Self-adaptive energy management method based on parameter difference matching of cell and fuel cell

Abstract

The invention relates to a self-adaptive energy management method based on parameter difference matching of a battery and a fuel cell. The method is used for solving the problem that an existing static strategy library depends on limited parameter samples collected in advance and cannot cope with continuously added hardware models in the market. And the existing online optimization algorithm needs long-time trial and error learning, so that the normal use of the vehicle is influenced. According to the scheme, key feature parameters of new hardware are captured in real time through a dynamic parameter clustering module, the key feature parameters are classified into an optimal strategy cluster by adopting an improved clustering algorithm, and the solidified parameter range limitation of a traditional strategy library is broken through; a layered reinforcement learning architecture is designed, a base strategy layer presets a high-frequency strategy to ensure second-level response, an increment strategy layer optimizes a new parameter combination online, real-time performance and precision are achieved, and the defect that online learning consumes too long time is overcome; and an incremental learning mechanism is synchronously constructed, and a strategy library is continuously and dynamically expanded during operation, so that energy management of adaptive cell and fuel cell parameter difference matching is realized.

Description

Technical Field

[0001] The present disclosure relates to the technical field of hybrid vehicle energy management, and in particular to an adaptive energy management method based on dynamic matching of battery and fuel cell parameter differences. Background Art

[0002] In the actual operation of hydrogen-electric hybrid vehicles, frequent replacement of battery packs and fuel cell stacks has become a daily necessity. For example, shared cars or logistics fleets often require rapid replacement of batteries of different models or batches to shorten charging times and improve operational efficiency. However, key parameters of the newly installed hardware (such as battery internal resistance and fuel cell power) differ significantly from the original system presets. This difference is not simply a numerical offset, but involves the coordinated changes of multiple parameters. Furthermore, when a vehicle's battery or fuel cell is replaced, traditional fixed-strategy energy management systems immediately expose their flaws: system efficiency decreases, response delays in the power distribution module increase, and in extreme cases, even lead to safety hazards such as circuit fluctuations and component overheating. Current solutions suffer from two major issues: First, the rigidity of static policy libraries, which rely on a limited number of pre-collected parameter samples and cannot adapt to the ever-increasing hardware models on the market. Second, existing online optimization algorithms require lengthy trial-and-error learning, which is equivalent to retraining the vehicle's "driving habits" after each battery replacement, seriously impacting normal vehicle operation. Therefore, it is crucial to design an energy management method that can rapidly and adaptively adapt to the differences in battery and fuel cell parameters. Summary of the Invention

[0003] The purpose of this disclosure is to propose an adaptive energy management method for multi-source hardware adaptation of hydrogen-electric hybrid vehicles to solve the control mismatch and efficiency loss problems caused by replacing batteries and fuel cells.

[0004] To achieve the above technical objectives, the present disclosure proposes, on the one hand, an adaptive energy management method based on matching battery and fuel cell parameter differences, the method comprising the following steps: constructing a parameter difference database, and dynamically clustering parameter combinations to generate multi-dimensional typical operating condition clusters; using reinforcement learning to perform independent offline training on each operating condition cluster in the typical operating condition cluster to generate the optimal energy management strategy corresponding to each cluster; constructing an adaptive strategy library with parameter interpretability for the optimal energy management strategy of each operating condition cluster; based on the battery and fuel cell parameters of the current vehicle, comparing the real-time parameter characteristics with the typical operating condition cluster, and judging in real time the cluster to which the current operating condition belongs; matching the adaptive strategy library based on the cluster to which the current operating condition belongs to determine the currently adopted energy management strategy.

[0005] In one implementation of the above technical solution, the dynamic clustering adopts a feature-weighted K-means algorithm.

[0006] In one embodiment of the above technical solution, the energy management strategy currently adopted is determined based on the cluster matching adaptive strategy library to which the current working condition belongs, including: if the cluster is known, then the corresponding strategy parameter θ is directly loaded from the strategy library according to the matched cluster ID. k , and then use the policy network Generate control actions; if it is a new cluster, select the two clusters with the smallest distance (k1, k2) according to the calculated weighted distance and perform weight distribution. The specific weight distribution formula is as follows: Next, parameter strategy fusion is performed. The specific strategy fusion formula is as follows: Then, temporary instructions are generated through the fused strategy parameters for control; where: are the distances between the two clusters (k1, k2) with the smallest distance, are the weights of the two clusters (k1, k2) with the smallest distance, are the strategy parameters of the two clusters (k1, k2) with the smallest distance.

[0007] In one embodiment of the above technical solution, while generating temporary instructions for control, the strategy library is updated, including the new cluster center μ new and policy parameters θ new Add to the strategy library, the calculation formula for the new cluster center is as follows: are the cluster centers of the two clusters (k1, k2) with the smallest distance.

[0008] In one embodiment of the above technical solution, constructing a parameter difference database includes: first generating a large number of parameter combinations by Latin hypercube sampling, covering the battery internal resistance R b 、Battery capacity C b 、Battery maximum discharge rate k dis , fuel cell rated power P fc and the slope of the efficiency curve β; each parameter dimension is divided into N equal probability intervals, and by randomly selecting each probability interval, a combination covering the entire parameter space is generated to obtain a parameter difference database.

[0009] In one embodiment of the above technical solution, the state space of the reinforcement learning is SOC is the battery state of charge, P req Indicates the real-time power demand of the vehicle. Indicates the internal resistance of the current cluster battery, k indicates the current cluster, represents the rated power of the current cluster fuel cell, η fc represents the current efficiency of the fuel cell, η b Represents the battery charging and discharging efficiency; the action space is U=[α fc ], αfc is the fuel cell power ratio; the reward function is r = μ·r energy +τ·r SOC , μ and τ represent the weight distribution coefficients, r energy and r SOC Represent the energy consumption term and SOC maintenance term respectively; among them, It represents the hydrogen consumption at the current moment. It represents the reference hydrogen consumption benchmark value, P b Indicates the battery power, P b 'Represents the battery rated power loss reference value; SOC' is the reference value, |SOC-SOC'| represents the linear penalty, and λ represents the out-of-bounds penalty coefficient. It represents the indicator function. When SOC is maintained in the safe range ([SOC min ,SOC max ]), the output of function f is 1, otherwise it is 0.

[0010] In one embodiment of the above technical solution, based on the cluster matching adaptive strategy library of the current working condition, it includes: assigning a unique and interpretable ID tag to the optimal energy management strategy of each cluster, and the coding structure of the ID is ID k =Param b _Param fc , Param b Indicates the range of key battery parameters, Param fc It represents the range of key parameters of fuel cells.

[0011] In one embodiment of the above technical solution, based on the battery and fuel cell parameters of the current vehicle, the real-time parameter characteristics are compared with the typical operating condition cluster, and the cluster to which the current operating condition belongs is determined in real time. The steps include: obtaining the battery and fuel cell parameters of the current vehicle, constructing X current =[R b ,C b ,k dis ,P fc ,β],R b Battery internal resistance, C b is the battery capacity, k dis is the maximum discharge rate of the battery, P fc is the rated power of the fuel cell, β is the slope of the efficiency curve; X current Each component in is processed by Min-Max normalization to obtain x norm,j , j=1,2,3,4,5; perform weighted distance calculation μ kis the center vector of the kth cluster, p is the order of the norm, which is used to define the distance metric. Here, the weighted form of Euclidean distance is used, so p = 2; the cluster to which the current parameter belongs is determined by nearest neighbor matching.

[0012] In order to achieve the above technical objectives, the present disclosure also proposes an adaptive energy management device based on parameter difference matching of batteries and fuel cells, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and execute any of the above methods.

[0013] In order to achieve the above technical objectives, the present disclosure also proposes a computer-readable storage medium storing a computer program that can be loaded by a processor and execute any of the above methods.

[0014] The technical effects of the present disclosure are as follows: (1) Traditional methods rely on fixed parameter thresholds or single-dimensional analysis, which makes it difficult to cope with scenarios where battery and fuel cell parameter differences are matched. This technology can automatically identify complex parameter difference combinations through dynamic weighted clustering and multi-parameter joint analysis, match the optimal control strategy in real time, and significantly improve the adaptability of working conditions and system stability. (2) Compared with existing technologies that rely on fixed strategy libraries or manual retraining, this technology uses an incremental learning mechanism to automatically integrate historical optimal strategies and update the database when new working conditions occur, thereby achieving dynamic expansion of the strategy library without the need for downtime or manual intervention, significantly improving the long-term adaptability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the flow chart of the energy management method based on adaptive battery and fuel cell parameter difference matching for hybrid electric vehicles. DETAILED DESCRIPTION

[0016] In response to the problems that the existing static policy library relies on limited pre-collected parameter samples and cannot cope with the continuous increase in new hardware models on the market; and the problem that the existing online optimization algorithm requires a long period of trial and error learning, which affects the normal use of the vehicle. The present invention uses a dynamic parameter clustering module to capture the key characteristic parameters of new hardware in real time, and adopts an improved clustering algorithm to classify them into optimal policy clusters, breaking through the parameter range limitations of the traditional policy library; designing a hierarchical reinforcement learning architecture, the base policy layer presets high-frequency strategies to ensure second-level response, and the incremental policy layer optimizes new parameter combinations online, combining real-time and precision, overcoming the defect of long online learning time; synchronously constructing an incremental learning mechanism, and continuously and dynamically expanding the policy library during operation, thereby realizing energy management that adaptively matches the parameter differences of batteries and fuel cells.

[0017] The following, combined with the accompanying drawings, provides a clear and complete description of how the technical solution of this case is implemented. Obviously, the described implementation methods are only part of the implementation methods of this case, not all of the implementation methods. Based on the implementation methods of this case, all other implementation methods obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0018] (1) Clustering parameter differences matching typical working condition clusters

[0019] First, a large number of parameter combinations are generated through Latin hypercube sampling to cover the battery internal resistance R b 、Battery capacity C b 、Battery maximum discharge rate k dis , fuel cell rated power P fc The specific logic is to divide each parameter dimension into N equal probability intervals, and through random selection of each probability interval, a combination covering the entire parameter space is generated to obtain a parameter difference database.

[0020] In the parameter difference database, various parameter matching combinations of batteries and fuel cells are simulated, covering extreme scenarios such as internal resistance changes, capacity changes, and power changes. Latin hypercube sampling technology is used to generate massive parameter combinations to ensure full dimensional coverage of the parameter space.

[0021] Secondly, clustering is performed based on parameter characteristics to obtain multiple typical operating condition clusters (e.g., high internal resistance battery + low-efficiency fuel cell). This disclosure uses an improved clustering algorithm for clustering, specifically a feature-weighted K-means algorithm, with weights assigned based on the results of parameter sensitivity analysis. When using the K-means clustering algorithm to classify operating conditions, the key characteristic parameters are mainly determined, such as battery internal resistance, capacity, and fuel cell efficiency offset. The characteristic parameter input matrix is ​​specifically shown in Equation (1).

[0022] X=[R b ,C b ,k dis ,P fc ,β] (1)

[0023] Before clustering, the data is preprocessed and the characteristic parameters of the input matrix are normalized using Min-Max to eliminate dimensional differences, as shown in formula (2).

[0024]

[0025] Clustering is performed based on parameter characteristics. The specific feature weighting is based on the degree of influence of the parameters on energy consumption, and different weights are assigned to ω jThe weighted distance formula is shown in formula (3). Where p is the order of the norm, which is used to define the distance measurement method. The weighted form of Euclidean distance is used here, so p = 2.

[0026]

[0027] For example, when When , the internal resistance difference contributes 40% to the distance. Through clustering in this feature-weighted manner, multiple typical working condition clusters are obtained.

[0028] (2) Offline training of optimal energy management strategies for each working condition

[0029] For the typical working condition clusters obtained by clustering parameter features above, each cluster obtained by clustering is trained independently in combination with reinforcement learning. The PPO algorithm is used to set a multi-objective reward function to obtain the optimal energy management strategy for each cluster.

[0030] Specifically, the specific framework of the reinforcement learning PPO algorithm is first determined, and the state space is specifically shown in formula (4).

[0031]

[0032] Among them, SOC is the battery state of charge, which indicates the percentage of remaining battery power; k represents the current cluster, so, Indicates the internal resistance of the current cluster battery, Represents the rated power of the current cluster fuel cell; η fc represents the current efficiency of the fuel cell; η b Indicates battery charge and discharge efficiency; P req It represents the real-time required power of the vehicle, as shown in formula (5).

[0033]

[0034] Among them, v represents the current speed of the vehicle, a represents the current acceleration of the vehicle, m represents the mass of the bus, g represents the acceleration of gravity, α represents the road slope, C d It represents the air resistance coefficient, A represents the frontal area, δ represents the rotation mass conversion coefficient, η t It indicates the working efficiency of the motor.

[0035] The second is the action space of the specific framework of the PPO algorithm. The action space is specifically shown in formula (6).

[0036] U=[α fc ] (6)

[0037] In order to simplify energy distribution and focus on energy distribution, the action space parameters only focus on the fuel cell power ratio αfc , α fc The value range is [0, 1], and the specific power of the battery and fuel cell (P b and P fc ) as shown in formula (7).

[0038]

[0039] The overall structure of the reward function consists of a weighted combination of energy consumption and SOC maintenance, as shown in formula (8).

[0040] r=μ·r energy +τ·r SOC (8)

[0041] Among them, μ and τ represent weight distribution coefficients. In order to meet the priority of energy consumption optimization and SOC tracking, μ should be set larger than τ in practical applications. energy and r SOC They represent the energy consumption term and the SOC maintenance term, respectively, as shown in Equations (9) and (10).

[0042]

[0043] The main goal of the energy consumption term is to minimize the total energy consumption of the system, which specifically includes the hydrogen consumption of the fuel cell and the charge and discharge losses of the battery. The reason for adding a negative sign is because the PPO reinforcement learning algorithm maximizes the cumulative reward, and energy consumption needs to be minimized, so it is negative; It represents the hydrogen consumption at the current moment. It represents the reference hydrogen consumption benchmark value. The reason for dividing by This is to eliminate the dimensionality effect, normalize the hydrogen consumption, and make the rewards under different working conditions comparable; at the same time, add a battery loss penalty, P b Indicates the battery power, P b ' indicates the rated power loss reference value of the battery. When the battery is charged and discharged at a high rate (|P b | large) or low efficiency (η b When the battery is used up (hours), the battery loss penalty increases.

[0044]

[0045] The goal of the SOC maintenance term is to stabilize the battery near the reference value SOC' while preventing overcharging or over-discharging. |SOC-SOC'| represents a linear penalty. When the SOC deviates from the reference value, the penalty increases linearly with the deviation. At the same time, a heavy penalty for out-of-bounds is introduced to force the strategy to correct the deviation. λ represents the out-of-bounds penalty coefficient. It represents the indicator function, which is 1 when the condition is met and 0 otherwise. Here, it means that when the SOC is maintained in the safe range ([SOCmin ,SOC max ]), the output of the f function is 1, otherwise it is 0, to ensure that the strategy will not cause overcharging or over-discharging of the battery.

[0046] Finally, the PPO algorithm framework also includes physical constraints, specifically constraints on motors, batteries, and fuel cells.

[0047] The physical constraints on the motor include the constraints on the motor output power and torque. As shown in formula (11):

[0048]

[0049] The physical constraints on the battery include constraints on the battery output power and state of charge. As shown in formula (12):

[0050]

[0051] The physical constraints on fuel cells are mainly the constraints on the output power of the fuel cell, as shown in formula (13):

[0052] P fc_min ≤P fc ≤P fc_max (13)

[0053] (3) Building an adaptive strategy library

[0054] Based on the optimal energy management strategy obtained through training, the optimal energy management strategy corresponding to each cluster is labeled with an ID to build an adaptive strategy library.

[0055] Specifically, the goal of the ID tag generation rule in this scheme is to assign a unique and interpretable ID tag to the optimal energy management strategy of each cluster, thereby supporting fast matching. The specific ID encoding structure is shown in Equation (14).

[0056] ID k =Param b _Param fc (14)

[0057] Among them, Param b It indicates the range of key battery parameters, such as R0.4-0.6 means the battery internal resistance range is 0.4~0.6Ω, C60-80 means the battery capacity is 60-80Ah; Param fc It represents the range of key fuel cell parameters. For example, P50-70 means the fuel cell power is 50-70kW, and η50-60 means the fuel cell efficiency is 50% to 60%.

[0058] (IV) Real-time determination of the cluster to which the current working condition belongs

[0059] Based on the current vehicle information, after inputting the battery and fuel cell parameters of the current vehicle, the real-time parameter characteristics are compared with the typical operating condition clusters obtained by clustering to determine the typical operating condition cluster to which the current state belongs.

[0060] Specifically, first input the current component parameters. The specific formula of the parameters involved is shown in formula (15):

[0061] X current =[R b ,C b ,k dis ,P fc ,β] (15)

[0062] In the real-time matching process, parameters also need to be pre-processed. current Each component in also adopts the Min-Max normalization parameter and uses the global extreme value of the original sampling data, as shown in formula (16).

[0063]

[0064] Then the weighted distance calculation is performed to obtain the weighted distance of the current situation. The specific weighted distance calculation formula is shown in formula (17).

[0065]

[0066] Where: p is the order of the norm, which is used to define the distance metric. The weighted form of Euclidean distance is used here, so p = 2.

[0067] It should be noted here that μ k is the center vector of the kth cluster and comes from the aforementioned clustering results.

[0068] Then the cluster to which the current parameter belongs is determined by nearest neighbor matching. Specifically, the weighted distance between the current parameter and all cluster centers is calculated by formula (17), and then the minimum distance D is selected. min , and record the corresponding cluster k * , then the minimum distance D min The distance value D is set to thershold In contrast, if D min ≤D thershold , then the cluster to which the current parameter belongs is determined to be k * Otherwise, it is determined to be a new cluster.

[0069] (5) Matching the adaptive strategy library for energy management

[0070] Based on the typical operating condition cluster ID of the current state and the matching strategy library, the energy management strategy adopted under the current operating condition is determined, the corresponding strategy parameters are called, and the real-time issuance of control instructions is completed, thereby realizing adaptive energy management.

[0071] If the cluster is known, the corresponding strategy parameter θ is loaded directly from the strategy library according to the matched cluster ID. k , and then use the policy network Generate control actions.

[0072] If it is a new cluster, then according to the calculated weighted distance D, the two clusters with the smallest distance (k1, k2) are selected for weight distribution. The specific weight distribution formula is shown in formula (18).

[0073]

[0074] Where: are the distances between the two clusters (k1, k2) with the smallest distance.

[0075] Then the parameter strategy fusion is performed. The specific strategy fusion formula is shown in formula (19).

[0076]

[0077] Where: are the weights of the two clusters (k1, k2) with the smallest distance, are the strategy parameters of the two clusters (k1, k2) with the smallest distance.

[0078] Then, temporary instructions are generated through the fused policy parameters for control, and the policy library is updated at the same time. Specifically, the new cluster center μ new and policy parameters θ new Add to the strategy library, and the calculation formula of the new cluster center is shown in formula (20).

[0079]

[0080] Where: are the cluster centers of the two clusters (k1, k2) with the smallest distance.

[0081] As can be seen from the above technical principles and implementation process, the adaptive energy management method based on dynamic parameter clustering and hierarchical reinforcement learning proposed in this invention constructs a full-dimensional parameter space covering battery internal resistance, capacity, and fuel cell efficiency offsets through Latin hypercube sampling. A feature-weighted K-means algorithm is used to achieve dynamic clustering of multiple parameters, accurately dividing typical operating condition clusters. Combined with an improved PPO reinforcement learning framework, the optimal policy integrating a multi-objective reward function is independently trained for each operating condition cluster, constructing an adaptive policy library with parameter interpretability. During real-time operation, a weighted distance algorithm is used to dynamically analyze hardware parameter features, achieving millisecond-level matching of known operating condition cluster policies and incremental learning adaptation of new clusters. Using a neighboring policy fusion and dynamic cluster center update mechanism, the policy library is autonomously expanded while ensuring real-time control accuracy. Compared with traditional fixed threshold matching and online retraining schemes, this method effectively solves the problem of dynamic adaptation of multi-source hardware parameters in hybrid power systems, thereby realizing an energy management method for adaptively matching battery and fuel cell parameter differences based on hybrid vehicles.

[0082] Through the description of the above implementation methods, technical personnel in the relevant field can clearly understand the method disclosed in the present invention, and the steps include: constructing a parameter difference database, and dynamically clustering the parameter combinations to generate multi-dimensional typical operating condition clusters; using reinforcement learning to perform independent offline training on each operating condition cluster in the typical operating condition cluster to generate the optimal energy management strategy corresponding to each cluster; for the optimal energy management strategy of each operating condition cluster, constructing an adaptive strategy library with parameter interpretability; based on the battery and fuel cell parameters of the current vehicle, comparing the real-time parameter characteristics with the typical operating condition cluster, and judging in real time the cluster to which the current operating condition belongs; matching the adaptive strategy library based on the cluster to which the current operating condition belongs, and determining the currently adopted energy management strategy.

[0083] Through the description of the above embodiments, those skilled in the art can clearly understand that the method of the present disclosure can be implemented by means of software plus necessary general-purpose hardware, and of course can also be implemented by means of dedicated hardware including dedicated integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. In general, all functions performed by computer programs can be easily implemented with corresponding hardware, and the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits, or dedicated circuits. However, for the present disclosure, software program implementation is a better implementation method in most cases. Therefore, those skilled in the art can clearly understand that according to the method of the present disclosure, an adaptive energy management device based on parameter difference matching of batteries and fuel cells can be implemented, including a memory and a processor, and the memory stores a computer program that can be loaded by the processor and execute any of the above methods.

[0084] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. An adaptive energy management method based on parameter difference matching between battery and fuel cell, characterized in that: The method comprises the following steps: Build a parameter difference database and dynamically cluster parameter combinations to generate multi-dimensional typical operating condition clusters; Each operating condition cluster in the typical operating condition cluster is trained independently offline using reinforcement learning to generate the optimal energy management strategy corresponding to each cluster; Build an adaptive strategy library with interpretable parameters for the optimal energy management strategy for each operating condition cluster; Based on the current vehicle battery and fuel cell parameters, the real-time parameter characteristics are compared with the typical operating condition cluster to determine in real time which cluster the current operating condition belongs to; Based on the cluster matching adaptive strategy library to which the current working condition belongs, the currently adopted energy management strategy is determined.

2. The method according to claim 1, characterized in that The dynamic clustering adopts the feature-weighted K-means algorithm.

3. The method according to claim 1, characterized in that Based on the cluster matching adaptive strategy library of the current working condition, the current energy management strategy is determined, including: If the cluster is known, the corresponding strategy parameter θ is loaded directly from the strategy library according to the matched cluster ID. k , and then use the policy network Generate control actions; If it is a new cluster, the two clusters with the smallest distance (k1, k2) are selected according to the calculated weighted distance and weight distribution is performed. The specific weight distribution formula is as follows: Perform parameter strategy fusion. The specific strategy fusion formula is as follows: Then, temporary instructions are generated through the fused strategy parameters for control; Where: are the distances between the two clusters (k1, k2) with the smallest distance, are the weights of the two clusters (k1, k2) with the smallest distance, are the strategy parameters of the two clusters (k1, k2) with the smallest distance.

4. The method according to claim 3, characterized in that While generating temporary instructions for control, the policy library is updated, including: The new cluster center μ new and policy parameters θ new Add to the strategy library, the calculation formula for the new cluster center is as follows: Where: are the cluster centers of the two clusters (k1, k2) with the smallest distance.

5. The method according to claim 1, wherein Construct parameter difference database, including: First, a large number of parameter combinations are generated through Latin hypercube sampling to cover the battery internal resistance R b 、Battery capacity C b 、Battery maximum discharge rate k dis , fuel cell rated power P fc and the slope of the efficiency curve, β; Each parameter dimension is divided into N equal probability intervals, and by randomly selecting each probability interval, a combination covering the entire parameter space is generated to obtain a parameter difference database.

6. The method according to claim 1, wherein The state space of the reinforcement learning is SOC is the battery state of charge, P req Indicates the real-time power demand of the vehicle. Indicates the internal resistance of the current cluster battery, k indicates the current cluster, represents the rated power of the current cluster fuel cell, η fc represents the current efficiency of the fuel cell, η b Represents the battery charging and discharging efficiency; the action space is U=[α fc ], α fc is the fuel cell power ratio; the reward function is r = μ·r energy +τ·r SOC , μ and τ represent the weight distribution coefficients, r energy and r SOC Represent the energy consumption item and SOC maintenance item respectively; in, It represents the hydrogen consumption at the current moment. It represents the reference hydrogen consumption benchmark value, P b Represents the battery power, P′ b Indicates the battery rated power loss reference value; SOC' is the reference value, |SOC-SOC'| represents the linear penalty, and λ represents the out-of-bounds penalty coefficient. It represents the indicator function. When SOC is maintained in the safe range ([SOC min ,SOC max ]), the output of function f is 1, otherwise it is 0.

7. The method according to claim 1, characterized in that Based on the cluster matching adaptive strategy library of the current working condition, including: assigning a unique and interpretable ID tag to the optimal energy management strategy of each cluster, the coding structure of the ID is ID k =Param b _Param fc , Param b Indicates the key parameter range of the battery, Param fc It represents the range of key parameters of fuel cells.

8. The method according to claim 1, characterized in that Based on the current vehicle battery and fuel cell parameters, the real-time parameter characteristics are compared with the typical operating condition cluster to determine in real time which cluster the current operating condition belongs to. The steps include: Get the battery and fuel cell parameters of the current vehicle and construct X current =[R b ,C b ,k dis ,P fc ,β],R b Battery internal resistance, C b is the battery capacity, k dis is the maximum discharge rate of the battery, P fc is the rated power of the fuel cell, β is the slope of the efficiency curve; X current Each component in is processed by Min-Max normalization to obtain x norm,j , j=1,2,3,4,5; Perform weighted distance calculation μ k is the center vector of the kth cluster, and p is the order of the norm; The cluster to which the current parameter belongs is determined through nearest neighbor matching.

9. An adaptive energy management device based on parameter difference matching between batteries and fuel cells, characterized by: The method comprises a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executes the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored which can be loaded by a processor and execute the method according to any one of claims 1 to 8.

Images