Capacitor performance prediction method based on machine learning

Through the capacitor performance prediction method based on machine learning, the problem of inaccurate performance prediction of biochar supercapacitor materials is solved, the prediction accuracy is improved, and the preparation of high-performance materials is promoted.

CN118981630AActive Publication Date: 2024-11-19JIANGXI MECHANICAL & ELECTRICAL VOCATIONAL & TECH COLLEGE
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411455783.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-18
Publication Date
2024-11-19
Estimated Expiration
2044-10-18

AI Technical Summary

Technical Problem

When designing and screening high-performance supercapacitor materials, it is difficult for the prior art to accurately capture the impact of the physical and chemical properties of the material on capacitance performance, resulting in inaccurate material performance prediction.

Method used

Using a capacitor performance prediction method based on machine learning, the target data of the biochar supercapacitor is collected, preprocessed and feature extraction is performed, the performance prediction model is established, the hyperparameters is optimized, and the optimal model is determined through Bayesian optimization and random search algorithms. and feature selection, thereby determining the standard range of value of target features.

Benefits of technology

Improve the prediction accuracy of capacitor performance, reveal the mechanism between material structure and capacitor performance, and promote the preparation and application of high-performance biochar supercapacitors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981630A_ABST
    Figure CN118981630A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of capacitor performance prediction, and particularly discloses a capacitor performance prediction method based on machine learning, and the method comprises the following steps: S1, carrying out the preprocessing of a target factor, obtaining a target feature, and determining a training set and a prediction set based on the target feature; s2, establishing an initial model based on a machine learning algorithm, determining an evaluation index of the initial model, and calculating a judgment value; optimizing hyper-parameters of the initial model, determining a judgment value again, generating coordinate points and determining a fitting curve; calculating a judgment value difference value, obtaining a maximum judgment value difference value, and calculating a selection value; s3, determining a target model, determining an optimal hyper-parameter and a corresponding selection value of the target model, and determining a precise model according to the selection value at the moment; and determining a value standard range of the target features based on the precise model and the prediction set. The mechanism between the material structure and the capacitor performance is disclosed through machine learning, and preparation of the high-performance biochar supercapacitor is promoted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of capacitor performance prediction, and in particular to a capacitor performance prediction method based on machine learning. Background Art

[0002] Integrated supercapacitors to prepare self-powered systems have been widely used in various portable microelectronic devices and sensing fields. In this case, how to design and screen high-performance supercapacitor materials is one of the current and future challenges.

[0003] Biochar is a carbon-rich material produced by pyrolysis of organic raw materials under oxygen-limited conditions. Modified biochar after special treatment has a high specific surface area and excellent conductivity, and is widely used as supercapacitor electrode material.

[0004] The capacitance performance of biochar supercapacitors is usually determined by the physical and chemical properties of the material, such as the pore structure of the material, and the detection conditions. From a synthetic perspective, the physical and chemical parameters of the material can be controlled by controlling the synthesis parameters. Therefore, accurately capturing the impact of the physical and chemical properties of the material on the capacitance performance is very helpful for the design and screening of high-performance biochar material supercapacitors.

[0005] As an emerging new material development strategy, machine learning can effectively explore the relationship between material structure and performance, accelerate the chemical synthesis of high-performance device materials, and develop a capacitor performance prediction method based on machine learning, which will help reveal the hidden mechanism between material structure and capacitor performance and promote the preparation of high-performance biochar supercapacitors due to its powerful ability to explore complex potential relationships. Summary of the invention

[0006] The purpose of the present invention is to provide a capacitor performance prediction method based on machine learning to solve the above technical problems.

[0007] The purpose of the present invention can be achieved through the following technical solutions: A capacitor performance prediction method based on machine learning comprises the following steps: S1: collecting target data of biochar supercapacitors, wherein the target data includes target factors and specific capacitance, and the target factors are factors that affect specific capacitance, preprocessing the target factors to obtain target features, and determining a training set and a prediction set based on the target features; S2: Establish a supercapacitor performance prediction model based on a machine learning algorithm, use it as an initial model, train the initial model based on a cross-validation method and the training set, and determine the evaluation index of the initial model after training, the evaluation index includes the determination coefficient ZB fir , root mean square error ZB secAnd the mean absolute error ZB thi , calculate the judgment value P = (ε-ZB fir )*(ZB sec +ZB thi ), ε is the preset correction constant and ε>1; The hyperparameters of the initial model are optimized based on a random search algorithm, and the judgment value after optimizing the hyperparameters is determined to generate coordinate points (a, P a ), P a represents the judgment value after the a-th optimization of the hyperparameter, and the coordinate points are fitted to obtain a fitting curve f(P); Calculate the judgment value difference PCZ a =|P a -P a-1 |, get the maximum judgment value difference PCZ max , calculate the selected value , where η is a preset correction coefficient, and N represents the total number of times the hyperparameter is optimized; S3: Sort the selected values ​​by size. The larger the selected value, the lower the ranking. The initial model corresponding to the top 30% of the selected values ​​in the sorting is used as the target model. The optimal hyperparameters of the target model are determined based on Bayesian optimization. The selected value of the target model at this time is determined, and the target model corresponding to the minimum selected value is used as the accurate model. The standard value range of the target feature is determined based on the precise model and the prediction set. When the value of the target feature falls within the standard value range, the specific capacitance of the biochar supercapacitor is the highest.

[0008] As a further solution of the present invention: in the step S3, the process of determining the standard range of the value of the target feature based on the accurate model and the prediction set specifically includes: Step 1: Determine the evaluation index of the accurate model based on the cross-validation method and the prediction set, determine the judgment value Pcur at this time, set the judgment value threshold Pzq, and when the judgment value Pcur>Pzq, execute step 2; Step 2: Determine the average absolute contribution rate of the target feature based on the precise model, and sort the target features according to the average absolute contribution rate. The larger the average absolute contribution rate, the lower the corresponding target feature is ranked. Step 3: removing the last target feature in the ranking from the prediction set, and repeating steps 1 to 2 until the corresponding judgment value is less than the judgment value threshold after a target feature is removed from the prediction set; Determine the average absolute contribution rate of the target features that have not been removed at this time, and sort the target features that have not been removed at this time according to the size of the average absolute contribution rate. The larger the average absolute contribution rate, the higher the corresponding target feature is ranked. The first C target features in the sorting are taken as the main features, and C is a preset number.

[0009] Step 4: Determine the standard range of values ​​of the main features based on the partial dependence diagram, the individual conditional expectation diagram and the precise model.

[0010] As a further solution of the present invention: in the step three, when the number of target features that have not been removed is C and the judgment value of the precise model is still greater than the judgment value threshold, a warning message is sent to the preset management personnel.

[0011] As a further solution of the present invention: the step S3 further includes the following steps: Determining the SHAP value of the main feature; When the SHAP value of the main feature is positive, the main feature is positively correlated with the specific capacitance of the biochar supercapacitor; When the SHAP value of the main feature is negative, the main feature is negatively correlated with the specific capacitance of the biochar supercapacitor.

[0012] As a further solution of the present invention: in step S1, the process of preprocessing the target factor to obtain the target feature specifically includes: Standardize the numerical target factors to eliminate the dimension; Encode non-numeric target factors through one-hot encoding; The outliers in the target factor are removed by the isolation forest algorithm.

[0013] As a further solution of the present invention: in step S1, the process of determining the training set and the prediction set based on the target feature specifically includes: Obtain the specific capacitance D corresponding to different target features, establish a target mapping between the target feature and the specific capacitance D, and establish a sample set YJH= (Y1, Y2, ..., Y m ), Y m represents the mth target mapping, and divides the sample set into a training set and a prediction set according to a preset ratio.

[0014] As a further solution of the present invention: in the step S2, in the process of training the initial model based on the cross-validation method and the training set and determining the evaluation index of the trained initial model, a five-fold cross-validation is adopted.

[0015] As a further solution of the present invention: in the step S1, in the process of establishing the supercapacitor performance prediction model based on the machine learning algorithm, the machine learning algorithm includes random forest, decision tree, artificial neural network, and support vector machine.

[0016] Beneficial effects of the present invention: In the present invention, firstly, target data is collected, and target factors in the target data are preprocessed to obtain target features, and a training set and a prediction set are obtained according to the target features; it is understandable that the target data can be obtained from published literature, and the target factors include but are not limited to the porous structure characteristics and chemical composition characteristics of biochar supercapacitor electrode materials, the porous structure characteristics include but are not limited to specific surface area, micropore specific surface area and pore size, and the chemical composition characteristics include but are not limited to carbon content and nitrogen content, which will not be repeated here; then an initial model is established based on machine learning, the evaluation index of the initial model is determined, and the judgment value is calculated; it is worth noting that the judgment value is used to comprehensively measure the performance of the initial model, and the judgment value The smaller it is, the better the performance of the model. Specifically, the closer the determination coefficient is to 1, the better the performance of the initial model. The smaller the root mean square error and the mean absolute error are, the better the performance of the initial model is. After that, the hyperparameters of the initial model are optimized based on the random search algorithm. The judgment value is re-determined after each optimization, and the target image is drawn. It can be understood that when the hyperparameters are not optimized, the judgment value of the initial model is P0, and the corresponding coordinate point is (0, P0). Considering that the hyperparameters will affect the performance of the model, the P0 corresponding to a certain model may be higher, but after optimizing the hyperparameters, the judgment value will decrease (the performance of the model is better), so it is necessary to optimize the hyperparameters first. Bayesian optimization can explore hyperparameters more effectively by constructing a probabilistic model of the objective function. Compared with random search and other methods, it can usually find the optimal hyperparameters faster and reduce the number of evaluations. However, Bayesian optimization requires the construction of a proxy model (such as a Gaussian process or a tree structure). This model needs to be updated after each new hyperparameter evaluation. The computational cost of training the model increases significantly when the dimension of the parameter space is high. Therefore, in this scheme, random search is first used to optimize the hyperparameters and determine the change in the judgment value of the initial model (that is, the change in performance). The lower the judgment value, the better the performance of the model. The selection value is used as a standard to measure the change in the judgment value. The larger the maximum judgment value difference, the higher the performance of the model after a certain optimization of the hyperparameters, and the smaller the selection value should be (because the subsequent selection value is smaller). The smallest initial model is used as the target model), and the relationship between the remaining parameters and the selection values ​​can refer to the above ideas; after that, after the target model is determined, the optimal hyperparameters of the target model are determined based on Bayesian optimization, the selection value of the target model at this time is determined, and the target model corresponding to the minimum selection value is used as the accurate model; it is worth noting that the initial model with better performance after optimizing the hyperparameters N times is screened out from the initial model, thereby reducing the computational cost brought by Bayesian optimization for all initial models; random search does not need to build complex models or update proxy models, it only evaluates by randomly sampling hyperparameter combinations, so the computational overhead of each iteration is small; finally, the standard range of values ​​of the target feature is determined based on the accurate model and the prediction set.The present invention can improve the prediction accuracy of capacitor performance by optimizing models and feature selection, and reveal the hidden mechanism between material structure and capacitor performance through machine learning, thereby promoting the preparation of high-performance biochar supercapacitors. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below in conjunction with the accompanying drawings.

[0018] Figure 1 It is a flow chart of a capacitor performance prediction method based on machine learning in the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] See also Figure 1 As shown, the present invention is a capacitor performance prediction method based on machine learning, comprising the following steps: S1: collecting target data of biochar supercapacitors, wherein the target data includes target factors and specific capacitance, and the target factors are factors that affect specific capacitance, preprocessing the target factors to obtain target features, and determining a training set and a prediction set based on the target features; S2: Establish a supercapacitor performance prediction model based on a machine learning algorithm, use it as an initial model, train the initial model based on a cross-validation method and the training set, and determine the evaluation index of the trained initial model, wherein the evaluation index includes a determination coefficient ZB fir , root mean square error ZB sec And the mean absolute error ZB thi , calculate the judgment value P = (ε-ZB fir )*(ZB sec +ZB thi ), ε is the preset correction constant and ε>1; The hyperparameters of the initial model are optimized based on a random search algorithm, and the judgment value after optimizing the hyperparameters is determined to generate coordinate points (a, P a ), P a represents the judgment value after the a-th optimization of the hyperparameter, and the coordinate points are fitted to obtain a fitting curve f(P); Calculate the judgment value difference PCZ a =|P a -P a-1 |, get the maximum judgment value difference PCZmax , calculate the selected value , where η is a preset correction coefficient, and N represents the total number of times the hyperparameter is optimized; S3: Sort the selected values ​​by size. The larger the selected value, the lower the ranking. The initial model corresponding to the top 30% of the selected values ​​in the sorting is used as the target model. The optimal hyperparameters of the target model are determined based on Bayesian optimization. The selected value of the target model at this time is determined, and the target model corresponding to the minimum selected value is used as the accurate model. The standard value range of the target feature is determined based on the precise model and the prediction set. When the value of the target feature falls within the standard value range, the specific capacitance of the biochar supercapacitor is the highest.

[0021] It should be noted that first, the target data is collected, and the target factors in the target data are preprocessed to obtain the target features, and the training set and the prediction set are obtained according to the target features; it is understandable that the target data can be obtained from published literature, and the target factors include but are not limited to the porous structure characteristics and chemical composition characteristics of biochar supercapacitor electrode materials, the porous structure characteristics include but are not limited to specific surface area, micropore specific surface area and pore size, and the chemical composition characteristics include but are not limited to carbon content and nitrogen content, which will not be repeated here; then the initial model is established based on machine learning, the evaluation index of the initial model is determined, and the judgment value is calculated; it is worth noting that the judgment value is used to comprehensively measure the performance of the initial model, and the smaller the judgment value, the better the performance of the model. Well, specifically, the closer the determination coefficient is to 1, the better the performance of the initial model, and the smaller the root mean square error and the mean absolute error, the better the performance of the initial model. After that, the hyperparameters of the initial model are optimized based on the random search algorithm, and the judgment value is re-determined after each optimization, and the target image is drawn. It can be understood that when the hyperparameters are not optimized, the judgment value of the initial model is P0, and the corresponding coordinate point is (0, P0). Considering that the hyperparameters will affect the performance of the model, the P0 corresponding to a certain model may be higher, but after optimizing the hyperparameters, the judgment value will decrease (the performance of the model is better), so it is necessary to optimize the hyperparameters first. Bayesian optimization can explore the hyperparameter space more effectively by constructing a probabilistic model of the objective function. Compared with random search, Methods such as random search can usually find the optimal hyperparameters faster and reduce the number of evaluations. However, Bayesian optimization requires the construction of a proxy model (such as a Gaussian process or a tree structure). This model needs to be updated after each new hyperparameter evaluation. The computational cost of training the model increases significantly when the dimension of the parameter space is high. Therefore, in this scheme, random search is first used to optimize the hyperparameters and determine the change in the judgment value of the initial model (that is, the change in performance). The lower the judgment value, the better the performance of the model. The selection value is used as a standard to measure the change in the judgment value. The larger the maximum judgment value difference, the higher the performance of the model after a certain optimization of the hyperparameters, and the smaller the selection value should be (because the initial model with a smaller selection value is used as the target model later). The target model), the relationship between the remaining parameters and the selection value can refer to the above ideas, f (N) represents the function value obtained by substituting N into the fitting curve f (P); after that, after the target model is determined, the optimal hyperparameters of the target model are determined based on Bayesian optimization, the selection value of the target model at this time is determined, and the target model corresponding to the minimum selection value is used as the accurate model; it is worth noting that the initial model (i.e., the target model) with better performance after using random search to optimize the hyperparameters N times is selected from the initial model, thereby reducing the computational cost of using Bayesian optimization for all initial models; random search does not need to build complex models or update proxy models, it only evaluates by randomly sampling hyperparameter combinations, so the computational overhead of each iteration is small;Finally, the standard range of values ​​of the target feature is determined based on the precise model and the prediction set. ;

[0022] In another preferred embodiment of the present invention, in step S3, the process of determining the standard range of values ​​of the target feature based on the precise model and the prediction set specifically includes: Step 1: Determine the evaluation index of the accurate model based on the cross-validation method and the prediction set, determine the judgment value Pcur at this time, set the judgment value threshold Pzq, and when the judgment value Pcur>Pzq, execute step 2; Step 2: Determine the average absolute contribution rate of the target feature based on the precise model, and sort the target features according to the average absolute contribution rate. The larger the average absolute contribution rate, the lower the corresponding target feature is ranked. Step 3: removing the last target feature in the ranking from the prediction set, and repeating steps 1 to 2 until the corresponding judgment value is less than the judgment value threshold after a target feature is removed from the prediction set; Determine the average absolute contribution rate of the target features that have not been removed at this time, and sort the target features that have not been removed at this time according to the size of the average absolute contribution rate. The larger the average absolute contribution rate, the higher the corresponding target feature is ranked. The first C target features in the sorting are taken as the main features, and C is a preset number.

[0023] Step 4: Determine the standard range of values ​​of the main features based on the partial dependence diagram, the individual conditional expectation diagram and the precise model.

[0024] It is worth noting that when Pcur>Pzq, it means that the performance of the model has not met expectations. Therefore, we choose to remove the target features with small average absolute contribution rates. These features have limited contribution to the predictive performance of the model and may introduce noise, thus affecting the overall performance of the model. By removing low-contribution features, we can simplify the model, reduce the risk of overfitting, and improve the generalization ability of the model. After the performance of the model meets expectations, we select the main features with the average absolute contribution rate and determine the standard range of the main features based on the partial dependence diagram, individual conditional expectation diagram, and precise model. The partial dependence diagram and individual conditional expectation diagram are important tools for visualizing and analyzing the impact of features in machine learning models. Their main function is to help understand the relationship between features and prediction results. By combining the analysis of partial dependence diagrams and individual conditional expectation diagrams, we can more comprehensively understand the role of the main features, thereby providing a reliable standard value range for the performance optimization of biochar supercapacitors, which will not be elaborated here. Interpretability tools are used to explain the prediction results, understand the input feature attributes that have a greater impact on the target attribute (specific capacitance), and how they affect the energy storage performance of supercapacitors, and obtain the controllable biochar supercapacitor electrode material synthesis knowledge system learned during the model establishment process, which will help accelerate the R&D cycle of high-performance supercapacitors and promote innovative development in the application field of biochar supercapacitors.

[0025] In another preferred embodiment of the present invention, in step three, when the number of target features that have not been removed is C and the judgment value of the precise model is still greater than the judgment value threshold, a warning message is sent to a preset manager.

[0026] In another preferred embodiment of the present invention, the step S3 further includes the following steps: Determining the SHAP value of the main feature; When the SHAP value of the main feature is positive, the main feature is positively correlated with the specific capacitance of the biochar supercapacitor; When the SHAP value of the main feature is negative, the main feature is negatively correlated with the specific capacitance of the biochar supercapacitor.

[0027] It should be noted that the SHAP value considers all possible feature combinations and evaluates the impact of each feature on the prediction results under different combinations, thereby accurately calculating the actual contribution of a single feature to the final output; when the SHAP value is positive, it means that an increase in the feature will lead to an increase in the model's predicted results, indicating a positive correlation. For example, an increase in the content of a certain chemical component may increase the specific capacitance; when the SHAP value is negative, it means that an increase in the feature will lead to a decrease in the model's predicted results, indicating a negative correlation; the introduction of the SHAP value enhances the interpretability of the model, allowing researchers to clearly understand how each major feature specifically affects the performance of the capacitor, thereby providing more precise guidance on optimizing design and material selection.

[0028] In another preferred embodiment of the present invention, in step S1, the process of preprocessing the target factor to obtain the target feature specifically includes: Standardize the numerical target factors to eliminate the dimension; Encode non-numeric target factors through one-hot encoding; The outliers in the target factor are removed by the isolation forest algorithm.

[0029] It should be noted that for numerical target factors, a standardization method is used to eliminate the dimension effect. This process unifies the values ​​of different scales into a standard range (specifically set based on experience), thereby improving the efficiency and stability of model training; for non-numerical target factors, one-hot encoding is used to convert them into numerical form, thereby converting each category feature into a binary vector to ensure that the model can effectively utilize these category information without introducing false order relationships; the isolation forest algorithm is used to identify and remove outliers in the target factor. Isolation forest is a tree-based algorithm that can effectively identify outliers in the data, thereby improving the robustness of the model and reducing the negative impact on model performance.

[0030] In another preferred embodiment of the present invention, in step S1, the process of determining the training set and the prediction set based on the target feature specifically includes: Obtain the specific capacitance D corresponding to different target features, establish a target mapping between the target feature and the specific capacitance D, and establish a sample set YJH= (Y1, Y2, ..., Y m ), Y m represents the mth target mapping, and divides the sample set into a training set and a prediction set according to a preset ratio.

[0031] It can be understood that the ratio of the training set to the prediction set is 7:3.

[0032] In another preferred embodiment of the present invention, in step S2, in the process of training the initial model based on the cross-validation method and the training set and determining the evaluation index of the trained initial model, five-fold cross-validation is used.

[0033] In another preferred embodiment of the present invention, in step S1, in the process of establishing the supercapacitor performance prediction model based on a machine learning algorithm, the machine learning algorithm includes a random forest, a decision tree, an artificial neural network, and a support vector machine.

[0034] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A capacitor performance prediction method based on machine learning, characterized in that: The following steps are involved: S1: collecting target data of biochar supercapacitors, wherein the target data includes target factors and specific capacitance, and the target factors are factors that affect specific capacitance, preprocessing the target factors to obtain target features, and determining a training set and a prediction set based on the target features; S2: Establish a supercapacitor performance prediction model based on a machine learning algorithm, use it as an initial model, train the initial model based on a cross-validation method and the training set, and determine the evaluation index of the trained initial model, wherein the evaluation index includes a determination coefficient ZB fir , root mean square error ZB sec And the mean absolute error ZB thi , calculate the judgment value P = (ε-ZB fir )*(ZB sec +ZB thi ), ε is the preset correction constant and ε>1; The hyperparameters of the initial model are optimized based on a random search algorithm, and the judgment value after optimizing the hyperparameters is determined to generate coordinate points (a, P a ), P a represents the judgment value after the a-th optimization of the hyperparameter, and the coordinate points are fitted to obtain a fitting curve f(P); Calculate the judgment value difference PCZ a =|P a -P a-1 |, get the maximum judgment value difference PCZ max , calculate the selected value , where η is a preset correction coefficient, and N represents the total number of times the hyperparameter is optimized; S3: Sort the selected values ​​by size. The larger the selected value, the lower the ranking. The initial model corresponding to the top 30% of the selected values ​​in the sorting is used as the target model. The optimal hyperparameters of the target model are determined based on Bayesian optimization. The selected value of the target model at this time is determined, and the target model corresponding to the minimum selected value is used as the accurate model. The standard value range of the target feature is determined based on the precise model and the prediction set. When the value of the target feature falls within the standard value range, the specific capacitance of the biochar supercapacitor is the highest.

2. The capacitor performance prediction method based on machine learning according to claim 1, characterized in that: In step S3, the process of determining the standard range of values ​​of the target feature based on the precise model and the prediction set specifically includes: Step 1: Determine the evaluation index of the accurate model based on the cross-validation method and the prediction set, determine the judgment value Pcur at this time, set the judgment value threshold Pzq, and when the judgment value Pcur>Pzq, execute step 2; Step 2: Determine the average absolute contribution rate of the target feature based on the precise model, and sort the target features according to the average absolute contribution rate. The larger the average absolute contribution rate, the lower the corresponding target feature is ranked. Step 3: removing the last target feature in the ranking from the prediction set, and repeating steps 1 to 2 until the corresponding judgment value is less than the judgment value threshold after a target feature is removed from the prediction set; Determine the average absolute contribution rate of the target features that have not been removed at this time, and sort the target features that have not been removed at this time according to the size of the average absolute contribution rate. The larger the average absolute contribution rate, the higher the corresponding target feature is ranked. The first C target features in the sorting are taken as the main features, where C is a preset number; Step 4: Determine the standard range of values ​​of the main features based on the partial dependence diagram, the individual conditional expectation diagram and the precise model.

3. The capacitor performance prediction method based on machine learning according to claim 2, characterized in that: In the step three, when the number of target features that have not been removed is C and the judgment value of the precise model is still greater than the judgment value threshold, a warning message is sent to the preset management personnel.

4. The capacitor performance prediction method based on machine learning according to claim 2, characterized in that: The step S3 further includes the following steps: Determining the SHAP value of the main feature; When the SHAP value of the main feature is positive, the main feature is positively correlated with the specific capacitance of the biochar supercapacitor; When the SHAP value of the main feature is negative, the main feature is negatively correlated with the specific capacitance of the biochar supercapacitor.

5. The capacitor performance prediction method based on machine learning according to claim 1, characterized in that: In step S1, the process of preprocessing the target factor to obtain the target feature specifically includes: Standardize the numerical target factors to eliminate the dimension; Encode non-numeric target factors through one-hot encoding; The outliers in the target factor are removed by the isolation forest algorithm.

6. The capacitor performance prediction method based on machine learning according to claim 1, characterized in that: In step S1, the process of determining the training set and the prediction set based on the target feature specifically includes: Obtain the specific capacitance D corresponding to different target features, establish a target mapping between the target feature and the specific capacitance D, and establish a sample set YJH= (Y1, Y2, ..., Y m ), Y m represents the mth target mapping, and divides the sample set into a training set and a prediction set according to a preset ratio.

7. The capacitor performance prediction method based on machine learning according to claim 1, characterized in that: In the step S2, five-fold cross validation is used in the process of training the initial model based on the cross validation method and the training set and determining the evaluation index of the trained initial model.

8. The capacitor performance prediction method based on machine learning according to claim 1, characterized in that: In the step S1, in the process of establishing the supercapacitor performance prediction model based on the machine learning algorithm, the machine learning algorithm includes random forest, decision tree, artificial neural network, and support vector machine.

Citation Information

Patent Citations

  • Anode aluminum foil performance prediction method for aluminum electrolytic capacitor

    CN112289392A

  • Machine learning model optimization method and device based on model hyper-parameters

    CN115952418A

  • Specific capacitance prediction method of supercapacitor based on biomass derived carbon material

    CN118395834A

  • Capacitor residual life prediction method and system based on machine learning

    CN118484680A

  • Failure prediction apparatus for electric device and failure prediction method using the same

    US20170206459A1