Hydrogen fuel cell bus energy management decision method and system

Abstract

The application discloses a hydrogen fuel cell bus energy management decision method and system, relates to the battery management technical field, and comprises the following steps: defining the state space and action space of a hybrid power system; based on the defined state space and action space, constructing an energy management model comprising a Q network, a target network double network architecture and an experience replay mechanism; adding an uncertainty protection module to the constructed energy management model to inhibit the instantaneous fluctuation of fuel cell output power; deploying the energy management model on a vehicle controller, collecting real-time running state parameters of the vehicle, inputting the model, outputting optimal power distribution instructions of the fuel cell and the power battery, and completing the whole vehicle energy management decision. The application quantifies the working condition uncertainty based on the full-action Q value distribution, dynamically adjusts the fuel cell aging protection strength, and realizes the dynamic optimal balance of fuel economy and battery durability.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and in particular to an energy management decision-making method and system for hydrogen fuel cell buses. Background Technology

[0002] With the development of new energy vehicle technology and the promotion of the hydrogen energy industry, hydrogen fuel cell buses, with their advantages of zero emissions, long driving range, and short refueling time, have become one of the core technological routes for low-carbon transformation in urban public transportation. Currently, mainstream hydrogen fuel cell buses generally adopt a hybrid power system consisting of a fuel cell and a power battery. The fuel cell serves as the primary power source, providing basic driving power, while the power battery acts as an auxiliary power source, handling peak power demands and recovering braking energy. The two systems coordinate their output power through energy management strategies to meet the dynamic power requirements of the entire vehicle. Energy management strategy is the core control technology of the hybrid power system, and its performance directly determines the vehicle's fuel economy, power source lifespan, and overall operating costs.

[0003] Existing energy management strategies for hydrogen fuel cell hybrid power systems can be mainly divided into three categories: First, rule-based strategies, such as thermostat strategies and power-following strategies. These strategies set power allocation thresholds based on engineering experience, are simple in structure, highly reliable, and easy to implement in engineering. However, the strategy parameters are highly dependent on manual calibration, have poor adaptability to operating conditions, and are difficult to achieve multi-objective optimization under complex urban public transport conditions. Second, offline optimization strategies, such as dynamic programming and Pontryagin's minimum principle. These strategies can find the global optimal solution under known complete operating conditions. However, the calculation process depends on all future operating condition information and cannot be directly applied to online decision-making in actual vehicles. Third, instantaneous optimization strategies, such as the equivalent hydrogen consumption minimization strategy. These strategies can achieve online real-time optimization. However, the value of the equivalent factor has a significant impact on the optimization effect and is difficult to dynamically adjust with operating condition fluctuations. The optimization effect drops significantly in scenarios with high operating condition uncertainty.

[0004] Urban public transport operation is characterized by frequent start-stop cycles, large power fluctuations, and significant uncertainties in passenger flow and road conditions. Fuel cells are subjected to prolonged aging conditions such as idling, start-stop cycles, and load variations, resulting in a significantly higher degradation rate than in steady-state operation. Existing energy management strategies struggle to simultaneously address the requirements of low hydrogen consumption and long lifespan in urban public transport scenarios, hindering the operational economics and large-scale application of hydrogen fuel cell buses. Therefore, there is an urgent need for an energy management decision-making method for hydrogen fuel cell buses that can dynamically adjust lifespan protection levels based on real-time operational uncertainties, achieving a dynamic optimal balance between fuel economy and fuel cell durability. Summary of the Invention

[0005] This invention provides an energy management decision-making method for hydrogen fuel cell buses, comprising: Step S10: Define the state space and action space of the hybrid power system; Step S20: Based on the defined state space and action space, construct an energy management model that includes a dual-network architecture of Q-network and target network and an experience playback mechanism; Step S30: Add an uncertainty protection module to the constructed energy management model to suppress instantaneous fluctuations in fuel cell output power; Step S40: Deploy the energy management model on the vehicle controller, collect real-time vehicle operating status parameters into the model, output the optimal power allocation command for the fuel cell and power battery, and complete the vehicle energy management decision.

[0006] The energy management decision-making method for hydrogen fuel cell buses described above involves constructing an energy management model based on a defined state space and action space, incorporating a dual-network architecture (Q-network and target network) and an experience playback mechanism. This process is divided into the following sub-steps: Step S21: Construct a dual-network architecture that supports Q-value distribution output across all actions; Step S22: Configure an extended experience playback mechanism with working condition tags; Step S23: Design a multi-objective initial reward function and set training hyperparameters and exploration mechanisms for urban public transportation scenarios; Step S24: Perform offline training until the model converges.

[0007] The energy management decision-making method for hydrogen fuel cell buses described above includes adding an uncertainty protection module to the constructed energy management model, which specifically consists of the following sub-steps: The basic value of instantaneous uncertainty at the current moment is calculated using the full-action Q-value distribution output by the Q-network of the energy management model. Step S32: Map the current instantaneous uncertainty baseline value to the aging equivalent cost factor of the fuel cell; Step S33: Solve for the optimal output power of the fuel cell based on the aging equivalent cost factor; Step S34: Integrate the solution logic of steps S31 to S33 into the output of the Q network to form an uncertainty protection module.

[0008] The energy management decision-making method for hydrogen fuel cell buses described above includes the following complete online decision-making process for the energy management model: Real-time data collection of the vehicle's required power and the battery's state of charge (SOC) is input into the Q network. The Q-network outputs the Q-values ​​corresponding to all possible actions in the current state; Calculate the basic value of instantaneous uncertainty according to step S31. ; Calculate the aging equivalent cost factor at the current moment according to step S32. ; The optimal output power with the minimum overall cost is determined in step S33 and then sent to the power system for execution. Waiting for the next decision cycle, the above process is repeated to achieve rolling online energy management decisions.

[0009] The energy management decision-making method for hydrogen fuel cell buses described above involves deploying an energy management model on the vehicle controller, collecting real-time vehicle operating status parameters as input to the model, and outputting optimal power allocation instructions between the fuel cell and the power battery to complete the vehicle's energy management decision. Specifically, it includes the following sub-steps: Step S41: Onboard deployment and parameter solidification of the energy management model; Step S42: Real-time vehicle operating status parameter acquisition and standardized preprocessing; Step S43: Generation of power allocation instructions for dual power sources and verification of system constraints; Step S44: Power command issuance and energy management decision-making closed-loop execution.

[0010] The present invention also provides an energy management decision system for hydrogen fuel cell buses, comprising: a space definition module, a management model training module, and an instruction generation and distribution module; The space definition module is used to define the state space and action space of the hybrid power system. The management model training module is used to construct an energy management model based on a defined state space and action space, which includes a dual-network architecture of Q-network and target network and an experience playback mechanism, and adds an uncertainty protection module to the constructed energy management model. The instruction generation and distribution module is used to deploy the energy management model on the vehicle controller, collect real-time vehicle operating status parameters into the model, output the optimal power allocation instruction for fuel cells and power batteries, and complete the vehicle energy management decision.

[0011] The beneficial effects achieved by this invention are as follows: Based on the full-action Q-value distribution, the uncertainty of the working conditions is quantified, and the aging protection strength of the fuel cell is dynamically adjusted to achieve a dynamic optimal balance between fuel economy and battery durability; the solution reuses the network output, eliminating the need for additional sensors and working condition identification algorithms, resulting in low on-board deployment costs, effectively reducing hydrogen consumption in bus operations, extending the service life of fuel cells, and adapting to the complex operating conditions of urban buses. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0013] Figure 1 This is a flowchart of an energy management decision-making method for hydrogen fuel cell buses provided in Embodiment 1 of this application; Figure 2 This is a schematic diagram of an energy management decision system for a hydrogen fuel cell bus provided in Embodiment 2 of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example 1 like Figure 1 As shown, Embodiment 1 of this application provides an energy management decision-making method for hydrogen fuel cell buses. The hydrogen fuel cell bus adopts a hybrid power system topology with the fuel cell and power battery indirectly connected in parallel. An energy management strategy coordinates and allocates the output power of the fuel cell and power battery to meet the power requirements of the entire vehicle. The method includes the following steps: Step S10: Define the state space and action space of the hybrid power system; This embodiment applies to a hydrogen fuel cell bus hybrid power system employing a fuel cell-power battery indirect parallel topology. The fuel cell, after being boosted by a unidirectional DC / DC converter, is connected in parallel with the power battery to the DC bus, jointly outputting electrical energy to the drive motor. This step is used to define the input dimension, action output dimension, and feasible solution constraint boundary for the subsequent energy management model. The specific implementation process is as follows: First, the state space of the hybrid power system is defined; in this embodiment, the vehicle power demand and the state of charge (SOC) of the power battery are selected as state variables to form a two-dimensional state space. ,in The power demand of the entire vehicle is used to characterize the instantaneous power demand under the current driving conditions, covering both driving and regenerative braking. Its value range is set to [-85kW, 200kW], with negative values ​​corresponding to the regenerative power under regenerative braking and positive values ​​corresponding to the driving power required for vehicle operation. The State of Charge (SOC) characterizes the remaining energy level of the power battery, calculated in real time using the ampere-hour integral method. The calculation step size is consistent with the system decision cycle. The basic parameters of the power battery match the specifications of commonly used lithium iron phosphate power batteries for urban buses. Considering the operating characteristics of urban buses and the health protection requirements of the power battery, its effective operating range is set to [0.6, 0.8]. The combination of the above two state variables can completely characterize the real-time operating state of the system required for energy management decisions, satisfying the state representation requirements of the Markov decision process, and can be directly used as input parameters for subsequent energy management models.

[0016] Secondly, the action space of the hybrid power system is defined. Based on the vehicle power balance relationship, once the vehicle's required power and the fuel cell's output power are determined, the output power or charging power of the power battery can be uniquely derived through the power balance relationship. Therefore, this embodiment selects the fuel cell output power. As the sole control variable for actions, it constitutes a one-dimensional action space. Based on the rated parameters and efficient operating range of the fuel cell, the output power range of the fuel cell is set to [5kW, 68kW]. The lower limit is used to avoid the fuel cell operating in the inefficient idling range for a long time, and the upper limit is used to match the rated peak output capability of the fuel cell. In order to match the discrete action output form of the subsequent Q network, the above power range is discretized into 64 power nodes with a uniform step size. The power interval between adjacent nodes is about 1kW, and each node corresponds to a discrete control action. This action space is the output decision quantity of the subsequent energy management model.

[0017] Finally, system operation constraints are set to limit the feasible domain for subsequent energy management decisions. This scheme uniformly defines the power sign rule: the output power of the power battery is positive under discharge conditions and negative under charging conditions, consistent with the sign rule of the vehicle's required power. The following constraints apply to all driving and braking energy recovery conditions and specifically include four constraints: 1. Vehicle power balance constraint: satisfies the following relation ,in The coefficient of performance (COP) of the DC / DC converter is set to 0.95, matching the steady-state operating efficiency level of commonly used boost DC / DC converters in urban public transportation. To ensure the power output of the power battery, the output power of the dual power sources is matched with the power demand of the vehicle in real time; 2. Power battery SOC constraint: Meets requirements ,in The value is 0.6. The value is 0.8, which is used to limit the depth of charge and discharge of the power battery to avoid damage to the battery life caused by deep charge and discharge. 3. Fuel cell output power constraint: meets the requirements. ,in The value is set to 5kW. The value is set to 68kW to ensure that the fuel cell output is always within the rated operating range. 4. Power battery power constraint: Meets requirements ,in The maximum discharge power of the power battery is set to 180kW. The maximum charging power of the power battery is set to -90kW to match the power battery's power performance parameters and avoid overcharging and over-discharging.

[0018] The above four constraints are all hard constraints of the system. If the power command output by a single-step decision exceeds any constraint boundary, it is determined to be an inaction and must be truncated or directly eliminated according to the constraint boundary. Under extreme power demand, priority is given to ensuring the vehicle's drive power demand. For the braking energy recovery condition, the regenerative power is preferentially used to charge the power battery, and the fuel cell maintains a minimum output power of 5kW and does not participate in braking energy absorption; when the regenerative power exceeds the maximum charging power of the power battery, the excess is consumed through mechanical braking.

[0019] Step S20: Based on the defined state space and action space, construct an energy management model that includes a dual-network architecture of Q-network and target network and an experience playback mechanism; This model is based on a general deep Q-network (DQN) architecture and undergoes targeted modifications to provide architectural support for the subsequent integration of uncertainty protection modules and the extension of fuel cell lifetime constraints. Specifically, it includes the following sub-steps: Step S21: Construct a dual-network architecture that supports Q-value distribution output across all actions; The input layer of the Q network corresponds to the two-dimensional state variables defined in step S10, namely the vehicle power demand. The hidden layer is configured with a three-layer fully connected structure, with 128, 128, and 64 neurons respectively. The ReLU activation function is used to fit the nonlinear mapping relationship between state and action value under complex urban bus operating conditions. The number of neurons in the output layer corresponds one-to-one with the action space dimension defined in step S10, and each neuron outputs one fuel cell output power. The action value function value corresponding to the action forms a complete Q-value distribution vector of the whole action, and all output actions are strictly limited to the upper and lower limits of fuel cell output power set in step S10.

[0020] The dual-network architecture comprises a Q-network and a target network with identical structures: the Q-network outputs the Q-value distribution of all actions in the current state in real time, providing a computational basis for subsequent uncertainty quantification and performing ε-greedy action selection; the target network calculates the temporal difference target value, and its parameters are copied and updated from the Q-network every 200 training steps, rather than being synchronized in real time. This avoids model oscillations caused by frequent fluctuations in the target value during training and improves the convergence stability of policy training. The network parameters are optimized using the Adam optimizer, and the loss function is the mean squared error loss, used to fit the temporal difference error.

[0021] Step S22: Configure an extended experience playback mechanism with working condition tags; A 50,000-record experience replay buffer is set up. To address the needs of subsequent fuel cell lifespan penalty calculations, the standard DQN quad-tuple storage structure is expanded to a quintuple structure. Each experience record is stored sequentially: current system state, fuel cell output power action, instantaneous reward value, next-moment system state, and fuel cell operating condition label. The fuel cell operating condition label is divided into four categories: idling, start-stop, variable load, and heavy load, corresponding one-to-one with the aging conditions in the subsequent degradation model. When the buffer is full, new experiences automatically overwrite the oldest historical experiences, achieving data rolling updates.

[0022] During the model training phase, 64 experiences are randomly selected from the experience replay buffer each time to form a mini-batch sample for iterative updates of network parameters. Random sampling breaks the temporal correlation of experience data under continuous operation conditions of urban public transport, reducing the interference of sample autocorrelation on gradient updates. At the same time, the pre-stored operation condition labels in the experience can be directly called to calculate the lifetime penalty value, avoiding repeated operation condition identification during online decision-making and improving the computational efficiency of real-time decision-making. All actions and states corresponding to samples must meet the system operation constraints set in step S10. Samples outside the feasible region are directly removed to ensure the legality of training data.

[0023] Step S23: Design a multi-objective initial reward function and set training hyperparameters and exploration mechanisms for urban public transportation scenarios; The multi-objective initial reward function comprises three terms: equivalent hydrogen consumption, SOC deviation penalty, and power fluctuation penalty, expressed as follows: ,in , , All are positive weighting coefficients, corresponding to the optimization priorities of the equivalent hydrogen consumption term, the SOC deviation penalty term, and the power fluctuation penalty term, respectively. The three penalty terms achieve dimensional normalization and optimization priority configuration through weighting coefficients, and the weighted sum constitutes a unified real-time value evaluation index; this embodiment Take 1, Take 10, Set to 1; t represents the current time. The equivalent hydrogen consumption penalty term at time t, in kg / s, is calculated as follows: First, calculate the instantaneous hydrogen consumption of the fuel cell at time t. ,in The lower heating value of hydrogen is taken as [value]. , Let t be the efficiency of the fuel cell system at time t. It is obtained by looking up the table based on the current output power from the fuel cell efficiency MAP. The efficiency MAP has a parabolic characteristic, with a maximum efficiency of about 0.58, corresponding to an output power of about 35kW. The efficiency is about 0.45 under 5kW idle condition and about 0.52 under 68kW rated condition. Let t be the output power of the fuel cell at time t; Then calculate the equivalent hydrogen consumption of the power battery at time t. This is used to convert the change in electrical energy of the power battery into an equivalent amount of hydrogen consumption. In this embodiment, it is uniformly stipulated that the output power of the power battery is denoted as... The unit is kW. The value is positive under discharge conditions and negative under charging conditions. Its range is limited by the power battery power constraint set in step S10. Based on the differences in energy transfer paths during charging and discharging, a differential equivalence factor is used for conversion: when That is, when the power battery discharges, ;when That is, when the power battery is charging, ;in As the discharge equivalence factor, The charging equivalence factor is a constant, expressed in kg / (kJ), and is derived from the average efficiency of the fuel cell, the average efficiency of the DC / DC converter, and the average charging and discharging efficiency of the power battery. In this embodiment, the average efficiency of the fuel cell is taken as 0.55, the average efficiency of the DC / DC converter as 0.95, the average discharging efficiency of the power battery as 0.92, and the average charging efficiency as 0.90. The calculated values ​​are... , ; It is derived from the vehicle power balance constraint in step S10; Final equivalent hydrogen consumption penalty: .

[0024] The SOC deviation penalty term at time t is based on the squared deviation between the current SOC and the target SOC, constraining the fluctuation of the power battery SOC within the working range set in step S10. The calculation formula is as follows: ,in The state of charge (SOC) of the power battery at time t is calculated using the ampere-hour integration method. is the target SOC value, and is a constant, which is taken as 0.7 in this embodiment, corresponding to the midpoint of the optimal operating range of the power battery.

[0025] This is the power fluctuation penalty term at time t. This term is based on the absolute value of the change in fuel cell output power between adjacent times, suppressing frequent power jumps in the fuel cell and reducing the adverse effects of variable load conditions on fuel cell lifespan. The calculation formula is: ,in The output power of the fuel cell at time t-1 is expressed in kW. The power fluctuation penalty term is set to 0 at the initial time. During the training phase, the term can be weighted differently based on the operating condition labels of the experience samples. The penalty weight can be appropriately increased under start-stop and variable load conditions to strengthen the basic life protection.

[0026] To address the characteristics of frequent starts and stops and large power fluctuations in urban bus operation, the model training hyperparameters were configured as follows: a total of 3000 training epochs were set, with a learning rate of 0.0005 and a discount rate of 0.95. An ε-greedy mechanism was adopted for the exploration strategy, with an initial exploration rate of 0.9, which linearly decayed to a final exploration rate of 0.05 in the first 2000 epochs, and remained constant at 0.05 in subsequent epochs. In the early stages of training, the model ensured sufficient coverage of various bus operation conditions to avoid getting trapped in local optima, and gradually converged to a stable optimal power allocation strategy in the later stages of training. During training, all output actions were verified by the system constraints in step S10; actions exceeding the feasible region were directly penalized with low values ​​to ensure that the learned strategy always met the requirements for safe operation of the power system.

[0027] Step S24: Perform offline training until the model converges; Based on the aforementioned dual-network architecture, experience replay mechanism, and multi-objective reward function, the energy management model is trained offline iteratively. During training, the system operation constraints set in step S10 are used as boundaries, and an ε-greedy exploration strategy is employed for action sampling. The Q-network parameters are updated through backpropagation of temporal difference errors, while the target network synchronizes its parameters at fixed steps. A single training round corresponds to a complete urban bus cycle, containing 1200 time steps, each lasting 1 second. The model is considered converged when the cumulative reward fluctuation over 50 consecutive training rounds is less than 1% and the reward curve tends to stabilize. After training, the converged Q-network and target network parameters are output as the basic energy management strategy.

[0028] Step S30: Add an uncertainty protection module to the constructed energy management model to suppress instantaneous fluctuations in fuel cell output power; This step, based on the energy management model constructed in step S20, connects an uncertainty protection module to the decision output. It should be noted that the power fluctuation penalty in step S23, which is a globally fixed-weight basic smoothing constraint, is replaced by an online dynamic correction mechanism in this module. The two are hierarchically complementary, and there is no duplicate penalty. Addressing the technical shortcomings of existing technologies that rely solely on empirical rules for fixed power limiting and cannot achieve a dynamic quantitative optimal balance between fuel economy and fuel cell durability, the core approach is to construct a quantitative mapping model between operating condition uncertainty and fuel cell aging increment. This model dynamically solves for the optimal power output with the goal of minimizing the overall cost, suppressing sudden power fluctuations in fuel cell under high-uncertainty operating conditions while avoiding economic losses caused by excessive limiting under stable operating conditions. Specifically, it includes the following sub-steps: Step S31: Calculate the basic value of instantaneous uncertainty at the current moment using the full-action Q-value distribution output by the energy management model Q-network; There is a clear correspondence between operating condition uncertainty and Q-value distribution: when the operating condition is stable and power demand is highly predictable, the energy management model has high confidence in judging the value of actions, the Q-value distribution is concentrated, and the information entropy is low; when the operating condition fluctuates drastically and power demand is highly uncertain, the dispersion of action value increases, and the Q-value variance and information entropy increase simultaneously. Therefore, the uncertainty level of the current operating condition can be quantified by the statistical characteristics of the Q-values ​​of all actions. Suppose there are N possible actions in the current state, and the Q-value corresponding to the i-th action is... ( In this embodiment, N is 64, which corresponds one-to-one with the 64 power nodes uniformly discretely obtained in the range of [5kW, 68kW] in step S10, and the power interval between adjacent actions is about 1kW.

[0029] First, calculate the variance of the Q-value for the entire action. This characterizes the degree of dispersion of action value. In the formula The average Q-values ​​of all possible actions at the current moment are used; then, the softmax function is used to normalize the Q-values ​​of each action into a selection probability. Calculate information entropy Characterizing the confidence level of a decision: In the formula, This is a temperature coefficient used to adjust the smoothness of the probability distribution; in this embodiment, it is set to 1.

[0030] Finally, the variance and information entropy are normalized separately and then weighted and fused to obtain the basic value of instantaneous uncertainty. The value range is 0~2: In the formula, , The weighting coefficients for the two indicators are both set to 0.5 in this embodiment; , These are the variance of Q-values ​​and the average value of information entropy for all states under the global training condition, respectively. They are obtained by statistically analyzing all valid samples in the experience replay buffer after offline training converges, and are used as normalization benchmarks online without needing to be updated in real time.

[0031] Step S32: Map the current instantaneous uncertainty baseline value to the aging equivalent cost factor of the fuel cell; To address the shortcomings of existing fuel cell life protection technologies that rely on setting fixed aging penalty weights during offline training—which fail to dynamically adjust the protection level based on real-time decision confidence, leading to economic losses due to overprotection under stable operating conditions and accelerated aging due to insufficient protection under fluctuating conditions—this step constructs an aging equivalent cost factor using the uncertainty of the full-action Q-value distribution at the energy management model output as the driving variable. This is used to quantify the aging increment of fuel cells, directly transforming the uncertainty at the decision-making level into a lifespan cost that can be measured in unison with hydrogen consumption. The calculation formula is: ,in The baseline aging cost factor represents the equivalent hydrogen consumption per unit time for fuel cell aging under steady operating conditions, corresponding to unit power fluctuation. The unit is [unit missing]. This embodiment is based on the calibration of four types of aging coefficients in the fuel cell degradation model: idling, variable load, start-stop, and heavy load. The calibration process is as follows: Degradation rate data for the four operating conditions are obtained through accelerated aging tests of the fuel cell; the aging rate under unit power variable load conditions is converted into equivalent hydrogen consumption according to the energy equivalence principle; finally, the baseline aging cost factor is obtained, with a value of [value missing]. ; The aging sensitivity coefficient is used to adjust the intensity of the impact of uncertainty on the aging increment. In this embodiment, the value is set to 1.5 to ensure that the aging cost has a sufficient growth gradient when uncertainty increases, which can cover the fluctuation range of the entire working condition of urban public transport. The uncertainty benchmark threshold corresponds to the uncertainty level under stable operating conditions. In this embodiment, the value is 1.0, which is obtained through offline statistics of stable operating conditions of typical urban public transport.

[0032] This step does not require additional operating condition identification sensors and algorithms. The full-action Q-value distribution output by the Q network can be directly reused to generate dynamic aging weights online. When the decision confidence is high and the operating condition is stable, the aging cost weight automatically decreases to prioritize fuel economy. When the decision confidence is low and the operating condition fluctuates drastically, the aging cost weight increases exponentially, automatically strengthening the life protection and achieving an online dynamic optimal trade-off between economy and durability.

[0033] Step S33: Solve for the optimal output power of the fuel cell based on the aging equivalent cost factor; With the goal of minimizing the combined cost of "instantaneous equivalent hydrogen consumption + power fluctuation aging cost", the optimal output power at the current moment is dynamically solved within the feasible power range of the fuel cell, replacing the fixed power limit in the existing technology.

[0034] First, the fuel cell power corresponding to the maximum Q value from the Q-values ​​of the entire operation output by the Q-network is selected as the original optimal target power. This power is the theoretically optimal solution obtained from offline training of the energy management model, which balances long-term fuel economy and SOC maintenance, and serves as the economic benchmark for online correction; simultaneously, the actual output power at the previous moment is obtained. Define the output power at the current moment as The comprehensive cost function at that time is : ,in for The instantaneous equivalent hydrogen consumption term is obtained by reusing the equivalent hydrogen consumption calculation method in step S23, which characterizes the instantaneous hydrogen consumption cost under the current power, in units of ; The power fluctuation aging term is determined by the aging equivalent cost factor. With single-step power change The product of these components, with units uniformly expressed as kg / s, represents the equivalent cost of fuel cell lifespan degradation caused by power fluctuations. For strategy deviation penalty terms, The policy deviation penalty coefficient is a constant, and its unit is 1000 kJ / m². The value in this embodiment is... This is used to constrain the deviation between the power after online correction and the original optimal power, to avoid a significant deviation from the global optimal trajectory obtained by model training due to excessive pursuit of power smoothness, and to prevent long-term SOC imbalance and increase in global hydrogen consumption.

[0035] Before iterating through the solution, the feasibility of all candidate power outputs needs to be verified: the corresponding power battery power and SOC change are derived through the vehicle power balance constraint. If the change exceeds the SOC working range or power battery power constraint range set in step S10, the candidate power is directly eliminated. Then, the feasible range of fuel cell output power defined in step S10 is determined. Within this process, iterate through the power values ​​corresponding to all possible actions to solve for the overall cost. When the minimum value is obtained This serves as the final power command issued. The technical advantage of this solution mechanism is that when the uncertainty of the operating conditions is low, Approaching the baseline value, the aging cost has a lower weighting, and the overall cost is mainly determined by hydrogen consumption and strategy consistency. The output power is close to the original optimal value, ensuring that fuel economy is not compromised. However, when operating condition uncertainty increases... As the power fluctuations increase exponentially, the aging cost weight increases rapidly, and the output power automatically smooths around the original optimal value. The higher the uncertainty, the stronger the smoothing force, achieving a dynamic optimal balance between economy, strategy consistency and durability.

[0036] Step S34: Integrate the solution logic of steps S31 to S33 into the output of the Q network to form an uncertainty protection module; After integration, the complete online decision-making process of the energy management model is as follows: 1. Collect the current vehicle power demand and power battery SOC in real time and input them into the Q network; 2. The Q-network outputs the Q-values ​​for all possible actions in the current state; 3. Calculate the basic value of instantaneous uncertainty according to step S31. ; 4. Calculate the aging equivalent cost factor at the current moment according to step S32. ; 5. Solve for the optimal output power with the minimum overall cost according to step S33, and send the solution to the power system for execution; 6. Wait for the next decision cycle and repeat the above process to achieve rolling online energy management decisions.

[0037] The uncertainty protection module only operates during the online decision-making stage and does not modify the network parameters and training process of the energy management model during offline training, resulting in low engineering implementation costs.

[0038] Step S40: Deploy the energy management model on the vehicle controller, collect real-time vehicle operating status parameters and input them into the model, output the optimal power allocation command between the fuel cell and the power battery, and complete the vehicle energy management decision. The energy management model, which has been trained and converged offline, is integrated with the uncertainty protection module and deployed in the vehicle controller. Taking the system state variables defined in step S10 as input, the system outputs a dual-power source power allocation command that satisfies the system operation constraints after network inference, uncertainty correction, and constraint verification, thereby realizing real-time closed-loop energy management decision-making for the hydrogen fuel cell bus. The specific steps include the following: Step S41: Onboard deployment and parameter solidification of the energy management model; The energy management model trained offline to convergence in step S20, along with the uncertainty protection module calibrated in step S30, are deployed to the vehicle control unit (VCU); simultaneously, the global Q-value variance benchmark value is... Global information entropy benchmark value Baseline aging cost factor Aging sensitivity coefficient Uncertainty benchmark threshold Strategy deviation penalty coefficient Once all calibration parameters are stored in the controller's storage unit, online real-time calibration is no longer required.

[0039] In this embodiment, the DQN network is a three-layer fully connected lightweight structure with a single-step inference computation of less than 100,000 floating-point operations, adapting to the computing power level of the on-board controller; the decision cycle is set to 100ms to match the power change rate of urban public transport conditions. After deployment, the model and protection module execute the energy management decision logic cyclically at a fixed period.

[0040] Step S42: Real-time vehicle operating status parameter acquisition and standardized preprocessing; At the beginning of each decision cycle, the vehicle controller collects real-time vehicle operation data, which is then filtered and preprocessed to obtain the two-dimensional state variables defined in step S10, which serve as the input for subsequent model inference.

[0041] Among them, the power required for the whole vehicle The target drive torque is obtained by analyzing the opening of the driver's accelerator and brake pedals, and calculated in conjunction with the current vehicle speed, covering both drive and regenerative braking conditions; the state of charge of the power battery. The data is collected and uploaded in real time by the Battery Management System (BMS). The raw acquired signal undergoes first-order low-pass filtering to eliminate sensor noise and transient interference. The preprocessed state parameters are strictly limited to the range set in step S10 to ensure the validity of the input.

[0042] Step S43: Generation of power allocation instructions for dual power sources and verification of system constraints; The real-time state variables obtained from the preprocessing in step S42 are input into the deployed energy management model. The Q-network outputs the Q-value distribution corresponding to all possible actions in the current state. Then, the uncertainty protection module is called to complete the power correction solution, and the optimal output power of the fuel cell with the minimum overall cost is obtained. ; The power battery output power is derived by reusing the vehicle power balance constraints from step S10, and the system operation constraints are verified item by item to generate a valid dual power source power allocation command. First, calculate the target output power of the power battery based on the vehicle power balance formula. ; Subsequently, the fuel cell output power constraint, power battery power constraint, four system constraints, power battery SOC constraint, and vehicle power balance constraint were verified sequentially: If any constraint is not met, the power command is truncated and corrected within the constraint boundary to ensure that the issued command is always within the safe operating range of the system.

[0043] Step S44: Power command issuance and closed-loop execution of energy management decisions; The verified target power command for the fuel cell is sent to the fuel cell DC / DC converter to control the fuel cell to output power according to the target power; the target power command for the power battery is sent to the battery management system to control the power battery to perform the corresponding discharge / charge action, thus completing a single energy management decision.

[0044] At the same time, the actual output power of the fuel cell at the current moment will be... The data is stored in the controller's cache unit as the reference value for the previous moment in the power fluctuation calculation of the next decision cycle, forming a complete closed loop of "acquisition-inference-correction-verification-execution-feedback". The vehicle controller repeats the above process at a fixed period of 100ms to realize real-time energy management throughout the vehicle's operation.

[0045] Example 2 like Figure 2 As shown, Embodiment 2 of this application provides an energy management decision system for hydrogen fuel cell buses, including: a space definition module 21, a management model training module 22, and an instruction generation and issuance module 23; Space definition module 21 is used to define the state space and action space of the hybrid power system; The management model training module 22 is used to construct an energy management model based on the defined state space and action space, which includes a dual network architecture of Q network and target network and an experience playback mechanism, and to add an uncertainty protection module to the constructed energy management model. The instruction generation and distribution module 23 is used to deploy the energy management model on the vehicle controller, collect real-time vehicle operating status parameters into the model, output the optimal power allocation instruction for fuel cells and power batteries, and complete the vehicle energy management decision.

[0046] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A decision-making method for energy management of hydrogen fuel cell buses, characterized in that, include: Step S10: Define the state space and action space of the hybrid power system; the state space includes the vehicle's required power and the state of charge of the power battery, and the action space is the output power of the fuel cell; Step S20: Based on the defined state space and action space, construct an energy management model that includes a dual-network architecture of Q-network and target network and an experience playback mechanism; Step S30: Add an uncertainty protection module to the constructed energy management model to suppress instantaneous fluctuations in fuel cell output power; adding an uncertainty protection module to the constructed energy management model is specifically divided into the following sub-steps: Step S31: Calculate the basic value of instantaneous uncertainty at the current moment using the full-action Q-value distribution output by the energy management model Q-network; Step S32: Map the current instantaneous uncertainty baseline value to the aging equivalent cost factor of the fuel cell; Step S33: Solve for the optimal output power of the fuel cell based on the aging equivalent cost factor; Step S34: Integrate the solution logic of steps S31 to S33 into the output of the Q network to form an uncertainty protection module; Step S40: Deploy the energy management model on the vehicle controller, collect real-time vehicle operating status parameters into the model, output the optimal power allocation command for the fuel cell and power battery, and complete the vehicle energy management decision.

2. The energy management decision-making method for hydrogen fuel cell buses according to claim 1, characterized in that, Based on the defined state space and action space, an energy management model is constructed, which includes a dual-network architecture of Q-network and target network and an experience playback mechanism. The specific steps are as follows: Step S21: Construct a dual-network architecture that supports Q-value distribution output across all actions; Step S22: Configure an extended experience playback mechanism with working condition tags; Step S23: Design a multi-objective initial reward function and set training hyperparameters and exploration mechanisms for urban public transportation scenarios; Step S24: Perform offline training until the model converges.

3. The energy management decision-making method for hydrogen fuel cell buses according to claim 1, characterized in that, The complete online decision-making process of the energy management model is as follows: Real-time data collection of the vehicle's required power and the battery's state of charge (SOC) is input into the Q network. The Q-network outputs the Q-values ​​for all possible actions in the current state. Calculate the basic value of instantaneous uncertainty according to step S31. ; Calculate the aging equivalent cost factor at the current moment according to step S32. ; The optimal output power with the minimum overall cost is determined in step S33 and then sent to the power system for execution. Waiting for the next decision cycle, the above process is repeated to achieve rolling online energy management decisions.

4. The energy management decision-making method for hydrogen fuel cell buses according to claim 1, characterized in that, The energy management model is deployed on the vehicle controller, and real-time vehicle operating status parameters are collected and input into the model. The model outputs the optimal power allocation command between the fuel cell and the power battery to complete the vehicle energy management decision. The specific steps include the following: Step S41: Onboard deployment and parameter solidification of the energy management model; Step S42: Real-time vehicle operating status parameter acquisition and standardized preprocessing; Step S43: Generation of power allocation instructions for dual power sources and verification of system constraints; Step S44: Power command issuance and energy management decision-making closed-loop execution.

5. A hydrogen fuel cell bus energy management decision-making system, executing the hydrogen fuel cell bus energy management decision-making method as described in any one of claims 1-4, characterized in that, The system includes: a space definition module, a management model training module, and an instruction generation and distribution module; The space definition module is used to define the state space and action space of the hybrid power system. The management model training module is used to construct an energy management model based on a defined state space and action space, which includes a dual-network architecture of Q-network and target network and an experience playback mechanism, and adds an uncertainty protection module to the constructed energy management model. The instruction generation and distribution module is used to deploy the energy management model on the vehicle controller, collect real-time vehicle operating status parameters into the model, output the optimal power allocation instruction for fuel cells and power batteries, and complete the vehicle energy management decision.

Images