Improved single particle electrochemical modeling method based on gradient boosting decision tree

Abstract

The present invention discloses an improved single-particle electrochemical modeling method based on a gradient enhanced decision tree, and the improved single-particle electrochemical modeling method based on a gradient enhanced decision tree includes the following steps: S1, collecting open circuit voltage data and preprocessing; S2, establishing SPM and obtaining lithium ion concentration data on the positive and negative electrode surfaces; S3, establishing a model based on GBDT; S4, obtaining model input and output data to verify the model accuracy. The problem that the existing battery model cannot accurately track the state of charge of the battery and estimate its life is solved. In the present invention, the open circuit voltage is collected, SPM is established to obtain the lithium ion concentration on the positive and negative electrode surfaces, and a model based on GBDT is established, and finally the accuracy verification is performed. The method of the present invention greatly improves the accuracy of the model, and the calculation is relatively simple and efficient.

Description

Technical Field

[0001] The invention relates to the field of single particle electrochemical models, in particular to an improved single particle electrochemical modeling method based on a gradient enhanced decision tree. Background Art

[0002] In a large energy storage system, a single battery failure may pose a danger to the entire system. Therefore, a reliable battery management system (BMS) is needed to control the battery charging and discharging process to ensure the safe operation of the system. In addition to battery balancing and thermal management, the BMS also needs to accurately track the battery's state of charge and estimate its life, which requires an accurate battery model.

[0003] The commonly used equivalent models for lithium-ion batteries are equivalent circuit model (ECM), empirical model and electrochemical model. ECM approximates the structure of the battery through several resistor-capacitor networks. Due to its simple structure and low computational complexity, it is widely used in online SOC and SOH estimation and control applications. However, the disadvantage is that in order to fit the dynamic electrochemical process and aging effects, ECM needs to be extended to higher-order circuits. At the same time, many parameters in the model need to be determined by table lookup based on SOC, temperature, current amplitude and current direction. The empirical model describes its battery behavior by using various functions and formulas. Since these models are fitted with past experimental data, they have poor prediction performance under unknown working conditions and low universality.

[0004] The "Battery Life Prediction Method, Device, Cloud Server and Storage Medium" disclosed in the Chinese patent literature, with the publication number CN114330149A, discloses a battery life prediction method, device, cloud server and storage medium, wherein the method includes: determining an empirical model, and using the empirical model to learn the battery historical data trajectory; estimating the battery capacity state based on the battery historical data to obtain a first estimation result, and in the process of learning the battery historical data trajectory, correcting the parameters of the empirical model according to the first estimation result; determining a mechanism model, and estimating the battery capacity state according to the mechanism model to obtain a second estimation result; using the corrected empirical model to predict the battery life, and in the prediction process, correcting the corrected empirical model parameters again according to the second estimation result, and finally obtaining the battery life prediction result. However, the Chinese patent with the publication number CN114330149A has poor prediction performance and low universality. Summary of the invention

[0005] The present invention solves the problem that the existing battery model cannot accurately track the state of charge of the battery and estimate its life. An improved single-particle electrochemical model modeling method based on a gradient enhanced decision tree is proposed. The open circuit voltage is collected, and an SPM is established to obtain the lithium ion concentration on the surface of the positive and negative electrodes. A model based on GBDT is established, and finally an accuracy check is performed. The method of the present invention greatly improves the accuracy of the model, and the calculation is relatively simple and efficient.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme: an improved single particle electrochemical modeling method based on gradient enhanced decision tree, comprising the following steps:

[0007] S1, collect open circuit voltage data and perform preprocessing;

[0008] S2, establish SPM and obtain lithium ion concentration data on the positive and negative electrode surfaces;

[0009] S3, building a GBDT-based model;

[0010] S4, obtaining model input and output data to verify the accuracy of the model.

[0011] In the present invention, first, multiple groups of open circuit voltage data are collected and preprocessed. After the preprocessing is completed, the lithium ion concentration data on the positive and negative electrode surfaces are obtained, specifically through the established SPM, according to the relationship between the open circuit voltage data and the lithium ion concentration data on the positive and negative electrode surfaces; after determining the open circuit voltage data and the lithium ion concentration data on the positive and negative electrode surfaces, the two are used as input and output to create a model based on GBDT, and finally the model data is checked for accuracy to exclude abnormal data; the model of the present invention has high accuracy. The present invention establishes the relationship between lithium ion concentration and open circuit voltage through a gradient enhanced decision tree, thereby achieving more accurate battery terminal voltage fitting.

[0012] Preferably, the step S1 comprises the following steps:

[0013] S11, select multiple temperature conditions, charge and discharge the battery to the set SOC respectively, and collect open circuit voltage data after stabilization;

[0014] S12, the open circuit voltage data should be within the set threshold range, the abnormal data should be removed, and the open circuit voltage data under the same temperature conditions of the historical day should be added. In the present invention, for the conditions of collection, the temperature conditions can be controlled and set in a closed environment. After adjusting to the preset SOC, the collection work can be carried out after it stabilizes; in addition, the open circuit voltage data outside the set threshold range is eliminated.

[0015] Preferably, step S2 comprises the following steps:

[0016] S21, SPM is based on the P2D model, with simplified reaction process, approximated calculation formula and reduced parameters. The formula is as follows:

[0017] V=E-η act -η con -η ohm

[0018] in, V represents the battery open circuit voltage, E represents the battery open circuit potential, ηact represents the reaction polarization overpotential, ηcon represents the concentration polarization overpotential, ηohm represents the ohmic polarization overpotential;

[0019] S22, the open circuit potential is expressed as:

[0020]

[0021] in, Respectively represent the maximum solid phase lithium insertion concentration of the positive and negative electrode materials, Indicates the current surface lithium insertion concentration of the positive and negative electrode materials. Specifically, p represents the positive electrode and n represents the negative electrode;

[0022] S23, substitute the collected open circuit voltage data into the formula of steps S21 and S22 to obtain the lithium ion concentration data on the positive and negative electrode surfaces. In the present invention, the battery terminal voltage is summarized as the joint effect of four processes: open circuit voltage, concentration polarization overpotential, ohmic polarization overpotential and reaction polarization overpotential; after the SPM is established, the lithium ion concentration data on the positive and negative electrode surfaces are obtained based on the open circuit voltage data.

[0023] Preferably, the GBDT-based model has the following formula, specifically:

[0024]

[0025] In the above formula, x represents the input sample, ω represents the model parameter, h represents the decision tree, and α represents the weight of each tree. Given a training set T = {(x1, y1), (x2, y2), ..., (x N ,y N )}, the input space X satisfies The output space Y satisfies In the present invention, compared with traditional statistical models, the GBDT model can effectively quantify and capture the influence of independent variables of different units or orders of magnitude on dependent variables; at the same time, as an integrated model of decision trees, GBDT can find various eigenvalues ​​or eigenvalue combinations, adjust the weights of independent variables by learning from data, and thus improve the estimation accuracy of the model.

[0026] Preferably, the optimization iterative process of the GBDT-based model comprises the following steps:

[0027] S31, initialize the first decision tree model, as follows:

[0028]

[0029] Among them, L(y i ,c) represents the loss function, c represents the constant value that is estimated to minimize the loss function;

[0030] S32, create M decision trees, m = 1, 2, ..., M, and obtain the negative gradient r of each sample im And the new decision tree, then update the model, the update model is as follows:

[0031]

[0032] Where I represents the parameter controlling the negative gradient;

[0033] S33, the final output of the strong learning machine is as follows:

[0034]

[0035] S34, training the GBDT-based model; in the present invention, after the model is established, the lithium ion concentration on the surface of the positive and negative electrodes is used as input, and the open circuit voltage is used as output, which are input into the GBDT-based model.

[0036] Preferably, the step S32 comprises the following steps:

[0037] S321, find the negative gradient r of each sample im , as follows:

[0038]

[0039] S322, creating a new decision tree. In the present invention, the negative gradient of the sample is first determined, and then a new decision tree is established.

[0040] Preferably, the step S322 includes: taking the obtained residual as the new value of the sample, and converting the data (x i ,r im ) is used as the training data for the next tree to obtain a new regression decision tree f m (X), the corresponding leaf node area is represented by R jm , where j = 1, 2, ..., J m , J m Represents the number of leaf nodes of the mth regression tree, and its leaf area c jm The best fit values ​​of are as follows: In the present invention, the above-mentioned means are used to obtain the final best fitting value of the leaf area.

[0041] Preferably, the step S4 specifically extracts any GBDT-based model input and output data, i.e., the positive and negative electrode surface lithium ion concentration data and the open circuit voltage data, and compares them with the positive and negative electrode surface lithium ion concentration data and the open circuit voltage data under the same temperature conditions at a future time. The positive and negative electrode surface lithium ion concentration data at a future time are detected using a special lithium ion detection device. If the comparison error is greater than 10%, the model input and output data need to be manually calculated to determine the error step and record it. In the present invention, after the model is created, the data at the same temperature at the future time is detected. The future time here does not need to be predicted, but the time is a certain period after the model input and output data are extracted. In the present invention, the future time is generally the data of a certain period of the second day after the model input and output data are extracted, which is directly detected by the lithium ion detection equipment; for the comparison error, specifically the lithium ion concentration data of the positive and negative electrode surfaces of the model are compared with the lithium ion concentration data of the positive and negative electrode surfaces under the same temperature conditions at the future time, and the open circuit voltage data of the model are compared with the open circuit voltage data under the same temperature conditions at the future time. If the comparison results of the two are both greater than 10%, it is defined as the total comparison error is greater than 10%.

[0042] The beneficial effects of the present invention are as follows: the improved single-particle electrochemical modeling method based on gradient enhanced decision tree of the present invention collects open circuit voltage, establishes SPM for obtaining lithium ion concentration on the surface of positive and negative electrodes, and establishes a model based on GBDT, and finally performs accuracy verification. The method of the present invention greatly improves the accuracy of the model, and the calculation is relatively simple and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the improved single particle electrochemical modeling method based on gradient enhanced decision tree of the present invention;

[0044] Figure 2 It is a flow chart of the improved single particle electrochemical model modeling method based on gradient enhanced decision tree of the present invention based on the GBDT model. DETAILED DESCRIPTION

[0045] Example:

[0046] This embodiment proposes an improved single particle electrochemical modeling method based on a gradient boosting decision tree, referring to Figure 1, specifically including the following multiple steps, step S1, first collect open circuit voltage data, and then pre-process, specifically, this step also includes the following two sub-steps, step S11, select several temperature conditions, charge and discharge the battery to a preset SOC, and start collecting open circuit voltage data after stabilization. In this embodiment, for the conditions of collection, the temperature conditions can be controlled and set in a closed environment, and after adjusting to the preset SOC, it is necessary to wait until it is stable before collecting.

[0047] Then, step S12 is performed, the open circuit voltage data is set within the threshold range, the abnormal data is removed, and the open circuit voltage data under the same temperature conditions of the historical day is added. In this step, the open circuit voltage data is automatically judged by the threshold judgment unit, and the open circuit voltage data outside the set threshold range is automatically eliminated. If the eliminated position needs to be filled, the data under the same temperature conditions of the historical day can be selected to replace it.

[0048] Step S2, establishing SPM to obtain lithium ion concentration data on the positive and negative electrode surfaces; specifically, including the following steps, step S21, wherein the battery open circuit voltage is specifically:

[0049] V=E-η act -η con -η ohm

[0050] In the above formula, V is the battery open circuit voltage, E is the battery open circuit potential, ηact is the reaction polarization overpotential, ηcon is the concentration polarization overpotential, ηohm is the ohmic polarization overpotential; SPM is based on the P2D model, with simplified reaction process, approximated calculation formula and reduced parameters; among them, the open circuit voltage is the most important part of the terminal voltage, and its size is determined by the surface lithium ion concentration of the corresponding positive and negative electrode active materials.

[0051] Step S22, the formula of open circuit potential is as follows:

[0052]

[0053] In the above formula, are the maximum solid phase lithium insertion concentrations of the positive and negative electrode materials, is the current surface lithium insertion concentration of the positive and negative electrode materials, where p represents the positive electrode and n represents the negative electrode; however, the above formula is often a fitting result obtained through an empirical formula.

[0054] Step S23, according to the collected open circuit voltage data and the formulas of steps S21 and S22, the surface lithium ion concentration data of the positive and negative electrodes are calculated. In this embodiment, the battery terminal voltage is summarized as the joint effect of four processes: open circuit voltage, concentration polarization overpotential, ohmic polarization overpotential and reaction polarization overpotential; after the SPM is established, the surface lithium ion concentration data of the positive and negative electrodes are calculated based on the open circuit voltage data.

[0055] Step S3, build a model based on GBDT, refer to Figure 2 Specifically, the GBDT-based model refers to the following formula: Among them, x is the input sample, ω is the model parameter, h is the decision tree, α is the weight of each tree, given a training set T = {(x1, y1), (x2, y2), ..., (x N ,y N )}, the input space X satisfies The output space Y satisfies In the present invention, GBDT is an iterative decision tree algorithm, which consists of multiple weak learners (decision trees) and is trained in the direction of reducing residuals using a gradient enhancement algorithm.

[0056] Step S3 includes the following sub-steps: Step S31, initializing the first decision tree model, referring to the following formula:

[0057]

[0058] Among them, L(y i ,c) represents the loss function, and c represents the constant value estimated to minimize the loss function.

[0059] Step S32: Specifically, M decision trees need to be established. When m = 1, 2, ..., M, the negative gradient r of each sample is first calculated. im

[0060]

[0061] Then create a new decision tree, specifically: take the obtained residual as the new value of the sample, and replace the data (x i ,r im ) as the training data for the next tree to obtain a new regression decision tree f m (X), the corresponding leaf node area is R jm , with j = 1, 2, ..., J m , J m is the number of leaf nodes of the mth regression tree, so the leaf area c jm The best fitting value of is:

[0062]

[0063] Finally, update the model and get the following formula:

[0064]

[0065] In the above formula, I is the parameter that controls the negative gradient.

[0066] Step S33, finally output the strong learning machine, the specific formula is:

[0067] The mapping function established by GBDT can be obtained from the above formula. Compared with traditional statistical models, GBDT can effectively quantify and capture the influence of independent variables of different units or orders of magnitude on dependent variables; at the same time, as an integrated model of decision trees, GBDT can find various eigenvalues ​​or eigenvalue combinations, adjust the weights of independent variables by learning from data, and thus improve the estimation accuracy of the model.

[0068] Step S34, training a model based on GBDT; obtaining a final model.

[0069] Step S4, obtaining model input and output data to verify the accuracy of the model; in detail, extracting any GBDT-based model input and output data, i.e., the positive and negative electrode surface lithium ion concentration data and open circuit voltage data, and comparing them with the positive and negative electrode surface lithium ion concentration data and open circuit voltage data under the same temperature conditions at a future time. The positive and negative electrode surface lithium ion concentration data at a future time are detected using a special lithium ion detection device. If the comparison error is greater than 10%, the model input and output data need to be manually calculated, and the error step is determined and recorded.

[0070] In this embodiment, after the model is created, the data at the same temperature in the future is detected. The future here does not need to be predicted, but the time is a period after the model input and output data are extracted. In the present invention, the future is generally the data of a period of time on the second day after the model input and output data are extracted, which is directly detected by the lithium ion detection equipment; for the comparison error, specifically the lithium ion concentration data of the positive and negative electrode surfaces of the model are compared with the lithium ion concentration data of the positive and negative electrode surfaces under the same temperature conditions in the future, and the open circuit voltage data of the model is compared with the open circuit voltage data under the same temperature conditions in the future. If the comparison results of the two are both greater than 10%, it is defined as the total comparison error is greater than 10%.

[0071] In the present invention, SPM is the abbreviation of single particle model, and GBDT is the abbreviation of gradient boosting decision tree. The improved single particle electrochemical model modeling method based on gradient boosting decision tree of the present invention, first, collects and obtains multiple groups of open circuit voltage data and performs preprocessing. After the preprocessing is completed, the lithium ion concentration data on the surface of the positive and negative electrodes is obtained, specifically through the established SPM, according to the relationship between the open circuit voltage data and the lithium ion concentration data on the surface of the positive and negative electrodes; after determining the open circuit voltage data and the lithium ion concentration data on the surface of the positive and negative electrodes, the two are used as input and output to create a model based on GBDT, and finally the model data is checked for accuracy to exclude abnormal data; the model accuracy of the present invention is high. The present invention establishes the relationship between its lithium ion concentration and open circuit voltage through a gradient boosting decision tree, thereby achieving more accurate battery terminal voltage fitting. The improved single particle electrochemical model modeling method based on gradient boosting decision tree of the present invention collects open circuit voltage, establishes SPM for obtaining lithium ion concentration on the surface of the positive and negative electrodes, and establishes a model based on GBDT, and finally performs accuracy verification. The method of the present invention greatly improves the accuracy of the model, and the calculation is relatively simple and efficient.

[0072] The above embodiments are further elaborations and illustrations of the present invention for ease of understanding, and are not limitations of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An improved single-particle electrochemical modeling method based on gradient boosting decision tree, characterized in that: The following steps are involved: S1, collect open circuit voltage data and perform preprocessing; S2, establishing SPM and obtaining lithium ion concentration data on the positive and negative electrode surfaces; obtaining data based on the relationship between open circuit voltage data and lithium ion concentration data on the positive and negative electrode surfaces through the established SPM; S3, building a GBDT-based model; S4, obtain the model input and output data to verify the model accuracy; extract any GBDT-based model input and output data, i.e., the lithium ion concentration data and open circuit voltage data on the positive and negative electrode surfaces, and compare them with the lithium ion concentration data and open circuit voltage data on the positive and negative electrode surfaces under the same temperature conditions at a future time. If the comparison error is greater than 10%, the model input and output data must be manually calculated to determine the error step and record it.

2. The improved single particle electrochemical modeling method based on gradient boosting decision tree according to claim 1 is characterized in that: The step S1 comprises the following steps: S11, select multiple temperature conditions, charge and discharge the battery to the set SOC respectively, and collect open circuit voltage data after stabilization; S12, the open circuit voltage data should be within the set threshold range, the abnormal data should be removed, and the open circuit voltage data under the same temperature conditions on the historical day should be added.

3. The improved single particle electrochemical modeling method based on gradient boosting decision tree according to claim 1 or 2, characterized in that: Described step S2 comprises the following steps: S21, SPM is based on the P2D model, with simplified reaction process, approximated calculation formula and reduced parameters. The formula is as follows: V=E-η act -or con -or ohm Where V represents the open circuit voltage of the battery, E represents the open circuit potential of the battery, and η act represents the reaction polarization overpotential, η con represents the concentration polarization overpotential, η ohm represents the ohmic polarization overpotential; S22, the open circuit potential is expressed as: in, Respectively represent the maximum solid phase lithium insertion concentration of the positive and negative electrode materials, Indicates the current surface lithium insertion concentration of the positive and negative electrode materials. Specifically, p represents the positive electrode and n represents the negative electrode; S23, substituting the collected open circuit voltage data into the formulas of steps S21 and S22 to obtain the lithium ion concentration data on the positive and negative electrode surfaces.

4. The improved single particle electrochemical modeling method based on gradient boosting decision tree according to claim 3 is characterized in that: The GBDT-based model has the following formula, specifically: In the above formula, x represents the input sample, ω represents the model parameter, h represents the decision tree, and α represents the weight of each tree. Given a training set T = {(x1, y1), (x2, y2), ..., (x N ,y N )}, the input space X satisfies The output space Y satisfies 5. The improved single particle electrochemical modeling method based on gradient boosting decision tree according to claim 4 is characterized in that: The optimization iterative process of the GBDT-based model includes the following steps: S31, initialize the first decision tree model, as follows: Among them, L(y i ,c) represents the loss function, c represents the constant value that is estimated to minimize the loss function; S32, create M decision trees, m = 1, 2, ..., M, and obtain the negative gradient r of each sample im And the new decision tree, then update the model, the update model is as follows: Where I represents the parameter controlling the negative gradient; S33, the final output of the strong learning machine is as follows: S34, train the GBDT-based model.

6. The improved single particle electrochemical modeling method based on gradient boosting decision tree according to claim 5, characterized in that: The step S32 comprises the following steps: S321, find the negative gradient r of each sample im , as follows: S322, create a new decision tree.

7. The improved single particle electrochemical modeling method based on gradient boosting decision tree according to claim 6 is characterized in that: The step S322 includes: taking the obtained residual as the new value of the sample, and converting the data (x i ,r im ) is used as the training data for the next tree to obtain a new regression decision tree f m (X), the corresponding leaf node area is represented by R jm , where j = 1, 2, ..., J m , J m Represents the number of leaf nodes of the mth regression tree, and its leaf area c jm The best fit values ​​of are as follows:

8. The improved single particle electrochemical modeling method based on gradient boosting decision tree according to claim 1, characterized in that: In step S4, the lithium ion concentration data of the positive and negative electrode surfaces at a future time is detected using a dedicated lithium ion detection device.

Images