated by ampere-hour integration at time k+1; SOC k represents the SOC value at time k, I k is the current value at time k, represents a time interval; SOH is the estimated battery health status value; C is the rated capacity of the battery; ω is the Gaussian state noise, υ is the Gaussian measurement noise; LSTM k Represents the SOC value output by the CNN-LSTM network at time k. The standard form of the state space expression is:

[0097]

[0098] in, x k is the state variable at time k; y k is the output variable at time k; A is the transfer matrix, B is the system input matrix, C is the system output matrix, D is the feedforward matrix; ω k is process noise, which is used to describe the state noise caused by insufficient accuracy of system modeling and the error caused by unknown input during the state transfer process; υ k It is a measurement of noise that describes the error caused by inaccurate input measured by the system sensors or external interference.

[0099] The following is an introduction to the implementation of the Kalman filter (KF) in this form, which is divided into two steps: time update and measurement update:

[0100] Time update:

[0101]

[0102] in, is the prior state estimate at time k, which is the intermediate calculation result of filtering, that is, the result at time k predicted based on the optimal estimate at the previous time (time k-1), which is the result of the prediction equation; is the posterior state estimate at time k, which is one of the filtering results, that is, the updated result, also called the optimal estimate; is the prior estimated covariance at time k ( The covariance of is the intermediate calculation result of filtering; is the control input; P k-1 is the posterior estimated covariance at time k-1; is the transpose of the state transfer matrix; Q is the process excitation noise covariance;

[0103] Measurement Update:

[0104]

[0105] in, K k is the Kalman gain; is the prior estimated covariance at time k ( covariance of ); H is the transformation matrix from state variables to measurements (observations); for H The transpose of R is the measurement noise covariance; is the estimated value of the posterior state at time k; is the prior state estimate at time k, which is the intermediate calculation result of filtering, that is, the result at time k predicted based on the optimal estimate at the previous time (time k-1), which is the result of the prediction equation; is the measured value (observation value), which is the input of the filter; Pk is the optimal estimate of the error covariance matrix at time point k; I Indicates the current. Through this step, the Kalman filter can be used to fuse the SOC value calculated by the ampere-hour integration method and the SOC value output by the CNN-LSTM network to achieve the purpose of smooth output.

[0106] The above describes the working process of the Kalman filter in the form and parameters of the standard Kalman formula. When applied to the state space expression (Equation 6 and Equation 7), the output is ,Right now:

[0107]

[0108] Step 6: Setting evaluation indicators;

[0109] The root mean square error (RMSE) indicator is set to evaluate the performance of the model:

[0110]

[0111] in, SOC k represents the SOC value at time k; SOC * k is the predicted SOC value at time k; N is the sample size. RMSE The smaller the value, the smaller the difference between the model's predicted value and the actual value, and the higher the model's prediction accuracy.

[0112] In order to verify the effectiveness of the SOC and SOH joint estimation method proposed in this invention, a comparison is made with the traditional neural network SOC estimation method in a 90% SOH dataset at 25°C. Figure 6 As shown in the figure, a shows the SOC estimation result of the traditional neural network SOC estimation method; b shows the SOC estimation result of the method of the present invention, that is, the SOC estimation method considering the influence of SOH (after Kalman filter correction); c shows the SOC estimation result of the SOC estimation method not considering the influence of SOH (after Kalman filter correction). RMSE comparison Figure 7 As shown. Figure 6 and Figure 7It can be seen that the SOC estimation method that does not consider the influence of SOH will cause a large deviation, and accordingly, the estimation error will also increase. Compared with the SOC estimation method that does not consider the influence of SOH, the RMSE level of the method of the present invention in the 90% SOH data set at 25°C is reduced by 59.36%, and compared with the traditional neural network SOC estimation method in the 90% SOH data set at 25°C, the RMSE level is reduced by 24.15%, which verifies the effectiveness of the method proposed in the present invention.

[0113] The above are only preferred embodiments of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention. These should also be regarded as the protection scope of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A lithium-ion battery charge and health state estimation method based on multi-model fusion, characterized in that: The following steps are involved: Step 1: Data collection and processing, including: Data collection: In the SOC estimation task, the input is voltage, current and temperature; in the SOH estimation task, the input is the time span value of the 4.0V-3.2V discharge window; Data processing: normalize, batch and window the data; Step 2: Battery SOH estimation based on equal voltage interval discharge time; Step 3: Building and training the battery SOC estimation model; Step 4: Ampere-hour integration method based on capacity correction: Apply the SOH value calculated in step 2 to the Ampere-hour integration method to adjust the estimated SOC value in real time; Step 5: Estimation of the battery SOC under the current battery SOH: The SOC output by the neural network in step 3 and the SOC calculated by the ampere-hour integration method in step 4 are fused by using a Kalman filter; Step 6: Setting evaluation indicators; In step 1, the specific process of the data processing step is as follows: Normalize the acquired data: Use the maximum and minimum normalization method to scale the eigenvalues ​​of the index to the interval [0,1]; Suppose the input data is l , the normalization process is as follows: in, l min is the minimum value in the feature sequence, l max is the maximum value in the feature sequence, l * is the normalized value; After normalization, the data is processed in batches and windows. After batch processing, the time series data is converted into a format suitable for neural network processing. The window length is selected as 60. The actual prediction process is to use the input features of the first 60 time points to predict the 61st target value. The time window is gradually moved forward to predict the entire time series, realizing dynamic monitoring and prediction of battery SOC and SOH. The specific process of step 2 is as follows: At room temperature, the charging is performed at a constant current of 1.5A, the charging cut-off voltage is 4.2V, and the discharging is performed at a constant current of 2A, the discharging cut-off voltage is 2.4V, and the discharging window is defined as 3.8V-2.8V, and the discharging time t is intercepted. 100 ,t 300 and t 600 , the battery SOH is estimated based on the change of discharge time at equal voltage intervals as the number of cycles increases; After obtaining the time series of equal voltage interval discharges of different cycles, the normalization, batching and windowing in step 1 are first performed, and then a SOH estimation model with LSTM network as the core is constructed, with the processed time series as input and SOH as output; The specific process of step three is as follows: The SOH window is selected as 100%-80%, and the SOC estimation model is built under 100% SOH, 95% SOH, 90% SOH, 85% SOH and 80% SOH respectively; the input variables of the model are voltage, current and temperature; the collected data is first normalized, batched and windowed in step 1, and then the processed input data enters the CNN-LSTM network. The CNN-LSTM network includes a CNN layer, an LSTM layer, a fully connected layer and an output layer. The CNN layer performs preliminary feature extraction, and then the preliminary extracted features are passed to the LSTM layer. The LSTM layer further analyzes the dynamic changes of the features in the time series, captures the evolution of battery performance over time, and outputs the SOC after passing through the fully connected layer and the output layer; In step 4, the ampere-hour integration method based on capacity correction is expressed by the following formula: in, SOC k+1 Represents the SOC at the next moment; SOC k Represents the SOC at the current moment; I Indicates current; Δ t represents the time interval; C is the rated capacity of the battery; SOH is the estimated battery health status value; The specific process of step five is as follows: First, calculate the estimated SOC value of the CNN-LSTM network under the current SOH: determine the current SOH, assuming its value is x, and then calculate the difference between the current SOH value and the five fixed SOH nodes, namely 100% SOH, 95% SOH, 90% SOH, 85% SOH and 80% SOH, assuming the difference values ​​are m1, m2, m3, m4 and m5 respectively, and assuming that the estimated SOC under the five fixed SOH nodes are SOC 1, SOC 2, SOC 3. SOC 4 and SOC 5, then the estimated SOC value under the current SOH is: Secondly, the state space expression is established: For the SOC estimation problem, the state vector x is the SOC value calculated by the ampere-hour integration method, and the measurement vector z is the SOC value output by the CNN-LSTM network, that is, the SOC estimation value of the CNN-LSTM network under the current SOH calculated in Formula 4. Therefore, the established state space expression is as follows: in, SOC k+1 Represents the SOC value calculated by ampere-hour integration at time k+1; SOC k represents the SOC value at time k, I k is the current value at time k, represents a time interval; SOH is the estimated battery health status value; C is the rated capacity of the battery; ω is the Gaussian state noise, υ is the Gaussian measurement noise; LSTM k Represents the SOC value output by the CNN-LSTM network at time k; the standard form of the state space expression is: in, x k is the state variable at time k; y k is the output variable at time k; A is the transfer matrix, B is the system input matrix, C is the system output matrix, D is the feedforward matrix, ω k is the process noise, υ k To measure noise; Finally, the implementation of the Kalman filter is divided into two steps: time update and measurement update: Time update: in, is the prior state estimate at time k; is the estimated value of the posterior state at time k; is the prior estimated covariance at time k; is the control input; P k-1 is the posterior estimated covariance at time k-1; is the transpose of the state transfer matrix; Q is the process excitation noise covariance; Measurement Update: in, K k is the Kalman gain; is the prior estimated covariance at time k; H is the conversion matrix from state variables to measurements; for H The transpose of R is the measurement noise covariance; is the estimated value of the posterior state at time k; is the prior state estimate at time k; is the measured value; P k is the optimal estimate of the error covariance matrix at time point k; I Indicates current; Applying Equation 8 and Equation 9 to Equation 5 and Equation 6, the output is ,Right now: The specific process of step six is ​​as follows: set up RMSE The indicator is used to evaluate the performance of the model. The calculation formula is as follows: in, SOC k represents the SOC value at time k; SOC * k is the predicted SOC value at time k; N is the sample size.

Images