A lithium battery state-of-charge assessment method based on LightGBM optimized square root central difference Kalman filter.

Abstract

This invention discloses a lithium battery state-of-charge (SOC) assessment method based on LightGBM optimized square root central difference Kalman filter, comprising the following steps: (A) establishing a second-order equivalent circuit model of the lithium battery; (B) identifying parameters of the second-order equivalent circuit model of the lithium battery using the forgetting factor recursive least squares method; (C) estimating SOC using square root central difference Kalman filter (SRCDKF); (D) using the historical information of SCDKF as training data for the LightGBM algorithm in the low SOC region, then using the LightGBM algorithm to predict new information, and merging the new information into the SCDKF. Compared with the prior art, this invention introduces the LightGBM algorithm into SCDKF to further improve the prediction accuracy, resulting in a SOC assessment method with good convergence even in the low SOC region.

Description

Technical Field

[0001] This invention relates to the field of lithium battery state of charge assessment technology, specifically to a lithium battery state of charge assessment method based on LightGBM optimized square root central difference Kalman filter. Background Technology

[0002] Accurate and efficient monitoring of the state of charge (SOC) of a power battery is fundamental to ensuring its safe, reliable, and efficient operation. Precise SOC estimation can balance differences between individual cells, optimize charging and discharging strategies, and prevent overheating, overcharging, and over-discharging. Currently, the depth of discharge of electric vehicle batteries is typically limited to 90% or 80%, meaning that 10% or 20% of the battery capacity is not effectively utilized. This results in significant energy waste and shortens the battery's driving range. Therefore, achieving accurate state estimation on existing power battery platforms, especially in the low SOC region, and fully utilizing available capacity is of great significance for improving the driving range of electric vehicles and alleviating range anxiety.

[0003] Accurate estimation of the state of charge (SOC) of a power battery depends on the chosen SOC estimation algorithm. Currently, the widely used algorithms are the Extended Kalman Filter (EKF) and the Center Difference Kalman Filter (CDKF). The EKF suffers from truncation errors during linearization, typically achieving only a first-order Taylor expansion, leading to insufficient accuracy. The CDKF, by utilizing the mean and covariance of the sampling points, can theoretically achieve second-order Taylor accuracy, but may suffer from rounding errors that cause computational divergence. While both the existing EKF and CDKF methods can estimate SOC, their accuracy and stability require further improvement. Summary of the Invention

[0004] This invention provides a lithium battery state of charge (SOC) estimation method based on LightGBM optimized square root central difference Kalman filter, aiming to improve the accuracy of battery SOC estimation, especially providing a battery SOC estimation method in the low SOC region.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0006] The lithium battery state-of-charge assessment method based on LightGBM optimized square root central difference Kalman filter includes the following steps:

[0007] (A) Establish a second-order equivalent circuit model for a lithium battery;

[0008] (B) The parameters of the second-order equivalent circuit model of lithium battery are identified by the forgetting factor recursive least squares method.

[0009] (C) SOC estimation is performed using square root central difference Kalman filter;

[0010] (D) In ​​the low SOC region, the historical information of SRCCKF is used as the training data of the LightGBM algorithm. Then, the LightGBM algorithm is used to predict the new information and the new information is merged into SRCCKF, thereby improving the estimation accuracy of SRCCKF.

[0011] Furthermore, the parameters of the second-order equivalent circuit model of the lithium battery described in step (A) include the ohmic internal resistance R0, the polarization resistance R1, the polarization capacitance C1, the concentration polarization resistance R2, and the concentration polarization capacitance C2.

[0012] Furthermore, the state-space model of the lithium battery is as follows:

[0013]

[0014] U(k)=U ocv (k)-R0I(k)-U1(k)-U2(k)+v k

[0015] Where Δt is the sampling time interval, k is the sampling time, I is the operating current, R0 is the battery's internal resistance in ohms, R1 and C1 are the polarization resistor and polarization capacitor, R2 and C2 are the concentration polarization resistor and concentration polarization capacitor, U1 and U2 correspond to the voltages of the two circuits, η is the charge / discharge efficiency, Q is the battery capacity, SOC(k) represents the battery SOC at time k, and w k =(w 1,k ,w 2,k ,w 3,k ) T For system noise, v k To observe the noise, U(k) is the terminal voltage at time k. ocv (k) is the open-circuit voltage at time k.

[0016] Furthermore, in step (B), the forgetting factor recursive least squares method is used to identify the parameters of the second-order equivalent circuit model of the lithium battery. The formula for the forgetting factor recursive least squares method is as follows:

[0017]

[0018] In the formula For data vectors, The predicted result of the observation; y(k) is the actual observation; K(k) is the gain term; λ is the forgetting factor, usually in the range of 0.95≤λ≤1; I is the identity matrix of the corresponding dimension.

[0019] Furthermore, in step (C), the square root central difference Kalman filter algorithm is used for charge state estimation. Compared with the central difference Kalman filter, the square root central difference Kalman filter uses the square root form of the covariance instead of the covariance form for updating. This square root is obtained by QR decomposition.

[0020] Furthermore, the low SOC region mentioned in step (D) is a region with an SOC of less than 20%.

[0021] Furthermore, when SOC < 20%, the historical innovations of the square root central difference Kalman filter are used as training data for the LightGBM algorithm. Then, the LightGBM algorithm is used to predict new innovations, which are then used to replace the innovations in the SRCCKF for state of charge estimation. This includes the following steps:

[0022] Step 1.1: Normalize the historical information and divide it into training data and test data;

[0023] The normalization formula uses min-max standardization, and the formula is as follows:

[0024]

[0025] Where e is the new information and y is the normalized new information, where e max e is the maximum value of the sample. min The minimum value in the sample;

[0026] Step 1.2: Based on the LightGBM algorithm, train the model on the new information data and determine the model parameters;

[0027] Step 1.3: Input the LightGBM model parameters and historical updates, and predict the updates for the next time step;

[0028] Step 1.4: Replace the original information with the predicted information for filtering calculation, and then complete the SOC estimation.

[0029] Furthermore, based on the LightGBM algorithm, model training is performed on the new information data to determine the model parameters, including the following steps:

[0030] Step 1.2.1: Set the initial gradient value of the model gradient to g. i The initial value is zero or a random value;

[0031]

[0032] Where λ is the learning rate, y i Let f(x) be the normalized array, and f(x) be the weak classifier.

[0033] Step 1.2.2: Build a decision tree, knowing the limit on the number of leaves or that all leaves cannot be further divided;

[0034] Step 1.2.3: Calculate the histogram, obtain the splitting gain from the histogram, select the optimal splitting feature G, and the classification threshold I;

[0035]

[0036]

[0037] Among them, W j P represents the optimal split point. j Threshold for the optimal split point, l ij C represents the loss value when the i-th tree splits at point j. ij This represents the feature weight of the i-th tree at point j;

[0038] Step 1.2.4: Establish the root node and segment the samples according to the optimal splitting feature and splitting threshold;

[0039] s = argmax(W j ) 1≤i≤m

[0040] Node = (s, W) s ,P s )

[0041] Where s is the new decision tree formed after splitting the i-th tree at point s, and W s P represents the splitting feature of the decision tree at point s. s Let be the splitting threshold of the decision tree at point s, and Node be the root node of the decision tree at point s.

[0042] Step 1.2.5: Update the gradient g of the decision tree. i ;

[0043] Step 1.2.6: Repeat steps 1.2.2 to 1.2.5 until all decision trees are built.

[0044] Compared with the prior art, the present invention has the following advantages:

[0045] 1. Compared with EKF, the square root central difference Kalman filter algorithm of the present invention does not require the calculation of the Jacobian matrix, and has higher accuracy and stability.

[0046] 2. Compared with CDKF, SRDKF can guarantee the non-negativity of the covariance matrix and improve numerical stability.

[0047] 3. This invention combines SRCDKF and LightGBM to obtain a state of charge assessment method that still has good convergence in the low SOC (SOC < 20%) region. Attached Figure Description

[0048] Figure 1 This is a flowchart of a lithium battery state-of-charge assessment method based on LightGBM optimized square root central difference Kalman filter.

[0049] Figure 2 This is a diagram of a second-order equivalent circuit.

[0050] Figure 3 This is a diagram showing the state of charge estimation of SRCDKF-Light SOC.

[0051] Figure 4 A comparison chart of SOC estimation errors. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to specific embodiments and accompanying drawings. It should be noted that the embodiments described herein are illustrative, but not limiting, and therefore this invention is not limited to the above embodiments. Based on the principles of this invention, all other implementation methods obtained by those skilled in the art without creative effort are considered to be within the protection scope of this invention.

[0053] like Figure 1 As shown, this embodiment provides a method for assessing the state of charge of lithium batteries based on LightGBM optimized square root center difference Kalman filter, including the following steps:

[0054] (A) Establish a second-order equivalent circuit model for a lithium battery;

[0055] (B) The parameters of the second-order equivalent circuit model of lithium battery are identified by the forgetting factor recursive least squares method.

[0056] (C) SOC estimation is performed using square root central difference Kalman filter;

[0057] (D) In ​​the low SOC region, the historical information of SRDKF is used as the training data of the LightGBM algorithm, and then the LightGBM algorithm is used to predict the new information, thereby improving the estimation accuracy of SRDKF.

[0058] Furthermore, the specific process of the lithium battery state-of-charge assessment method based on LightGBM optimized square root central difference Kalman filter is as follows:

[0059] Step 1.1: Establish a second-order equivalent circuit model of the lithium battery, such as... Figure 2 As shown;

[0060] Step 1.2: Use the forgetting factor recursive least squares method to identify the parameters of the second-order equivalent circuit model established in Step 1.1.

[0061] Step 1.3: Based on the model parameters obtained in Step 1.2, SOC estimation is performed using square root central difference Kalman filtering;

[0062] Step 1.4: In the low SOC region (SOC < 20%), use the historical innovations of SRCCKF as training data for the LightGBM algorithm. Then, use the LightGBM algorithm to predict new innovations and incorporate them into SRCCKF, thereby improving the estimation accuracy of SRCCKF. The SOC estimation results are as follows: Figure 3 As shown, the estimation error is as follows: Figure 4 As shown.

[0063] I. Battery Modeling and Parameter Identification

[0064] Using a second-order equivalent circuit model, the state-space model of the battery is as follows:

[0065]

[0066] U(k)=U ocv (k)-R0I(k)-U1(k)-U2(k)+v k (2)

[0067] Where Δt is the sampling time interval, k is the sampling time, I is the operating current, R0 is the battery's internal resistance in ohms, R1 and C1 are the polarization resistor and polarization capacitor, R2 and C2 are the concentration polarization resistor and concentration polarization capacitor, U1 and U2 correspond to the voltages of the two circuits, η is the charge / discharge efficiency, Q is the battery capacity, SOC(k) represents the battery SOC at time k, and w k =(w 1,k ,w 2,k ,w 3,k ) T For system noise, v k To observe the noise, U(k) is the terminal voltage at time k. ocv (k) is the open-circuit voltage at time k.

[0068] The forgetting factor recursive least squares method is used to identify parameters of the equivalent circuit model. The formula for the forgetting factor recursive least squares method is as follows:

[0069]

[0070] In the formula For data vectors, The predicted result of the observation; y(k) is the actual observation; K(k) is the gain term; λ is the forgetting factor, usually in the range of 0.95≤λ≤1; I is the identity matrix of the corresponding dimension.

[0071] 2. SOC estimation is performed using the square root central difference Kalman filter algorithm. EKF approximates the nonlinear system with first-order Taylor accuracy, while SRCDKF can approximate the nonlinear system with at least second-order Taylor accuracy. The SRCDKF algorithm can also enhance the numerical stability of the filter.

[0072] Suppose the state equation and measurement equation of the stochastic discrete system are as follows:

[0073]

[0074] In the formula, f and h are the state function and observation function of the system, respectively, and w k-1 and v k These are system noise and observation noise, respectively.

[0075] (1) Initialization. The initial values ​​of the state variables and the initial error covariance matrix are as follows:

[0076]

[0077] P0 represents process noise, and Q0 represents measurement noise; performing Cholesky decomposition on P0 and Q0, we get:

[0078]

[0079]

[0080] (2) Time Update. Following the Sigma point-symmetric sampling strategy, by... and Calculate the Sigma point ξ i,kk-1 (i = 1, 2, ..., n) can be represented as

[0081]

[0082] The corresponding weighting coefficients are:

[0083]

[0084] Where h is the step size of the central difference, n represents the state dimension, and W0 m W i m , and These are the weights corresponding to the Sigma points;

[0085] xi,k-1 Through the nonlinear state function f k-1 (·) propagates as g i,kk-1 , by g i,kk-1 Further state predictions can be obtained. and the square root of the error covariance matrix

[0086]

[0087]

[0088] in

[0089]

[0090]

[0091]

[0092] (3) Measurement update. Utilizing and The Sigma point x is calculated using a symmetric sampling strategy. i,kk-1 (i = 0, 1, ..., 2n), then x i,kk-1 It can be represented as

[0093]

[0094] x i,kk-1 Through the nonlinear measurement function h k (·) propagation is c i,kk-1 , by c i,kk-1 Output prediction can be obtained and the square root of the autocovariance matrix and mutual covariance matrix

[0095] χ i,kk-1 =h k+1 (ξ i,kk-1 ), i = 0, 1, ..., 2n (17)

[0096]

[0097]

[0098] in

[0099]

[0100]

[0101]

[0102]

[0103] After obtaining new measurement z k Then, the measurement was updated.

[0104]

[0105]

[0106]

[0107] In the formula K k This is the filter gain matrix.

[0108] 3. When SOC < 20%, use the historical innovations of SRCCKF as training data for the LightGBM algorithm, then use the LightGBM algorithm to predict new innovations and replace the innovations in SRCCKF for state of charge estimation, including the following steps:

[0109] Step 1.1: Normalize the historical information and divide it into training data and test data;

[0110] The normalization formula uses min-max standardization, and the formula is as follows:

[0111]

[0112] Where y is the normalized information, max is the maximum value of the sample, and min is the minimum value of the sample.

[0113] Step 1.2: Based on the LightGBM algorithm, train the model on the new information data and determine the model parameters;

[0114] Step 1.2.1: Set the initial gradient value of the model gradient to g. i The initial value is zero or a random value;

[0115]

[0116] Where λ is the learning rate, y i Let f(x) be the normalized array, and f(x) be the weak classifier.

[0117] Step 1.2.2: Build a decision tree, knowing the limit on the number of leaves or that all leaves cannot be further divided;

[0118] Step 1.2.3: Calculate the histogram, obtain the splitting gain from the histogram, select the optimal splitting feature G, and the classification threshold I;

[0119]

[0120]

[0121] Among them, W j P represents the optimal split point. j Threshold for the optimal split point, l ij C represents the loss value when the i-th tree splits at point j. ij This represents the feature weight of the i-th tree at point j;

[0122] Step 1.2.4: Establish the root node and segment the samples according to the optimal splitting feature and splitting threshold;

[0123] s = argmax(W j ) 1≤i≤m (31)

[0124] Node = (s, W) s ,P s (32)

[0125] Where s is the new decision tree formed after splitting the i-th tree at point s, and W s P represents the splitting feature of the decision tree at point s. s Let be the splitting threshold of the decision tree at point s, and Node be the root node of the decision tree at point s.

[0126] Step 1.2.5: Update the gradient g of the decision tree. i ;

[0127] Step 1.2.6: Repeat steps 1.2.2 to 1.2.5 until all decision trees are built;

[0128] Step 1.3: Input the LightGBM model parameters and historical updates, and predict the updates for the next time step;

[0129] Step 1.4: Replace the original information with the predicted information for filtering calculation, and then complete the SOC estimation.

[0130] Table 1 shows the statistical indicators of SOC estimation by the algorithm at FUDS 25°. All are based on the second-order equivalent circuit model and FFRLS parameter identification. The maximum errors of EKF, CDKF, SRDKF, and SRDKF-LightGBM are 11.26%, 10.48%, 9.76%, and 2.55%, respectively; the mean absolute errors are 1.32%, 1.16%, 1.06%, and 0.39%, respectively; and the root mean square errors are 3.04%, 2.57%, 2.39%, and 0.64%, respectively. The SRCCKF-LightGBM algorithm shows improvements across all three metrics. CDKF achieves some accuracy improvements over EKF, and SRCCKF also shows some improvement over CDKF. The SRCCKF-LightGBM algorithm demonstrates significant improvements: in maximum error, SRCCKF-LightGBM improves by 73.8% compared to SRCCKF, 75.6% compared to CDKF, and 77.3% compared to EKF; in mean absolute error, SRCCKF-LightGBM improves by 63.2% compared to SRCCKF, 66.3% compared to CDKF, and 70.4% compared to EKF; and in root mean square error, SRCCKF-LightGBM improves by 73.2% compared to SRCCKF, 75% compared to CDKF, and 78.9% compared to EKF. These results demonstrate the superior tracking performance of the algorithm presented in this invention, highlighting the significant advancements in the technology.

[0131] Table 1

[0132] algorithm MAX MAE RMSE EKF 11.26％ 1.32％ 3.04％ CDKF 10.48％ 1.16％ 2.57％ SRCDKF 9.76％ 1.06％ 2.39％ SRCDKF-LightGBM 2.55％ 0.39％ 0.64％

[0133] Although exemplary embodiments of the invention have been described and illustrated, those skilled in the art will understand that various changes and substitutions can be made thereto without departing from the spirit of the invention. Furthermore, many modifications can be made to adapt specific situations to the doctrine of the invention without departing from the central concepts of the invention described herein. Therefore, the invention is not limited to the specific embodiments disclosed herein, but may include all embodiments and equivalents that fall within the scope of the invention.

Claims

1. A lithium battery state of charge estimation method based on LightGBM optimized square root central difference Kalman filter, characterized in that, Includes the following steps: (A) Establish a second-order equivalent circuit model for a lithium battery; (B) The parameters of the second-order equivalent circuit model of lithium battery are identified by the forgetting factor recursive least squares method. (C) SOC estimation is performed using square root central difference Kalman filtering; (D) In ​​the low SOC region, the historical information of SRCCKF is used as the training data of the LightGBM algorithm. Then, the LightGBM algorithm is used to predict the new information and the new information is merged into SRCCKF, thereby improving the estimation accuracy of SRCCKF. The low SOC region mentioned in step (D) is a region with an SOC of less than 20%; When SOC < 20%, the historical innovations of the square root central difference Kalman filter are used as training data for the LightGBM algorithm. Then, the LightGBM algorithm is used to predict new innovations and replace the innovations in the SRCDKF for state of charge estimation. This includes the following steps: Step 1.1: Normalize the historical information and divide it into training data and test data; The normalization formula uses min-max standardization, and the formula is as follows: ； Where e is the new information and y is the normalized new information, The maximum value of the sample. The minimum value in the sample; Step 1.2: Based on the LightGBM algorithm, train the model on the new information data and determine the model parameters; Step 1.3: Input the LightGBM model parameters and historical updates, and predict the updates for the next time step; Step 1.4: Replace the original information with the predicted information for filtering calculation, and then complete the SOC estimation; Based on the LightGBM algorithm, model training is performed on the new information data, and model parameters are determined, including the following steps: Step 1.2.1: Set the initial gradient value of the model gradient to... The initial value is zero or a random value; ； in, For learning rate, The normalized array, It is a weak classifier; Step 1.2.2: Build a decision tree, knowing the limit on the number of leaves or that all leaves cannot be further divided; Step 1.2.3: Calculate the histogram, obtain the splitting gain from the histogram, select the optimal splitting feature G, and the classification threshold I; ； in, Represents the optimal split point. Threshold for the optimal split point This represents the loss value when the i-th tree splits at point j. This represents the feature weight of the i-th tree at point j; Step 1.2.4: Establish the root node and segment the samples according to the optimal splitting feature and splitting threshold; ； Where s is the new decision tree formed after splitting the i-th tree at point s. The splitting feature of the decision tree at point s, Let be the splitting threshold of the decision tree at point s, and Node be the root node of the decision tree at point s. Step 1.2.5: Update the gradient of the decision tree. ; Step 1.2.6: Repeat steps 1.2.2 to 1.2.5 until all decision trees are built.

2. The lithium battery state-of-charge assessment method based on LightGBM optimized square root central difference Kalman filter according to claim 1, characterized in that, The parameters of the second-order equivalent circuit model of the lithium battery described in step (A) include the ohmic internal resistance R0, the polarization resistance R1, the polarization capacitance C1, the concentration polarization resistance R2, and the concentration polarization capacitance C2.

3. The lithium battery state-of-charge assessment method based on LightGBM optimized square root central difference Kalman filter according to claim 2, characterized in that, The state-space model of a lithium battery is as follows: ； ； in Let k be the sampling time interval, I be the operating current, R0 be the battery's internal resistance in ohms, R1 and C1 be the polarization resistor and polarization capacitor, R2 and C2 be the concentration polarization resistor and concentration polarization capacitor, U1 and U2 be the voltages of the two circuits, η be the charge / discharge efficiency, Q be the battery capacity, and SOC(k) represent the battery SOC at time k. For system noise, To observe the noise, Let be the terminal voltage at time k. Let be the open-circuit voltage at time k.

4. The lithium battery state-of-charge assessment method based on LightGBM optimized square root central difference Kalman filter according to claim 1, characterized in that, In step (B), the forgetting factor recursive least squares method is used to identify the parameters of the second-order equivalent circuit model of the lithium battery. The formula for the forgetting factor recursive least squares method is as follows: ； In the formula For data vectors, Predicted results of the observed values; These are actual observed values; λ is the gain term; λ is the forgetting factor, which is usually in the range of 0.95≤λ≤1. is the identity matrix of the corresponding dimension.

5. The lithium battery state-of-charge assessment method based on LightGBM optimized square root central difference Kalman filter according to claim 1, characterized in that, In step (C), the square root central difference Kalman filter algorithm is used for charge state estimation. Compared with the central difference Kalman filter, the square root central difference Kalman filter uses the square root form of the covariance instead of the covariance form for updating. This square root is obtained by QR decomposition.

Images