An adaptive configuration method and system for fuel cell vehicle energy management parameters

Abstract

The application discloses a kind of adaptive configuration method and system of fuel cell automobile energy management parameter, belong to fuel cell hybrid vehicle control technical field.Method includes: collecting driving segment and extracting operating condition characteristic parameter, clustering obtains operating condition category and in-class mahalanobis radius;For each class, solve the pareto optimal solution set with equivalent hydrogen consumption and power fluctuation rate as target;Mahalanobis distance of current characteristic and clustering center is calculated when online running, according to the relationship between distance and in-class mahalanobis radius, soft boundary processing is executed, and the corresponding pareto solution set is obtained;Then, according to the current state of charge deviation, a solution is dynamically selected from the solution set, and the output fuel cell power of fuzzy controller is configured.The soft boundary mechanism driven by mahalanobis distance avoids operating condition boundary parameter jump, and the adaptive trade-off of economy and life protection is realized by state of charge guided parameter selection.The application also introduces incremental learning based on matching confidence, so that the system can continuously evolve.

Description

Technical Field

[0001] This invention belongs to the field of fuel cell hybrid vehicle control technology, specifically relating to an adaptive configuration method and system for energy management parameters of fuel cell vehicles. Background Technology

[0002] Fuel cell vehicles typically use a hybrid power supply of fuel cells and batteries. The energy management strategy determines the power distribution between the two energy sources, which directly affects the vehicle's fuel economy, the battery's ability to maintain its state of charge, and the lifespan of the fuel cell system.

[0003] To address the complex and ever-changing operating conditions in real-world driving, existing technologies have proposed various energy management methods that combine operating condition identification with controller parameter optimization. The basic idea behind these methods is as follows: In the offline phase, typical driving modes are extracted from historical operating condition data, and controller parameters are optimized offline for different modes, establishing a parameter library; in the online phase, real-time vehicle driving data is collected through sensors to identify the current driving state, and the corresponding control parameters are retrieved from the parameter library to configure the controller and achieve power distribution. These methods have achieved certain results in improving adaptability to different operating conditions.

[0004] However, existing technologies still have the following common problems: First, offline optimization results are limited and lack the flexibility of online optimization. Existing methods typically employ genetic algorithms, dynamic programming, and other optimization techniques to solve for the optimal control parameters under each type of operating condition, but the optimization results often only retain a compromise solution. Since the vehicle's state constantly changes during actual driving, a single parameter solution cannot meet the differentiated needs of different scenarios.

[0005] Second, the online controller parameters are unrelated to the battery's state of charge (SOC) and cannot respond to changes in SOC. Current methods select parameters based solely on operating condition category or membership degree during the online phase, failing to incorporate the battery's current SOC into the parameter selection feedback loop. When the SOC is low, priority should be given to reducing hydrogen consumption to replenish the battery; when the SOC is high, priority should be given to reducing power fluctuations to extend the stack's lifespan. Fixed parameters cannot simultaneously meet both requirements.

[0006] Third, there is a lack of smooth transition mechanisms at the boundaries of operating condition categories. Existing methods use discrete hard classification to identify operating conditions. When the actual driving condition characteristics are located in the boundary region of a predefined category, the identification results repeatedly jump between adjacent categories, causing frequent switching of controller parameters and discontinuous oscillations in power output.

[0007] The core of the above problem is that existing methods fail to recognize the differences in optimal control objectives corresponding to different states of charge under the same operating conditions, and also fail to provide a technical solution that can retain the diversity of offline optimization while adaptively adjusting online according to the state of charge.

[0008] Therefore, there is an urgent need in this field for an adaptive configuration method and system for energy management parameters that can solve the above problems. Summary of the Invention

[0009] This invention aims to provide an adaptive configuration method and system for energy management parameters of fuel cell vehicles, in order to solve the technical problems of abrupt changes in operating condition boundary parameters and the independence of controller parameters on state of charge in the prior art. To achieve the above objectives, the present invention provides an adaptive configuration method for energy management parameters of a fuel cell vehicle, comprising the following steps: Step S1: Collect multiple driving segments, extract the working condition feature parameters of each driving segment, use a clustering algorithm to classify similar segments into the same working condition category, and record the cluster center and intra-cluster Mahalanobis radius of each working condition category; Step S2: For each operating condition category, with equivalent hydrogen consumption and fuel cell power fluctuation rate as dual optimization objectives, a multi-objective optimization algorithm is used to solve for the Pareto optimal solution set for that category. Each solution corresponds to a set of fuzzy controller parameters. Step S3: During online operation, vehicle driving data is collected in real time, current operating condition feature parameters are extracted, the Mahalanobis distance between the current feature vector and each cluster center is calculated, and soft boundary processing is performed based on the relationship between the Mahalanobis distance and the Mahalanobis radius within the corresponding cluster to obtain the currently available Pareto optimal solution set. Step S4: Obtain the deviation between the current state of charge and the target state of charge, and select a solution from the obtained Pareto optimal solution set according to the deviation. When the deviation indicates that the state of charge is low, the solution with smaller equivalent hydrogen consumption is selected, and when the deviation indicates that the state of charge is high, the solution with smaller power fluctuation rate is selected. Step S5: Configure the fuzzy controller using the fuzzy controller parameters corresponding to the selected solution, and output the target power of the fuel cell.

[0010] Furthermore, the soft boundary processing includes: When there exists a unique class such that the Mahalanobis distance is less than the Mahalanobis radius within that class, obtain the Pareto optimal solution set for that class; When there are multiple categories such that the Mahalanobis distance is less than the Mahalanobis radius within each category, obtain the union of the Pareto optimal solution sets of these categories; When the Mahalanobis distance of all classes is greater than or equal to the Mahalanobis radius within the corresponding class, obtain a weighted combination of the Pareto optimal solution sets of at least the two classes with the closest Mahalanobis distance.

[0011] Furthermore, in step S3, a sliding window of fixed duration is used when extracting the current working condition feature parameters. The duration of the sliding window is 120 seconds, and it slides once every 1 second, moving in the latest 1 second data and moving out the earliest 1 second data each time.

[0012] Furthermore, the formula for calculating the Mahalanobis distance is: ; in, For the current feature vector, The cluster center of class 1, For the first j The covariance matrix of the class; the in-class Mahalanobis radius is the first... All training samples in the class are grouped to cluster centers. The maximum value of the Mahalanobis distance.

[0013] Furthermore, the method for selecting a solution based on the deviation in step S4 is as follows: First, construct the dynamic weighting coefficient α = 0.5 + k*ΔSOC, where ΔSOC=SOC target -SOC current k is the calibration coefficient; Secondly, calculate the normalized equivalent hydrogen consumption H for each solution in the current Pareto optimal solution set. norm and normalized power volatility F norm : , ; in , These are the maximum and minimum equivalent hydrogen consumption values ​​in the current solution set, respectively. , These are the maximum and minimum power volatility values ​​in the current solution set, respectively. Finally, the solution that minimizes the overall score S = α * Hnorm + (1 − α) * Fnorm is selected.

[0014] Furthermore, it also includes: correcting the dynamic weighting coefficients according to the currently identified operating condition category, wherein the economic weight is increased under high-speed cruising conditions and the lifespan weight is increased under low-speed congestion conditions.

[0015] Furthermore, it also includes: identifying the current driving style and correcting the dynamic weighting coefficient, wherein an aggressive driving style reduces the fuel economy weight and a conservative driving style increases the fuel economy weight.

[0016] Furthermore, the fuzzy controller has a dual-input single-output structure, with the input being the state-of-charge error and the required power, and the output being the fuel cell power; the fuzzy controller parameters include the input quantization factor, the output scaling factor, and the peak position of the membership function.

[0017] Furthermore, it also includes: During online operation, the ratio of the minimum Mahalanobis distance to the corresponding intra-class Mahalanobis radius is recorded as a confidence index for each working condition matching. When the confidence level is lower than the preset threshold for several consecutive periods, the current feature vector and the actual optimal control parameters in the subsequent time window will be used as new samples. The newly added samples are used to perform incremental clustering updates on the offline knowledge base.

[0018] The present invention also provides an adaptive configuration system for energy management parameters of a fuel cell vehicle, comprising: An offline build module is used to perform steps S1 and S2 as described in claim 1; An online matching module is used to perform step S3 as described in claim 1; The parameter selection module is used to perform step S4 as described in claim 1; A control execution module is used to execute step S5 as described in claim 1; Compared with the prior art, the present invention has the following beneficial effects: By employing a soft boundary processing mechanism, Pareto optimal solution sets of weighted combinations or unions are output at the operating condition boundaries, thus avoiding parameter jumps and power oscillations.

[0019] By using a sliding window of fixed duration for extracting operating condition features, the contradiction between single-point sampling being susceptible to noise interference and excessively long windows causing response delays is resolved. The window duration is set to 120 seconds after a trade-off, which ensures the statistical significance of features while enabling rapid parameter response after changes in operating conditions.

[0020] Third, by selecting parameters guided by the state of charge, the controller parameters change continuously with the battery's state of charge. When the battery charge is low, hydrogen consumption is reduced first, and when the battery charge is high, power fluctuations are reduced first, thus achieving adaptive adjustment of the control objective.

[0021] Fourth, by introducing operating condition category correction and driving style correction, the energy management strategy can adapt to both driving scenarios and driver habits, achieving multi-dimensional personalized optimization.

[0022] Fifth, by introducing confidence monitoring and incremental learning mechanisms, the system can adaptively update the knowledge base during long-term operation without manual intervention, thus improving the system's continuous adaptability. Attached Figure Description

[0023] Figure 1 This is a flowchart of an adaptive configuration method for energy management parameters of a fuel cell vehicle according to the present invention.

[0024] Figure 2 This is a schematic diagram of an adaptive configuration system for energy management parameters of a fuel cell vehicle according to the present invention. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention are clearly and completely described in the following description. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1: To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment takes a fuel cell logistics vehicle as an example.

[0028] like Figure 1 As shown, the method of the present invention is divided into two main stages: offline knowledge base construction and online operation. The offline stage of offline knowledge base construction includes steps S1 and S2; the online operation stage includes steps S3, S4, and S5.

[0029] Step S1: Collect driving segments and extract features to obtain the category and Mahalanobis radius of the class. Multiple driving segments were selected from a standard operating condition database and real-vehicle data, each with a fixed duration. Seven characteristic parameters were extracted from each segment: average vehicle speed, maximum vehicle speed, idle speed ratio, average acceleration, average deceleration, speed standard deviation, and power change rate per unit time. These parameters characterize the macroscopic statistical properties and dynamic response features of the driving conditions from different dimensions.

[0030] Principal component analysis was used to reduce the seven-dimensional features to three dimensions. A clustering algorithm (such as K-means++) was then used to cluster the three-dimensional features of all samples. The optimal number of clusters was determined using the silhouette coefficient, resulting in multiple work condition categories. The cluster centers of each category were recorded. C j The covariance matrix Σ j Then, the cluster center is calculated for each training sample in that class. C j The Mahalanobis distance is calculated, and the maximum value is taken as the intra-class Mahalanobis radius of that category. : ; This radius reflects the natural distribution range of this type of operating condition in the feature space.

[0031] Step S2: Solve for the Pareto optimal solution set Multi-objective optimization is performed separately for each operating condition category. The optimization variables are configurable parameters of the fuzzy controller, including: the state-of-charge error quantization factor. K e Demand power quantization factor K p Output scaling factor K u And the peak position parameters of seven fuzzy sets (negative large, negative medium, negative small, zero, positive small, positive medium, positive large).

[0032] The two optimization objectives are defined as follows: Objective 1: Equivalent hydrogen consumption

[0033] in: Let λ be the hydrogen consumption rate of the fuel cell, and λ be the equivalence factor. P bat ( t () represents battery power (positive for discharging and negative for charging). Hydrogen has a low calorific value. T This refers to the duration of operation.

[0034] Objective 2: Fuel cell power volatility

[0035] in, for t fuel cell power at any given time This represents the average power of the fuel cell within this operating condition segment.

[0036] A multi-objective genetic algorithm (such as NSGA-II) is used to solve the problem, obtaining a Pareto optimal solution set for each working condition category. The cluster centers are then... C j Intraclass Markovial radius Covariance matrix Σ j It is associated with and stored in the corresponding Pareto optimal solution set to form an offline knowledge base.

[0037] Online operating condition matching and soft boundary handling Step S3: Calculate the Mahalanobis distance, perform soft boundary processing, and obtain the currently available Pareto optimal solution set.

[0038] After offline construction is complete, the system enters the online operation phase. During vehicle operation, the controller maintains a fixed-duration sliding window and periodically performs condition matching updates. In each update cycle, the same seven feature parameters as in the offline phase are extracted from the window data and subjected to the same principal component transformation to obtain a three-dimensional feature vector. X cur .

[0039] It should be noted that the duration of the sliding window is determined based on a trade-off between the accuracy of operating condition recognition and the real-time response. If the window duration is too short, the amount of driving data contained within the window will be insufficient, the extracted statistical feature parameters will lack statistical significance, and they will be susceptible to transient noise interference, leading to misclassification of the operating condition category. If the window duration is too long, the window will contain features from multiple different driving stages, and the features will be over-smoothed and lose their timeliness. When the actual driving condition changes, the recognition result will lag behind the actual change in operating condition, resulting in a delay in the update of controller parameters. In this embodiment, the duration of the sliding window is set to 120 seconds. This duration can ensure the effectiveness of feature statistics while enabling parameter updates to be completed within 30 seconds after the operating condition change, balancing recognition accuracy and response speed. The sliding window is updated by sliding once every 1 second, moving in 1 second of newly collected data each time and moving out the oldest 1 second of data, keeping the window duration constant.

[0040] Calculate the Mahalanobis distance between the current feature vector and each cluster center:

[0041] Let the first j The class matching threshold is ,in η The relaxation coefficient is determined through offline cross-validation. Based on the relationship between Mahalanobis distance and the threshold, the following three cases are handled.

[0042] Scenario 1 (Normal Identification): A unique one exists. j Make d j < T j At this point, the current operating condition is determined to be the [number missing]. j The system outputs the complete Pareto optimal solution set for that class. This is the most common scenario in system design.

[0043] Scenario 2 (Overlapping Boundaries): Multiple j satisfy d j < T jIn real-world vehicle operation, when the vehicle's driving characteristics fall precisely at the boundary between two known operating condition categories, a hard classification strategy can lead to minute measurement noise or natural fluctuations in the feature parameters causing the category judgment to change repeatedly within seconds. This frequent switching causes the controller parameters to jump accordingly, resulting in oscillations in the fuel cell power command. To suppress this phenomenon, this method does not force the selection of a category when boundaries overlap, but instead outputs the union of the Pareto optimal solution sets corresponding to these categories. Since the union contains the optimization parameters for each boundary category, the subsequent parameter selection module can find a smoothly transitioning solution from a richer candidate set, thus ensuring that the final output power command remains continuous in the boundary region.

[0044] Scenario 3 (Outside the known range): All d j≥ T j At this point, the current operating condition is not similar enough to any of the categories in the offline knowledge base. If the nearest category is still used, parameter mismatch may lead to a decrease in fuel economy or fuel cell lifespan. This method outputs a Pareto optimal solution set of several categories with the closest Mahalanobis distance and performs a weighted combination. Regarding the number of nearest categories, this embodiment selects two. If only one category is selected, it degenerates into hard classification, and the parameters will change abruptly when the operating condition is between two categories, failing to eliminate the aforementioned power oscillations; if three or more categories are selected, although the smoothing effect is slightly improved, the weights are distributed to more solutions, increasing the size of the merged solution set, which significantly increases the complexity of calculating the comprehensive score in the subsequent step S4, while the actual improvement in control effect is negligible. Therefore, selecting two categories can ensure the continuity of parameter transition while taking into account computational efficiency. In specific operation, the Pareto optimal solution set of the two categories with the closest Mahalanobis distance is obtained, and all solutions in these two solution sets are merged, with each solution having a weight. The weighted design ensures that categories that are closer in distance contribute more. This way, as the vehicle gradually enters an unknown operating area, the control parameters do not suddenly jump to a completely unfamiliar configuration. Instead, they borrow parameter characteristics from the two most similar known categories in a weighted manner, achieving a smooth transition.

[0045] The above three scenarios together constitute the soft boundary processing mechanism. Compared with traditional hard classification, this mechanism no longer generates discrete class jumps near the operating boundary, but instead outputs a continuously changing set of candidate parameters, thereby eliminating power oscillations caused by parameter switching.

[0046] Step S4: Obtain the SOC deviation and dynamically select a solution.

[0047] After obtaining the currently available Pareto optimal solution set, a specific solution needs to be selected to configure the controller. The traditional approach is to fix the midpoint of the Pareto front, but this cannot adapt to changes in the battery's state of charge. For example, when the battery charge is low, if a compromise parameter is still used, it may not be able to replenish the energy in time, causing the charge to drop further or even triggering undervoltage protection; while when the battery charge is high, if the emphasis is still on reducing hydrogen consumption, the fuel cell may need to frequently and significantly adjust the power to keep up with demand, which will accelerate the degradation of the stack performance.

[0048] To address this issue, this step introduces a state-of-charge (SOC)-guided selection mechanism. Let the target SOC be... SOC target Current state of charge SOC current Calculate the deviation: Δ SOC = SOC target - SOC current Constructing dynamic weight coefficients: α=0.5+ k *ΔSOC coefficient k Calibration was performed via offline simulation, limited to the [0,1] interval. For each solution in the current Pareto optimal solution set, the normalized equivalent hydrogen consumption was calculated. H norm and normalized power volatility F norm Let the solution set be... J 1 The maximum value is The minimum value is ,but:

[0049] J 1 The maximum value is The minimum value is Similarly, calculate F. norm，

[0050] Calculate the overall score for each solution:

[0051] Observing the formula, we can see that when ΔSOC>0 (low battery level), α>0.5, and the economic item in the overall score will be... H norm The weight of α increases; when ΔSOC < 0 (high power consumption), α < 0.5, and the power fluctuation term F normThe weight of increases.

[0052] Select the solution that minimizes the overall score S to configure the controller.

[0053] If the current solution set comes from the weighted merging of case three, then the final score for each solution is: Select the solution with the smallest final score.

[0054] Since α varies with Δ SOC As the battery level changes continuously, the position of the selected solution on the Pareto front also shifts continuously: when the battery level is low, it tends towards a region with better economic efficiency (i.e., lower hydrogen consumption), and when the battery level is high, it tends towards a region with smaller power fluctuations. This allows the same offline parameter library to automatically adjust the control bias according to the real-time status of the battery level, without the need to calibrate parameters separately for different battery levels.

[0055] To further enhance adaptability, this embodiment also modifies α based on operating conditions and driving style. During high-speed cruising, the vehicle speed is stable, and power demand changes gradually, naturally placing the fuel cell in a relatively stable operating state. In this case, appropriately increasing the economy weight can further reduce hydrogen consumption without significantly increasing power fluctuations. In low-speed congested conditions, frequent starts and stops cause drastic fluctuations in power demand. If economy is still prioritized, the fuel cell will frequently undergo large load changes, accelerating its degradation. Therefore, appropriately increasing the lifespan weight (i.e., decreasing α) is more reasonable. Similarly, under aggressive driving styles, power demand changes drastically, so the economy weight should be appropriately reduced to alleviate the pressure on the fuel cell to keep up; the opposite applies to conservative driving styles. These modifications enable the energy management strategy to adapt to both driving scenarios and driver habits, achieving personalized optimization.

[0056] Step S5: Configure the fuzzy controller Configure the fuzzy controller using the parameters contained in the selected solution. The input to the fuzzy controller is the state-of-charge error. e =SOC target -SOC currente and required power P req The output is the fuel cell power P. fc .

[0057] Quantification: e Multiply by K and map to the fuzzy universe, then P req Multiply K p Mapping to the same domain. Membership functions are triangular functions, dividing the data into seven fuzzy sets. The rule base is fixed. Defuzzification uses the centroid method. The inference result is multiplied by K. u The original fuel cell power P was obtained. fc,raw .

[0058] Directly output Pfc,raw This could lead to a step change in power command, impacting the membrane electrode assembly and bipolar plates of the fuel cell stack. Therefore, a first-order low-pass filter is used.

[0059] Where β is the filter coefficient. After filtering, P fc,out Within the safe operating range of the fuel cell, the battery power is calculated and limited, and instructions are issued for execution.

[0060] Optional, confidence monitoring and incremental updates to the knowledge base During long-term operation, vehicles may encounter new operating conditions not covered by the offline knowledge base (such as newly opened mountain roads or construction zones). If the knowledge base remains unchanged, the frequency of scenario three in step S3 will gradually increase as the operating environment changes. Although the weighted combination mechanism for scenario three can temporarily address the issue, relying on weighted combinations in the long term cannot achieve optimal control.

[0061] Therefore, this embodiment introduces a self-learning mechanism. During online operation, the confidence index is recorded for each working condition matching:

[0062] The smaller the ratio, the closer the current feature is to the cluster center of a known category, and the higher the matching confidence. Conversely, when the ratio is close to or exceeds 1, it indicates that the current working condition is at the boundary or beyond the known range.

[0063] When multiple consecutive cycles γ When the number of samples falls below a preset threshold, it indicates that the current driving condition is not well covered by existing categories. In this case, the feature vector of the current window and the actual controller parameters selected within that window (i.e., the solution selected in step S4) are stored as candidate samples in the cache. When the number of candidate samples in the cache reaches a preset capacity, an offline update process is triggered: all candidate samples are merged with historical samples, the cluster analysis in step S1 is re-executed, the cluster centers, covariance matrix, and intra-cluster Mahalanobis radii are updated, and the multi-objective optimization in step S2 is re-executed for newly added or significantly changed categories, thereby updating the offline knowledge base. The cache is then cleared after the update is complete.

[0064] This mechanism allows the offline knowledge base to evolve gradually as the vehicle's operating environment expands. Without human intervention, the system automatically incorporates new operating condition types into the knowledge base and re-optimizes the corresponding control parameters, thereby maintaining stable control performance over long-term operation. Example

[0065] This embodiment provides an adaptive configuration system for energy management parameters of a fuel cell vehicle, used to implement the method described in Embodiment 1. Figure 2As shown, the system of the present invention includes five modules, which are respectively deployed in the development computer and the vehicle electronic control unit (ECU).

[0066] Offline build module: Located on the development computer, it is used to execute steps S1 and S2, receive historical operating condition data, output offline parameter library files, and download them to the ECU for storage via the diagnostic interface.

[0067] Online matching module: Located in ECU, it is used to execute step S3, maintain the sliding window, calculate the Mahalanobis distance, and output the corresponding Pareto optimal solution set based on soft boundary processing.

[0068] Parameter selection module: Located in the ECU, it is used to execute step S4, read the current state of charge, calculate the dynamic weighting coefficient, select the solution with the smallest score from the received Pareto optimal solution set, and make corrections according to the operating condition category and driving style, and output the controller parameters.

[0069] Control execution module: Located in the ECU, it is used to execute step S5, receive parameters, configure the fuzzy controller, calculate the target power of the fuel cell, and send it to the actuator.

[0070] Self-learning update module: Located in ECU, it is used to perform confidence monitoring and incremental update steps, maintain candidate sample cache, and trigger offline update when conditions are met.

[0071] The above five modules work together to form a complete closed-loop adaptive energy management system.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive configuration method for energy management parameters of a fuel cell vehicle, characterized in that, include: Step S1: Collect multiple driving segments, extract the working condition feature parameters of each driving segment, use a clustering algorithm to classify similar segments into the same working condition category, and record the cluster center and intra-cluster Mahalanobis radius of each working condition category; Step S2: For each operating condition category, with equivalent hydrogen consumption and fuel cell power fluctuation rate as dual optimization objectives, a multi-objective optimization algorithm is used to solve for the Pareto optimal solution set for that category. Each solution corresponds to a set of fuzzy controller parameters. Step S3: During online operation, vehicle driving data is collected in real time, current operating condition feature parameters are extracted, the Mahalanobis distance between the current feature vector and each cluster center is calculated, and soft boundary processing is performed based on the relationship between the Mahalanobis distance and the Mahalanobis radius within the corresponding cluster to obtain the currently available Pareto optimal solution set. Step S4: Obtain the deviation between the current state of charge and the target state of charge, and select a solution from the obtained Pareto optimal solution set according to the deviation. When the deviation indicates that the state of charge is low, the solution with smaller equivalent hydrogen consumption is selected, and when the deviation indicates that the state of charge is high, the solution with smaller power fluctuation rate is selected. Step S5: Configure the fuzzy controller using the fuzzy controller parameters corresponding to the selected solution, and output the target power of the fuel cell.

2. The method according to claim 1, characterized in that, The soft boundary processing includes: When there exists a unique class such that the Mahalanobis distance is less than the Mahalanobis radius within that class, obtain the Pareto optimal solution set for that class; When there are multiple categories such that the Mahalanobis distance is less than the Mahalanobis radius within each category, obtain the union of the Pareto optimal solution sets of these categories; When the Mahalanobis distance of all classes is greater than or equal to the Mahalanobis radius within the corresponding class, obtain a weighted combination of the Pareto optimal solution sets of at least the two classes with the closest Mahalanobis distance.

3. The method according to claim 1, characterized in that, In step S3, a sliding window of fixed duration is used when extracting the current working condition feature parameters. The duration of the sliding window is 120 seconds, and it slides once every 1 second, moving in the latest 1 second data and moving out the earliest 1 second data each time.

4. The method according to claim 1, characterized in that, The formula for calculating the Mahalanobis distance is: ； in, For the current feature vector, The cluster center of class 1, For the first j The covariance matrix of the class; the in-class Mahalanobis radius is the first... All training samples in the class are grouped to cluster centers. The maximum value of the Mahalanobis distance.

5. The method according to claim 1, characterized in that, The method for selecting a solution based on the deviation in step S4 is as follows: First, construct the dynamic weighting coefficient α = 0.5 + k*ΔSOC, where ΔSOC=SOC target -SOC current k is the calibration coefficient; Secondly, calculate the normalized equivalent hydrogen consumption H for each solution in the current Pareto optimal solution set. norm and normalized power volatility F norm : ， ； in , These are the maximum and minimum equivalent hydrogen consumption values ​​in the current solution set, respectively. , These are the maximum and minimum power volatility values ​​in the current solution set, respectively. Finally, select the option that makes the overall score S = α* H The solution with the smallest norm+(1−α)*Fnorm.

6. The method according to claim 5, characterized in that, Also includes: The dynamic weighting coefficients are adjusted according to the currently identified operating condition category, wherein the economic weight is increased under high-speed cruising conditions and the lifespan weight is increased under low-speed congestion conditions.

7. The method according to claim 5, characterized in that, Also includes: The current driving style is identified and the dynamic weighting coefficient is adjusted, wherein the fuel economy weight is reduced for aggressive driving style and increased for conservative driving style.

8. The method according to claim 1, characterized in that, The fuzzy controller has a dual-input single-output structure. The inputs are the state-of-charge error and the required power, and the output is the fuel cell power. The parameters of the fuzzy controller include the input quantization factor, the output scaling factor, and the peak position of the membership function.

9. The method according to claim 1, characterized in that, Also includes: During online operation, the ratio of the minimum Mahalanobis distance to the Mahalanobis radius within the corresponding class is recorded as a confidence index for each working condition matching. When the confidence index is lower than a preset threshold for several consecutive periods, the current feature vector and the actual optimal control parameters within a subsequent time window are used as new samples. The offline knowledge base is then updated by incremental clustering using the new samples.

10. An adaptive configuration system for energy management parameters of a fuel cell vehicle, characterized in that, include: An offline build module is used to perform steps S1 and S2 as described in claim 1; An online matching module is used to perform step S3 as described in claim 1; The parameter selection module is used to perform step S4 as described in claim 1; A control execution module is used to execute step S5 as described in claim 1; A self-learning update module is used to perform the steps in the method of claim 9.

Images