Self-adaptive energy management method of fuel cell hybrid energy storage system

Abstract

The invention provides an adaptive energy management method for a fuel cell hybrid energy storage system, which comprises the following steps of: performing principal component analysis on working condition data, and selecting a plurality of characteristic parameters to obtain a sample data set; taking the feature data as input sample points to carry out K-means clustering to divide the feature data into a training set and a test set; inputting the training set into a support vector machine model for training; testing the trained support vector machine model through a cross validation method to obtain a trained support vector machine model; inputting the collected real-time working condition data into the trained support vector machine model to obtain an identified category, and inputting the identified category into the vehicle dynamics model to obtain the whole vehicle demand power; and inputting the required power of the whole vehicle and the SOC of the lithium ion battery into a fuzzy controller to obtain the power of the proton exchange membrane fuel cell so as to realize the power distribution of the hybrid energy storage system. According to the invention, the adaptability of the hybrid energy storage system to complex working conditions is effectively improved, and the power output of the proton exchange membrane fuel cell and the lithium ion battery can be effectively balanced.

Description

Technical Field

[0001] The present invention relates to the technical field of system control and motor drive, and in particular to an adaptive energy management method for a fuel cell hybrid energy storage system, which can be used in the fields of electric drive systems of electric vehicles, hybrid energy storage system optimization, etc. Background Art

[0002] As a highly efficient and clean energy conversion device, proton exchange membrane fuel cells (PEMFCs) can significantly improve energy efficiency and reduce greenhouse gas and pollutant emissions, playing a key role in addressing global climate change and environmental pollution. Their high energy conversion efficiency and rapid start-up at low temperatures make them a core technology driving sustainable development in transportation, distributed power generation, and renewable energy integration, and are crucial for achieving carbon neutrality.

[0003] Proton exchange membrane fuel cells have insufficient dynamic response capabilities in large-scale and high-power transportation and frequently changing load scenarios, making it difficult to achieve rapid state changes in working conditions. Therefore, proton exchange membrane fuel cells usually work together with lithium-ion batteries to form a hybrid energy storage system. The hybrid energy storage system needs to coordinate and optimize the power flow between the proton exchange membrane fuel cell and the lithium-ion battery based on multiple factors such as power demand and driving conditions to achieve rational and efficient energy distribution. Scientific energy management strategies can significantly improve the efficiency of hydrogen energy utilization, reduce operating costs, and extend the service life of fuel cells. In hybrid energy storage systems, the development of precise energy management strategies is crucial to optimizing system performance, improving energy utilization efficiency, and reducing operating costs. It is a key link in achieving efficient system operation.

[0004] The energy management strategies of hybrid energy storage systems can be mainly divided into two categories: rule-based control strategies and optimization-based control strategies. Rule-based control strategies achieve power distribution through preset thresholds or logical conditions, while optimization-based control strategies achieve optimal control of system performance through mathematical modeling and optimization algorithms. Rule-based control strategies are divided into two categories: deterministic rule control strategies and fuzzy logic control strategies. Deterministic rule control strategies have the characteristics of simple implementation and high computational efficiency, but lack adaptability to complex working conditions. Fuzzy logic control strategies introduce fuzzy reasoning mechanisms and use expert experience or historical data to build a fuzzy rule base. They can handle the uncertainty and nonlinear characteristics of the system to a certain extent and have good robustness. However, the design and optimization of its rules still rely on experience, and it is difficult to guarantee global optimality.

[0005] Patent application number 202210771603.6 discloses a multi-objective optimization method for hydrogen fuel cell hybrid vehicle energy management, belonging to the technical field of hybrid vehicle energy management. By integrating driving behavior into the hydrogen fuel cell hybrid vehicle energy management method, an online hybrid adaptive noise-resistant clustering algorithm and a heuristic self-learning labeling algorithm HSLSVM / NN are used to obtain a driving behavior recognition model. Based on the driving behavior recognition model, the Pontryagin Minimum Principle (PMP) is combined to obtain a multi-objective optimized A-ECMS energy management strategy that adapts to driving behavior. At the same time, the motor load power demand and its rate of change and / or the weighted SOC of the energy storage system are comprehensively considered in the hydrogen fuel cell hybrid vehicle energy management to obtain the corresponding optimal vehicle energy management system. This solves the technical problem that most existing energy management strategies only consider external driving conditions, resulting in an unreasonable allocation of demand power, energy waste, and a short power supply life. However, the above patent lacks adaptability to complex operating conditions. Summary of the Invention

[0006] In response to the technical problems that the energy management methods of existing hybrid energy storage systems have low adaptability to complex working conditions and low energy utilization efficiency, the present invention proposes an adaptive energy management method for fuel cell hybrid energy storage systems. The present invention fully considers the frequently changing power requirements, while taking into account the adaptability and robustness of the system under dynamic load conditions, ensuring that the vehicle maintains stable and efficient driving performance under complex and changeable working conditions; therefore, the present invention is an energy management strategy that can respond in real time and optimize energy distribution, which is of great significance for improving the overall performance and energy utilization efficiency of automobiles under complex working conditions.

[0007] In order to achieve the above object, the technical solution of the present invention is implemented as follows: a fuel cell hybrid energy storage system adaptive energy management method, the steps of which are as follows:

[0008] Step 1: Perform principal component analysis on the working condition data under known working conditions, select multiple characteristic parameters to reflect the working condition data, and obtain a sample data set;

[0009] Step 2: K-means cluster analysis: Use the characteristic data of the sample data set as input sample points for K-means clustering, divide the known working conditions into three modes, and divide them into training sets and test sets;

[0010] Step 3: Input the training set sample data in the training set into the support vector machine model for training; based on the test set, test the trained support vector machine model through the cross-validation method to obtain the optimal parameters of the support vector machine model and obtain the trained support vector machine model;

[0011] Step 4: Input the collected real-time operating condition data into the support vector machine model trained in step 3 to obtain the identified category, and input the operating condition speed of the identified category into the vehicle dynamics model to obtain the required power of the entire vehicle;

[0012] Step 5: Input the obtained vehicle demand power and the SOC of the lithium-ion battery into the fuzzy controller to obtain the proton exchange membrane fuel cell power and realize the power distribution of the hybrid energy storage system.

[0013] Preferably, the operating condition data is standardized before being subjected to principal component analysis, and the method for performing feature screening on the operating condition data by principal component analysis is:

[0014] (1) Constructing characteristic parameter matrix: Four working conditions, FTP, HWFIT, NYCC, and CLTP, were selected for research. The data were divided into multiple segments by fixed step size identification, and the time interval of each segment was 15s. The characteristic parameter matrix X was constructed based on the extracted working condition data, which includes 20 characteristic parameters: acceleration, maximum acceleration, maximum deceleration, acceleration time ratio, deceleration time ratio, uniform speed time ratio, parking time ratio, 0-20 speed time ratio, 20-40 speed time ratio, 40-60 speed time ratio, 60-80 speed time ratio, maximum speed, average speed, average speed excluding parking time, speed standard deviation, acceleration standard deviation, average acceleration in acceleration section, average deceleration in deceleration section, acceleration standard deviation in acceleration process, deceleration standard deviation in deceleration process, and driving distance.

[0015] (2) Calculate the sample correlation coefficient matrix C directly from the characteristic parameter matrix X;

[0016] (3) Calculate the eigenvalues ​​and eigenvectors of 20 characteristic parameters based on the sample correlation coefficient matrix C;

[0017] (4) Calculate the contribution rate and cumulative contribution rate of each principal component based on the eigenvalue;

[0018] (5) Construct principal component values: select principal components with cumulative contribution rates of more than 90% as representative characteristic parameters.

[0019] Preferably, the K-means clustering is implemented as follows:

[0020] (1) Randomly select three initial centroids, denoted as μ1, μ2, and μ3;

[0021] (2) Assign sample points to the nearest cluster:

[0022] Calculate each sample point x i , i = 1, 2, ..., n and the Euclidean distance of each centroid, and assign the sample point to the cluster C with the closest distance k ;

[0023] (3) Update the centroid: For each cluster C k , recalculate the center of mass μ k is the mean of all sample points in the cluster;

[0024] (4) Iterative optimization: Repeat steps (2) and (3) to recalculate the centroid of each cluster until the centroid position converges;

[0025] (5) Objective function minimization: Minimize the sum of squares from the sample points within the cluster to the centroid:

[0026] The K-means clustering algorithm uses the characteristic data obtained from principal component analysis as input sample points and divides the data into three different cluster categories: C1, C2, and C3. C1 represents the low-speed mode of urban congestion; C3 represents the high-speed mode; and C2 represents the medium-speed mode.

[0027] Preferably, the training set sample data is:

[0028]

[0029] Where x i is an n-dimensional feature vector; y i is the classification category, m is the number of samples;

[0030] The hyperplane of the support vector machine is obtained by the linear equation ω T x+b=0 to represent;

[0031] By applying the soft margin optimization method, the minimum distance problem from the sample point to the hyperplane and the model misclassification problem are transformed into a constrained optimization problem:

[0032]

[0033] sty i (ω·x i +b)≥1-ξ i

[0034] In the formula, ω is the weight factor; b is the bias term; p is the adjustment parameter used to change the penalty form for misclassified points; C is the penalty parameter; ξ i is a slack variable; λ is a regularization parameter used to control the complexity of feature weights, ω k is the weight of the kth feature.

[0035] The constrained optimization problem is solved using the sequential quadratic programming method, and the optimal solution obtained is the optimal hyperplane; the support vector machine uses the kernel function to transform the samples from the original feature space into a high-dimensional feature space; the penalty parameter of the support vector machine model is found through the cross-validation method, and the optimal combination of the optimal penalty parameter C and the kernel function parameter γ is found within the specified parameter range through grid search.

[0036] Preferably, the steps of solving the constrained optimization problem using the sequential quadratic programming method are:

[0037] 1). Initialization: Select the initial point (ω (0) , b (0) ,ξ (0) ), set the number of iterations t = 0 and the convergence threshold ε;

[0038] 2). At the current point (ω (t) , b (t) ,ξ (t) ), the Lagrangian function of the constrained optimization problem is:

[0039]

[0040] Where, α i and β i are the Lagrange multipliers associated with the classification constraints and slack variables, respectively;

[0041] Calculate the gradient of the objective function and the constraints, that is, the gradient of the Lagrangian function with respect to the weight factor ω, the bias term b, and the slack variable ξ:

[0042]

[0043] Among them, ω (t) represents the weight factor obtained at the tth iteration;

[0044] 3) Construct the quadratic programming sub-problem:

[0045]

[0046] The constraints are:

[0047]

[0048] 4) Solve the quadratic programming subproblem using the active set method and obtain the search direction (Δω, Δb, Δξ);

[0049] 5) Perform line search along the search direction and update the current point:

[0050] ω (t+1) =ω (t) +αΔω

[0051] b (t+1) =b (t) +αΔb

[0052] ξ (t) =ξ (t) +αΔξ

[0053] Where α is the step size; Δω, Δb, Δξ are increments; b (t+1) 、ω (t+1) ,ξ (t+1) is the next step value of the iteration;

[0054] Check the convergence condition: if the convergence condition is met, stop the iteration; otherwise, set the number of iterations t=t+1 and continue the iteration. Preferably, the radial basis function kernel function is used to construct the classification function, and the radial basis function kernel function is:

[0055] K(x i ,x j )=exp(-γ||x i -x j || 2 )

[0056] Where γ is the kernel function parameter, x i 、x j are two different eigenvectors in the sample;

[0057] The classification function is: Here, sgn represents the sign function.

[0058] Preferably, the method for finding the optimal combination of the optimal penalty parameter C and the kernel function parameter γ is:

[0059] ① Parameter grid generation: Generate a 51*51 two-dimensional parameter grid matrix, and set the range of penalty parameter C and kernel function parameter γ to: C=2 x ,γ=2 y ;Wherein, exponential parameters x,y∈{-10,-9.8,...,10}, step size 0.2;

[0060] ② Variable initialization: Create a 51*51 zero matrix, record the cross-validation accuracy of each parameter combination, the initial penalty factor is: C=1, the initial kernel function parameter is: γ=0.1, the initial best accuracy is: bestacc=0; set the minimum positive number ε=10 -4 ;

[0061] ③ Grid search and cross-validation: Traverse all possible combinations of penalty parameters and kernel function parameters, evaluate model performance through 5-fold cross-validation to find the optimal parameter pair: divide the training dataset evenly into 5 subsets D1, D2, ..., D5. For each round s∈{1,2,...,5}, use subset Ds as the validation set. The training dataset is:

[0062] D train =D1∪...∪D s-1 ∪D s+1 ∪...∪D5, use the current parameters to train the dataset D train The SVM model is trained on the subset Ds and the classification accuracy of the SVM model is calculated: The accuracy of the five rounds of verification is averaged as the final performance indicator of the current parameter pair:

[0063] ④ Dynamically update the optimal parameters and determine the optimal values: Compare the current accuracy with the previous accuracy. If the accuracy of the current parameter combination is higher than the historical best value, update the optimal parameters to the current penalty parameters and kernel function parameters, and finally use the optimal parameters to reintroduce the SVM model.

[0064] Preferably, a vehicle dynamics model is constructed based on the longitudinal dynamics characteristics of the vehicle, specifically as follows:

[0065] The rolling resistance of the vehicle is: F r =C r ·m·g·cosθ; where C r is the rolling resistance coefficient; m is the vehicle mass; g is the acceleration of gravity; θ is the road slope;

[0066] The air resistance is: Where ρ is the air density; C d is the air resistance coefficient; A f is the effective frontal area of ​​the vehicle; v is the vehicle speed; v w is the real-time wind speed in the direction of vehicle movement;

[0067] The climbing resistance caused by the road slope is: F g =m·g·sinθ;

[0068] The traction force is:

[0069] The required power of the vehicle is related to the traction and vehicle speed. The calculation formula is: P d =F t ·v.

[0070] Preferably, the implementation method of the fuzzy controller in step 5 is: establishing a judger and classifying the vehicle required power P according to the classification result of the support vector machine model. d The data is split and input into three different fuzzy logic control systems. The fuzzifier converts the input variables into fuzzy values, and the output is obtained through fuzzy reasoning and sent to the defuzzifier. The precise value of the proton exchange membrane fuel cell power is obtained through defuzzification.

[0071] The fuzzy reasoning process includes two core components: one is the membership function of the input and output variables, and the other is the fuzzy logic rules constructed based on the membership function; the judge transmits the vehicle's required power to three fuzzy detectors respectively by detecting categories 1, 2, and 3.

[0072] Preferably, in fuzzy logic control, the SOC of the lithium-ion battery has three membership functions, representing low L, medium M, and high H, which are trapezoidal, triangular, and trapezoidal membership functions respectively; the vehicle demand power P d It has five input membership functions, namely, very low VL, low L, medium M, high H, and very high VH, which are the membership functions of trapezoid, trapezoid, triangle, triangle, and trapezoid respectively;

[0073] Output variable PEMFC power P fc The membership functions are divided into nearly zero Z, positive small PS, medium PM, positive large PL, large PB, and maximum PV. The first five are triangular membership functions, and the sixth is a trapezoidal membership function.

[0074] When the SOC of the lithium-ion battery is low, the PEMFC will simultaneously power the load and the lithium-ion battery, regardless of the power demand of the vehicle. When the SOC of the lithium-ion battery is medium, the PEMFC and the lithium-ion battery will meet the power demand in different output ratios depending on the operating conditions. When the SOC of the lithium-ion battery is high, the lithium-ion battery alone will provide power in low-speed mode, and the PEMFC and lithium-ion battery will provide power together in medium- and high-speed modes.

[0075] Different fuzzy rules are designed for the three driving modes of low speed, medium speed and high speed:

[0076]

[0077] The fuzzy logic control system derives the proton exchange membrane fuel cell power under the corresponding driving mode at each moment according to the membership function and the set fuzzy rules, thereby realizing the power distribution of the hybrid energy storage system.

[0078] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a fuzzy logic energy management strategy based on support vector machine pattern recognition for a hybrid energy storage system of proton exchange membrane fuel cells and lithium-ion batteries, which includes three parts: ① Feature extraction of known operating condition data to construct a feature parameter matrix; then, principal component analysis is used to reduce the dimension of the matrix containing 20 feature parameters, and 7 principal components are screened out as representative eigenvalues; ② The driving cycle is classified by K-means cluster analysis, and the operating conditions are divided into three categories: low speed, medium speed, and high speed according to the 7 eigenvalues, and the classification results obtained by cluster analysis are used to train the support vector machine model; then the support vector machine classifies the real-time input operating condition data, and passes the classification results as input to the fuzzy logic control; ③ The fuzzy logic control takes the lithium-ion battery SOC and the required power of the whole vehicle as input, and presets different fuzzy logic rule bases for the three operating conditions of low speed, medium speed, and high speed. Fuzzy logic control uses the operating condition categories identified by the support vector machine to perform differentiated fuzzy logic reasoning and processing on the input power and lithium-ion battery SOC, and determines the PEMFC power. Principle analysis and experimental results demonstrate that this invention effectively improves the hybrid energy storage system's adaptability to complex operating conditions and can effectively balance the power output of the PEMFC and lithium-ion battery. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0080] Figure 1 Flowchart of the present invention.

[0081] Figure 2 Schematic diagram of the K-means cluster analysis results of the present invention.

[0082] Figure 3 Schematic diagram of the classification results of the test set of the present invention.

[0083] Figure 4 This is the FTP-HWFIT working condition identification result diagram.

[0084] Figure 5 Schematic diagram of the membership function of SOC of the present invention.

[0085] Figure 6 The vehicle power requirement P is d Schematic diagram of the membership function.

[0086] Figure 7 is the proton exchange membrane fuel cell power P fc Schematic diagram of the membership function.

[0087] Figure 8 The power of the lithium ion battery and the power of the proton exchange membrane fuel cell P of the present invention are fc Simulation diagram of .

[0088] Figure 9 The simulated vehicle power requirement and the proton exchange membrane fuel cell power P fc Simulation diagram of the membership function.

[0089] Figure 10 This is a comparison diagram of the SOC curves of the three strategies of the present invention. DETAILED DESCRIPTION

[0090] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0091] like Figure 1 As shown, a fuel cell hybrid energy storage system adaptive energy management method, its working steps are as follows:

[0092] Step 1: Principal component analysis: Collect working condition data under known working conditions, perform principal component analysis on the working condition data, select multiple characteristic parameters to reflect the working condition data, and obtain a sample data set.

[0093] The present invention performs principal component analysis to obtain a sample data set, and trains the subsequent support vector machine model based on this. When performing principal component analysis, the known working condition data is first standardized, and its main functions are as follows: 1. Eliminate the influence of dimension. Directly analyzing data of different dimensions and orders of magnitude may cause features with a larger numerical range to dominate the principal component analysis. Standardization can eliminate the influence of such dimensions and orders of magnitude, so that each feature is in an equal position in the analysis; 2. Ensure the rationality of the principal component: the standardized data enables the calculation of the principal component to reflect the actual correlation structure of the data, rather than being dominated by the numerical size of the data. If standardization is not performed, the principal component may be biased towards variables with a larger numerical range, thereby failing to accurately extract the main features and structural information of the data. 3. Improve the accuracy of principal component analysis: through standardization, the distribution of data in each feature direction is more uniform, which is conducive to the accuracy of the subsequent calculation of eigenvalues ​​and eigenvectors, thereby improving the reliability and accuracy of the results of principal component analysis.

[0094] The number of driving condition characteristic parameters can reach 62, of which speed and acceleration are the most important. However, using only speed and acceleration for analysis will lead to information loss, seriously affecting the accuracy of the classification results. In addition, using too many parameters to describe driving conditions will cause data redundancy and increase computational complexity. Principal component analysis is a multivariate statistical analysis method that filters multiple related feature quantities and retains the principal components, thereby simplifying the data structure. The present invention uses principal component analysis to perform feature screening on driving condition data. The specific process is as follows:

[0095] (1) Constructing the characteristic parameter matrix

[0096] The present invention selects four working conditions for research: FTP (Federal Test Procedure, U.S. Federal Test Procedure), HWFIT (Highway Fuel Economy Test, Highway Fuel Economy Test), NYCC (New York City Cycle, New Zealand City Cycle), and CLTP (China Light-duty Vehicle Test Cycle, China's independently developed light vehicle test cycle). A fixed-step recognition method is used to divide the four working conditions into multiple data segments, with a time interval of 15 seconds for each segment. This facilitates the separate processing, analysis, and storage of each data segment, making data management more refined and improving the efficiency and accuracy of data processing. Then, a characteristic parameter matrix is ​​constructed based on the extracted working condition data, which includes 20 characteristic parameters: acceleration, maximum acceleration, maximum deceleration, acceleration time ratio, deceleration time ratio, uniform speed time ratio, parking time ratio, 0-20 speed time ratio, 20-40 speed time ratio, 40-60 speed time ratio, 60-80 speed time ratio, maximum speed, average speed, average speed excluding parking time, speed standard deviation, acceleration standard deviation, average acceleration in acceleration section, average deceleration in deceleration section, acceleration standard deviation during acceleration, deceleration standard deviation during deceleration, and travel distance. Suppose there are n' samples, each sample has 20 characteristic parameters, and the constructed characteristic parameter matrix X is:

[0097]

[0098] Among them, x i'j represents the jth parameter of the i'th sample.

[0099] (2) Calculate the sample correlation coefficient matrix

[0100] The sample correlation coefficient matrix C is directly calculated from the feature parameter matrix X. The values ​​in the correlation coefficient matrix C represent the correlation coefficient between each two feature parameters. The sample correlation coefficient matrix C is:

[0101]

[0102] (3) Calculate eigenvalues ​​and eigenvectors:

[0103] Then, the eigenvalues ​​and eigenvectors of the 20 characteristic parameters are calculated according to the sample correlation coefficient matrix C, and the eigenvalue equation is solved:

[0104] det(C-λI)=0

[0105] Thus we can get 20 eigenvalues ​​λ1, λ2…λ 20 and the corresponding eigenvectors v1, v2…v 20 .

[0106] (4) Calculate the contribution rate and cumulative contribution rate of the principal component:

[0107] The contribution rate and cumulative contribution rate of each principal component are further calculated through the eigenvalue. The contribution rate of the kth principal component is:

[0108]

[0109] The cumulative contribution rate of the kth principal component is:

[0110]

[0111] The eigenvalues, contribution rates, and cumulative contribution rates of the principal component analysis are shown in Table 1. Table 1 shows that among the 20 principal components, the cumulative contribution rate of the characteristic parameters Y1-Y7 has reached 91.77%.

[0112] (5) Construct principal component values:

[0113] The principal component values ​​are constructed based on the eigenvectors to provide a data basis for subsequent classification analysis. Let V be the eigenvector matrix, whose column vectors are v1, v2…v 20 , then the principal component value matrix Y is:

[0114] Y=XV

[0115] We selected principal components Y1-Y7, whose cumulative contribution rates exceeded 90%, as representative feature parameters, mapping the high-dimensional data (20 dimensions) to a low-dimensional space (7 dimensions) to reduce computational complexity. These seven principal components can effectively reflect the main characteristics of the original data.

[0116] Table 1 Principal component analysis results

[0117] characteristic parameters Eigenvalue Contribution rate Cumulative contribution rate Y1 6.7510 0.3375 0.3375 Y2 4.3617 0.2181 0.5556 Y3 3.4031 0.1702 0.7258 Y4 1.3576 0.0679 0.7937 Y5 1.0253 0.0513 0.8449 Y6 0.8365 0.0418 0.8868 Y7 0.6178 0.0309 0.9177 … … … … Y19 eps eps 1 Y20 eps eps 1

[0118] Step 2: K-means cluster analysis: The characteristic data of the sample data set are used as input sample points for K-means clustering, and the known working conditions are divided into three modes and divided into training sets and test sets.

[0119] Taking into account the difficulty of identification and classification accuracy, this paper uses K-means cluster analysis to divide the working conditions into three modes. K-means clustering is an unsupervised learning algorithm based on partitioning. The specific steps are as follows:

[0120] (1) Initialize the center of mass:

[0121] First, three initial centroids are randomly selected and recorded as μ1, μ2, and μ3.

[0122] (2) Assign sample points to the nearest cluster:

[0123] For each sample point x i (i=1, 2, ..., n) calculates the Euclidean distance between it and each centroid and assigns the sample point to the cluster with the closest distance. The specific formula is:

[0124]

[0125] Among them, x ij is the sample point x i The jth eigenvalue of kj is the center of mass μ k The j-th eigenvalue of .

[0126] The sample point x i Assign to the closest cluster C k :

[0127]

[0128] (3) Update the centroid:

[0129] For each cluster C k , recalculate the center of mass μ k is the mean of all sample points in the cluster:

[0130]

[0131] (4) Iterative optimization:

[0132] Repeat steps (2) and (3). During each iteration, the algorithm recalculates the centroid of each cluster until the centroid position converges.

[0133] (5) Iteration objective: The objective function is minimized:

[0134] The goal of the K-means clustering algorithm is to minimize the sum of squares of the sample points to the centroid within the cluster, that is:

[0135]

[0136] The K-means clustering algorithm uses characteristic data obtained from principal component analysis as input. Based on the dimensionality-reduced characteristic data from principal component analysis, the K-means clustering algorithm can more efficiently perform pattern segmentation, ultimately identifying three distinct cluster categories: C1, C2, and C3. C1 represents the low-speed mode of urban congestion; C3 represents the high-speed mode; and C2 represents the medium-speed mode.

[0137] The classification results are as follows Figure 2 shown. Figure 2The black crosses in the figure represent the center points of each category, or cluster centers. The locations of these centers reflect the average characteristic positions of the data points within each cluster. There are clear boundaries between the three clusters, indicating that the K-means clustering algorithm can effectively distinguish traffic conditions with different speed patterns when partitioning urban traffic data after principal component analysis dimensionality reduction, resulting in a relatively good clustering effect.

[0138] Step 3: Support vector machine model training: Input the training set sample data in the training set into the support vector machine model for training; based on the test set, the trained support vector machine model is tested by the cross-validation method to obtain the optimal parameters of the support vector machine model and obtain the trained support vector machine model.

[0139] Support vector machines achieve accurate classification by finding the optimal hyperplane to maximize the gap between different categories of data. They have the advantages of high computational efficiency, excellent generalization ability, and good robustness. The classification results obtained by K-means cluster analysis are divided into training set sample data and test set sample data, where the training set samples are:

[0140]

[0141] Where x i is a 7-dimensional feature vector, n=7; y i is the classification category; m is the number of samples.

[0142] The hyperplane of the support vector machine can be expressed by the following linear equation:

[0143] ω T x+b=0

[0144] By applying the soft margin optimization method, the minimum distance problem from the sample point to the hyperplane and the model misclassification problem can be transformed into a constrained optimization problem as follows:

[0145]

[0146] sty i (ω·x i +b)≥1-ξ i

[0147] In the formula, ω is the weight factor; b is the bias term; p is the adjustment parameter used to change the penalty form for misclassified points; C is the penalty parameter; ξ i is a slack variable; λ is a regularization parameter used to control the complexity of feature weights. k is the weight of the kth feature.

[0148] The sequential quadratic programming (SQP) method is used to solve this constrained optimization problem. The basic steps are as follows:

[0149] 1. Initialization: Select the initial point (ω (0) , b (0) ,ξ (0) ), set the number of iterations t = 0 and the convergence threshold ε.

[0150] 2. At the current point (ω (t) , b (t) ,ξ (t) ), the Lagrangian function of the constrained optimization problem is,

[0151]

[0152] Where, α i and β i are the Lagrange multipliers associated with the classification constraints and slack variables, respectively.

[0153] Calculate the gradient of the objective function and the constraints, that is, the gradient of the Lagrangian function with respect to the weight factor ω, the bias term b, and the slack variable ξ:

[0154]

[0155] Among them, ω (t) Represents the weight factor obtained at the tth iteration.

[0156] 3. Construct the quadratic programming subproblem,

[0157]

[0158] The constraints are:

[0159]

[0160] 4. Solve the quadratic programming subproblem by the active set method and obtain the search direction (Δω, Δb, Δξ).

[0161] 5. Perform line search along the search direction and update the current point.

[0162] ω (t+1) =ω (t) +αΔω

[0163] b (t+1) =b (t) +αΔb

[0164] ξ (t) =ξ (t) +αΔξ

[0165] Where α is the step size; Δω, Δb, Δξ are increments; b(t+1) 、ω (t+1) ,ξ (t+1) The next value of the iteration.

[0166] Check the convergence conditions. If the convergence conditions are met (the gradient is small enough or the step size is small enough), stop the iteration; otherwise, set t = t + 1 and continue the iteration. By iteratively solving a series of quadratic programming subproblems, the optimal solution to the optimization problem, that is, the optimal hyperplane, can be obtained.

[0167] Support vector machines use kernel functions to convert samples from the original feature space into a high-dimensional feature space. The choice of kernel function determines the basic performance of the support vector machine. Common kernel functions include linear kernel functions, polynomial kernel functions, Gaussian radial basis function kernel functions, Sigmoid kernel functions, etc. Among them, the radial basis function kernel function can map data to an infinite-dimensional space and is suitable for processing complex nonlinear problems. Therefore, the present invention uses the radial basis function kernel function to construct the classification function, and the radial basis function kernel function is:

[0168] K(x i ,x j )=exp(-γ||x i -x j || 2 )

[0169] In the formula, γ is the kernel function parameter, which is used to control the similarity between samples. i 、x j are two different eigenvectors in the sample.

[0170] The classification function is:

[0171]

[0172] sgn represents a sign function that returns a corresponding class label based on the input value. The output of the sign function sgn is used to determine which class the sample should be classified into. To prevent overfitting and prioritize smaller penalty parameters for similar accuracy to enhance model robustness, the present invention uses a cross-validation method to find the optimal penalty parameter C and kernel function parameter γ for the support vector machine model. A grid search is then performed to find the optimal combination within the specified parameter range.

[0173] ① Parameter grid generation

[0174] Generate a 51*51 two-dimensional parameter grid matrix, and set the range of the penalty parameter C and the kernel function parameter γ to:

[0175] C=2 x ,γ=2 y

[0176] Among them, the exponential parameters x,y∈{-10,-9.8,...,10}, the step size is 0.2.

[0177] ②Variable initialization

[0178] Create a 51*51 zero matrix and record the cross-validation accuracy of each parameter combination. The initial penalty factor is:

[0179] C=1

[0180] The initial kernel function parameters are:

[0181] γ=0.1

[0182] The initial best accuracy is:

[0183] bestacc=0

[0184] Set the minimum positive number ε=10 -4 , used to determine whether the accuracy difference is negligible.

[0185] ③Grid search and cross validation

[0186] Within the above parameter range, we systematically traverse all possible combinations of penalty parameters and kernel function parameters, and evaluate model performance through 5-fold cross-validation to find the optimal parameter pair. The training dataset is evenly divided into 5 subsets D1, D2, ..., D5. For each round s∈{1,2,...,5}, subset Ds is used as the validation set. The training dataset is:

[0187] D train =D1∪...∪D s-1 ∪D s+1 ∪...∪D5

[0188] Use the current parameters to train The SVM model is trained on the subset Ds and the classification accuracy of the SVM model is calculated:

[0189]

[0190] The accuracy of the five rounds of verification is averaged as the final performance indicator of the current parameter pair:

[0191]

[0192] ④Dynamically update the optimal parameters and determine the optimal values;

[0193] Compare the current accuracy with the previous one. If the accuracy of the current parameter combination is higher than the historical best value, update the optimal parameters to the current penalty parameter and kernel function parameters. Finally, use the optimal parameters to reintroduce the SVM model.

[0194] Based on the above steps, we first determine the penalty parameters and kernel parameters, use the training set samples S to train the support vector machine model, and then input the test set samples to evaluate the classification accuracy of the model. The test set pattern recognition results are as follows: Figure 3 As shown in the figure, the red portion represents the actual sample value, and the green portion represents the predicted sample value. Overlapping red and green indicates that the predicted result is consistent with the input data category. Experimental calculations show that the support vector machine model has an accuracy rate of 93.27%.

[0195] In order to verify the effectiveness of the support vector machine classification method, the present invention uses a mixed driving cycle of FTP and HWFIT for verification. The total duration of this cycle is 3240s, and the average speed is 28.96km / h. The recognition results of the driving cycle are as follows: Figure 4 The results show that the trained support vector machine model can effectively identify the three types of driving modes.

[0196] Step 4: Input the collected real-time working condition data into the support vector machine model trained in step 3 to obtain the identified category, and input the working condition speed of the identified category into the vehicle dynamics model to obtain the required power of the entire vehicle.

[0197] In the subsequent fuzzy logic control system, real-time power is a key input parameter. The real-time operating condition is directly identified by the previously trained support vector machine model. The dynamics model then uses the corresponding power demand to determine the identified categories 1, 2, and 3, which serve as inputs for the three subsequent fuzzy logic systems.

[0198] In order to accurately derive the vehicle power requirement based on the operating speed, this step constructs a vehicle dynamics model based on the vehicle's longitudinal dynamic characteristics. The details are as follows:

[0199] The rolling resistance of the vehicle is:

[0200] F r =C r ·m·g·cosθ

[0201] Where C r is the rolling resistance coefficient; m is the vehicle mass; g is the acceleration due to gravity; θ is the road slope.

[0202] The air resistance is:

[0203]

[0204] Where ρ is the air density; C d is the air resistance coefficient; A f is the effective frontal area of ​​the vehicle; v is the vehicle speed; v w is the real-time wind speed in the direction of vehicle movement.

[0205] The climbing resistance caused by the road slope is:

[0206] F g =m·g·sinθ

[0207] The traction force is:

[0208]

[0209] The vehicle model parameters and values ​​are shown in Table 2:

[0210] Table 2 Vehicle model parameters

[0211] parameter Physical meaning Value m / kg Vehicle quality 1600 <![CDATA[g / (m / s 2 )]]> Gravity 9.81 <![CDATA[C r ]]> Rolling resistance coefficient 0.02 <![CDATA[C d ]]> Air resistance coefficient 0.5 <![CDATA[ρ / (kg / m 3 )]]> Air density 1.225

[0212] The required power of the vehicle is related to the traction and vehicle speed, and its calculation formula is:

[0213] P d =F t ·v

[0214] The vehicle's required power P d It is provided by proton exchange membrane fuel cells and lithium-ion batteries, and the relationship is as follows:

[0215] P d =η DC / DC ·P fc +P li

[0216] Where η DC / DC is the conversion efficiency of the DC / DC converter, which is 0.9; P fc is the output power of the proton exchange membrane fuel cell; P li is the output power of the lithium-ion battery. In summary, the required power of the vehicle is obtained from the operating speed.

[0217] Step 5: Input the obtained vehicle demand power and the SOC of the lithium-ion battery into the fuzzy controller to obtain the proton exchange membrane fuel cell power and realize the power distribution of the hybrid energy storage system.

[0218] The fuzzy logic control system uses the vehicle dynamics model to obtain the vehicle demand power P d The SOC of the lithium-ion battery is used as the input of the fuzzy controller, and the power of the proton exchange membrane fuel cell is used as the output of the fuzzy controller. First, a judge is established and the vehicle demand power P is calculated based on the classification results of the above support vector machine model. dThe input is then divided and fed into three different fuzzy logic control systems. The fuzzifiers then convert the input variables into fuzzy values. Fuzzy inference then generates the output, which is then sent to the defuzzifiers. Finally, defuzzification yields the precise value of the proton exchange membrane fuel cell power. The fuzzy inference process consists of two core components: membership functions for the input and output variables, and fuzzy logic rules constructed based on these membership functions. The discriminator detects categories 1, 2, and 3 and transmits the vehicle's required power to the three fuzzifiers, respectively. The SOC is calculated using real-time data such as the lithium-ion battery power.

[0219] (1) Membership function

[0220] Input variables: SOC of lithium-ion battery and vehicle power requirement P d The membership functions are as follows: Figure 5 and Figure 6 As shown. In fuzzy logic control, the SOC of lithium-ion batteries has three membership functions, representing low (L), medium (M), and high (H), which are trapezoidal, triangular, and trapezoidal membership functions respectively. The vehicle demand power P d It has five input membership functions, namely very low (VL), low (L), medium (M), high (H) and very high (VH), which are the membership functions of trapezoid, trapezoid, triangle, triangle and trapezoid respectively.

[0221] Output variable PEMFC power P fc The membership function is as follows Figure 7 As shown, the membership functions are divided into near zero (Z), positive small (PS), positive middle (PM), positive large (PL), relatively large (PB), and maximum (PV). The first five are triangular and the sixth is trapezoidal.

[0222] (2) Fuzzy logic control rules

[0223] The present invention designs different sets of fuzzy control rules for low-speed, medium-speed, and high-speed driving modes, respectively, to achieve precise control in different modes. Based on relevant experience, the fuzzy rule design is shown in Table 3. When the SOC is low, regardless of the power demand of the entire vehicle, the PEMFC will simultaneously power the load and the lithium-ion battery; when the SOC is medium, depending on the operating conditions, the PEMFC and the lithium-ion battery will meet the power demand in different output ratios; when the SOC is high, the lithium-ion battery alone is used for power supply in low-speed mode, and the PEMFC and lithium-ion battery are used together for power supply in medium- and high-speed modes.

[0224] Table 3 Fuzzy rules under three working conditions

[0225]

[0226] The fuzzy logic control system obtains the proton exchange membrane fuel cell power in the corresponding mode at each moment according to the above membership function and the set fuzzy rules, thereby realizing the power distribution of the hybrid energy storage system.

[0227] According to the above steps, the beneficial effect of the SVM-FLC energy management strategy of the present invention is verified under the FTP-HWFIT mixed driving condition. Figure 8 and Figure 9 As shown. Figure 8 and Figure 9 As can be seen, in the hybrid energy storage system, the PEM fuel cell serves as the primary power source, providing relatively stable power to the load, while the lithium-ion battery serves as an auxiliary power source, providing peak power. Furthermore, the support vector machine-fuzzy logic control system's energy management strategy dynamically adjusts the power distribution between the lithium-ion battery and the PEM fuel cell based on pattern recognition results, thereby eliminating the impact of complex operating conditions on power distribution. Simultaneously, the lithium-ion battery performs peak-shaving and valley-filling functions, avoiding high-current discharge in the PEM fuel cell and protecting it from interference from high-frequency currents and frequent load changes.

[0228] Comparing the control strategy of support vector machine-fuzzy logic control system with constant power control strategy and power following control strategy, the SOC curve of lithium-ion battery is shown as follows: Figure 10 shown.

[0229] Depend on Figure 10 It can be seen that the constant power strategy allows the fuel cell power to be maintained steadily at the rated value, thereby reducing the impact of power fluctuations on the fuel cell. The power following strategy allows the proton exchange membrane fuel cell power to change closely with the changes in the power demand of the operating conditions, serving as the main power source to cope with changes in power demand. The SOC fluctuation range under the constant power strategy is approximately 0.12, and the SOC fluctuation range under the power following strategy is approximately 0.21. In comparison, the fuel cell power under the support vector machine-fuzzy logic control system strategy of the present invention has a smaller SOC fluctuation range (0.06) while responding to changes in the operating power, and the loss of the lithium-ion battery will be less, thereby improving the life and economy of the hybrid energy storage system.

[0230] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fuel cell hybrid energy storage system adaptive energy management method, characterized in that: The steps are as follows: Step 1: Perform principal component analysis on the working condition data under known working conditions, select multiple characteristic parameters to reflect the working condition data, and obtain a sample data set; Step 2: K-means cluster analysis: Use the characteristic data of the sample data set as input sample points for K-means clustering, divide the known working conditions into three modes, and divide them into training sets and test sets; Step 3: Input the training set sample data in the training set into the support vector machine model for training; based on the test set, test the trained support vector machine model through the cross-validation method to obtain the optimal parameters of the support vector machine model and obtain the trained support vector machine model; Step 4: Input the collected real-time operating condition data into the support vector machine model trained in step 3 to obtain the identified category, and input the operating condition speed of the identified category into the vehicle dynamics model to obtain the required power of the entire vehicle; Step 5: Input the obtained vehicle demand power and the SOC of the lithium-ion battery into the fuzzy controller to obtain the proton exchange membrane fuel cell power and realize the power distribution of the hybrid energy storage system.

2. The adaptive energy management method for a fuel cell hybrid energy storage system according to claim 1, characterized in that: The operating condition data is standardized before principal component analysis. The method for performing feature screening on the operating condition data by principal component analysis is as follows: (1) Constructing characteristic parameter matrix: Four working conditions, FTP, HWFIT, NYCC, and CLTP, were selected for research. The data were divided into multiple segments by fixed step size identification, and the time interval of each segment was 15s. The characteristic parameter matrix X was constructed based on the extracted working condition data, which includes 20 characteristic parameters: acceleration, maximum acceleration, maximum deceleration, acceleration time ratio, deceleration time ratio, uniform speed time ratio, parking time ratio, 0-20 speed time ratio, 20-40 speed time ratio, 40-60 speed time ratio, 60-80 speed time ratio, maximum speed, average speed, average speed excluding parking time, speed standard deviation, acceleration standard deviation, average acceleration in acceleration section, average deceleration in deceleration section, acceleration standard deviation in acceleration process, deceleration standard deviation in deceleration process, and driving distance. (2) Calculate the sample correlation coefficient matrix C directly from the characteristic parameter matrix X; (3) Calculate the eigenvalues ​​and eigenvectors of 20 characteristic parameters based on the sample correlation coefficient matrix C; (4) Calculate the contribution rate and cumulative contribution rate of each principal component based on the eigenvalue; (5) Construct principal component values: select principal components with cumulative contribution rates of more than 90% as representative characteristic parameters.

3. The adaptive energy management method for a fuel cell hybrid energy storage system according to claim 1 or 2, characterized in that: The implementation method of the K-means clustering is: (1) Randomly select three initial centroids, denoted as μ1, μ2, and μ3; (2) Assign sample points to the nearest cluster: Calculate each sample point x i , i = 1, 2, ..., n and the Euclidean distance of each centroid, and assign the sample point to the cluster C with the closest distance k ; (3) Update the centroid: For each cluster C k , recalculate the center of mass μ k is the mean of all sample points in the cluster; (4) Iterative optimization: Repeat steps (2) and (3) to recalculate the centroid of each cluster until the centroid position converges; (5) Objective function minimization: Minimize the sum of squares from the sample points within the cluster to the centroid: The K-means clustering algorithm uses the characteristic data obtained from principal component analysis as input sample points and divides the data into three different cluster categories: C1, C2, and C3. C1 represents the low-speed mode of urban congestion; C3 represents the high-speed mode; and C2 represents the medium-speed mode.

4. The adaptive energy management method for a fuel cell hybrid energy storage system according to claim 3, characterized in that: The training set sample data is: Where x i is an n-dimensional feature vector; y i is the classification category, m is the number of samples; The hyperplane of the support vector machine is obtained by the linear equation ω T x+b=0 to represent; By applying the soft margin optimization method, the minimum distance problem from the sample point to the hyperplane and the model misclassification problem are transformed into a constrained optimization problem: In the formula, ω is the weight factor; b is the bias term; p is the adjustment parameter used to change the penalty form for misclassified points; C is the penalty parameter; ξ i is a slack variable; λ is a regularization parameter used to control the complexity of feature weights, ω k is the weight of the kth feature. The constrained optimization problem is solved using the sequential quadratic programming method, and the optimal solution obtained is the optimal hyperplane; the support vector machine uses the kernel function to transform the samples from the original feature space into a high-dimensional feature space; the penalty parameter of the support vector machine model is found through the cross-validation method, and the optimal combination of the optimal penalty parameter C and the kernel function parameter γ is found within the specified parameter range through grid search.

5. The adaptive energy management method for a fuel cell hybrid energy storage system according to claim 4, characterized in that: The steps for solving the constrained optimization problem using the sequential quadratic programming method are as follows: 1). Initialization: Select the initial point (ω (0) , b (0) ,ξ (0) ), set the number of iterations t = 0 and the convergence threshold ε; 2). At the current point (ω (t) , b (t) ,ξ (t) ), the Lagrangian function of the constrained optimization problem is: Where, α i and β i are the Lagrange multipliers associated with the classification constraints and slack variables, respectively; Calculate the gradient of the objective function and the constraints, that is, the gradient of the Lagrangian function with respect to the weight factor ω, the bias term b, and the slack variable ξ: Among them, ω (t) represents the weight factor obtained at the tth iteration; 3) Construct the quadratic programming sub-problem: The constraints are: 4) Solve the quadratic programming subproblem using the active set method and obtain the search direction (Δω, Δb, Δξ); 5) Perform line search along the search direction and update the current point: oh (t+1) =ω (t) +aDo b (t+1) =b (t) +αΔb x (t) =ξ (t) +αΔξ Where α is the step size; Δω, Δb, Δξ are increments; b (t+1) 、ω (t+1) ,ξ (t+1) is the next step value of the iteration; Check the convergence condition: If the convergence condition is met, stop the iteration; otherwise, set the number of iterations t = t + 1 and continue the iteration.

6. The adaptive energy management method for a fuel cell hybrid energy storage system according to claim 4 or 5, characterized in that: The radial basis function kernel function is used to construct the classification function. The radial basis function kernel function is: K(x i ,x j )=exp(-γ||x i -x j || 2 ) Where γ is the kernel function parameter, x i 、x j are two different eigenvectors in the sample; The classification function is: Here, sgn represents the sign function.

7. The adaptive energy management method for a fuel cell hybrid energy storage system according to claim 6, characterized in that: The method for finding the optimal combination of the optimal penalty parameter C and the kernel function parameter γ is: ① Parameter grid generation: Generate a 51*51 two-dimensional parameter grid matrix, and set the range of penalty parameter C and kernel function parameter γ to: C=2 x ,γ=2 y ;Wherein, exponential parameters x,y∈{-10,-9.8,...,10}, step size 0.2; ② Variable initialization: Create a 51*51 zero matrix, record the cross-validation accuracy of each parameter combination, the initial penalty factor is: C=1, the initial kernel function parameter is: γ=0.1, the initial best accuracy is: bestacc=0; set the minimum positive number ε=10 -4 ; ③ Grid search and cross-validation: Traverse all possible combinations of penalty parameters and kernel function parameters, evaluate model performance through 5-fold cross-validation to find the optimal parameter pair: divide the training dataset evenly into 5 subsets D1, D2, ..., D5. For each round s∈{1,2,...,5}, use subset Ds as the validation set. The training dataset is: D train =D1∪...∪D s-1 ∪D s+1 ∪...∪D5, use the current parameters to train the dataset D train The SVM model is trained on the subset Ds and the classification accuracy of the SVM model is calculated: The accuracy of the five rounds of verification is averaged as the final performance indicator of the current parameter pair: ④ Dynamically update the optimal parameters and determine the optimal values: Compare the current accuracy with the previous accuracy. If the accuracy of the current parameter combination is higher than the historical best value, update the optimal parameters to the current penalty parameters and kernel function parameters, and finally use the optimal parameters to reintroduce the SVM model.

8. The adaptive energy management method for a fuel cell hybrid energy storage system according to claim 7, characterized in that: The vehicle dynamics model is constructed based on the vehicle longitudinal dynamics characteristics, as follows: The rolling resistance of the vehicle is: F r =C r ·m·g·cosθ; where C r is the rolling resistance coefficient; m is the vehicle mass; g is the acceleration of gravity; θ is the road slope; The air resistance is: Where ρ is the air density; C d is the air resistance coefficient; A f is the effective frontal area of ​​the vehicle; v is the vehicle speed; v w is the real-time wind speed in the direction of vehicle movement; The climbing resistance caused by the road slope is: F g =m·g·sinθ; The traction force is: The required power of the vehicle is related to the traction and vehicle speed. The calculation formula is: P d =F t ·v.

9. The adaptive energy management method for a fuel cell hybrid energy storage system according to any one of claims 4, 5, 7, and 8, characterized in that: The implementation method of the fuzzy controller in step 5 is: establish a judger and classify the vehicle demand power P according to the classification result of the support vector machine model. d The data is split and input into three different fuzzy logic control systems. The fuzzifier converts the input variables into fuzzy values, and the output is obtained through fuzzy reasoning and sent to the defuzzifier. The precise value of the proton exchange membrane fuel cell power is obtained through defuzzification. The fuzzy reasoning process includes two core components: one is the membership function of the input and output variables, and the other is the fuzzy logic rules constructed based on the membership function; the judge transmits the vehicle's required power to three fuzzy detectors respectively by detecting categories 1, 2, and 3.

10. The adaptive energy management method for a fuel cell hybrid energy storage system according to claim 9, characterized in that: In fuzzy logic control, the SOC of lithium-ion batteries has three membership functions, representing low L, medium M, and high H, which are trapezoidal, triangular, and trapezoidal membership functions respectively; the vehicle demand power P d It has five input membership functions, namely, very low VL, low L, medium M, high H, and very high VH, which are the membership functions of trapezoid, trapezoid, triangle, triangle, and trapezoid respectively; Output variable PEMFC power P fc The membership functions are divided into nearly zero Z, positive small PS, medium PM, positive large PL, large PB, and maximum PV. The first five are triangular membership functions, and the sixth is a trapezoidal membership function. When the SOC of the lithium-ion battery is low, the PEMFC will simultaneously power the load and the lithium-ion battery, regardless of the power demand of the vehicle. When the SOC of the lithium-ion battery is medium, the PEMFC and the lithium-ion battery will meet the power demand in different output ratios depending on the operating conditions. When the SOC of the lithium-ion battery is high, the lithium-ion battery alone will provide power in low-speed mode, and the PEMFC and lithium-ion battery will provide power together in medium- and high-speed modes. Different fuzzy rules are designed for the three driving modes of low speed, medium speed and high speed: The fuzzy logic control system derives the proton exchange membrane fuel cell power under the corresponding driving mode at each moment according to the membership function and the set fuzzy rules, thereby realizing the power distribution of the hybrid energy storage system.

Images