Life cycle-oriented self-adjusting energy management method and system for fuel cell heavy truck

Abstract

This invention discloses a self-regulating energy management method and system for fuel cell heavy-duty trucks throughout their entire lifecycle, relating to the field of fuel cell energy management technology. The method includes: constructing and training a comprehensive energy management model; fitting the lifecycle efficiency curve and open-circuit voltage characteristic curve through a system characteristic identification module; offline training and transfer learning of a near-end policy optimization algorithm using a reinforcement learning decision module to obtain a transferred near-end policy network; then, constraint optimization using nonlinear model predictive control by a rolling optimization correction module to generate energy control commands; and finally, real-time acquisition of vehicle operation data input into the model to achieve adaptive adjustment of power distribution between the fuel cell and the power battery. This invention achieves adaptive adjustment of power distribution between the fuel cell and the power battery throughout their entire lifecycle by constructing a comprehensive energy management model including system characteristic identification, reinforcement learning decision-making, and rolling optimization correction.

Description

Technical Field

[0001] This invention relates to the field of fuel cell energy management technology, specifically to a self-regulating energy management method and system for fuel cell heavy-duty trucks throughout their entire lifecycle. Background Technology

[0002] Hydrogen fuel cell hybrid power systems, with their high energy density and long driving range, are more suitable for commercial heavy-duty trucks with heavy loads and long distances. However, under varying vehicle operating conditions, the output characteristics of both the fuel cell system and the power battery degrade as the vehicle's operating time and mileage increase. This continuous degradation of system characteristics leads to a gradual decrease in the optimization effectiveness of traditional energy management strategies trained based on factory characteristics. Model mismatch will further worsen the optimization effect and may even accelerate system degradation.

[0003] Currently, energy management strategies for fuel cell hybrid vehicles are mainly categorized into rule-based methods, optimization-based methods, and learning-based methods. Rule-based methods rely on expert experience and struggle to adapt to complex and variable operating conditions; optimization-based methods involve high computational costs and are limited in real-time performance; while reinforcement learning-based methods possess self-learning capabilities, traditional deep reinforcement learning energy management strategies suffer from low computational efficiency and poor robustness, especially lacking effective adaptive adjustment mechanisms after the system's characteristics degrade throughout its lifecycle. Therefore, there is an urgent need for a self-adjusting energy management method for fuel cell heavy-duty trucks that addresses the entire lifecycle, thereby improving the overall lifecycle performance of fuel cell hybrid systems under varying performance scenarios. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a self-regulating energy management method and system for fuel cell heavy-duty trucks throughout their entire lifecycle. By constructing a comprehensive energy management model that includes system characteristic identification, reinforcement learning decision-making, and rolling optimization correction, the method achieves adaptive adjustment of power distribution between the fuel cell and the power battery throughout the entire lifecycle.

[0005] On the one hand, a self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle includes:

[0006] S1, acquire historical vehicle operation data; the historical vehicle operation data includes current and voltage data of fuel cell stack under different aging conditions, state of charge and open circuit voltage data of power battery under different aging conditions, and WTVC operating condition data.

[0007] S2, Construct and train a comprehensive energy management model; the comprehensive energy management model includes a system characteristic identification module, a reinforcement learning decision-making module, and a rolling optimization and correction module;

[0008] The system characteristic identification module uses a genetic algorithm to perform degradation parameter fitting on the current and voltage data of the fuel cell stack under different aging states, and outputs the full life cycle fuel cell system efficiency curve. It also uses the least squares method to perform polynomial fitting on the state of charge and open circuit voltage data of the power battery under different aging states, and outputs the full life cycle power battery open circuit voltage characteristic curve.

[0009] The reinforcement learning decision-making module constructs a reinforcement learning interactive environment based on the efficiency curve of the fuel cell system and the open-circuit voltage characteristic curve of the power battery throughout the entire life cycle. Taking the WTVC operating condition data as the state space input, the module uses the proximal policy optimization algorithm to train the policy network offline in the reinforcement learning interactive environment, outputs the trained policy network, and performs transfer training on the trained policy network to obtain a transfer proximal policy network for the entire life cycle. The transfer proximal policy network is used to calculate the fuel cell power change rate based on the WTVC operating condition data.

[0010] The rolling optimization correction module uses the fuel cell power change rate as the initial value of the nonlinear model predictive control solver through a migration near-end policy network oriented towards the entire life cycle, and generates energy control commands based on the constrained fuel cell power change rate.

[0011] S3 inputs real-time collected vehicle operation data into a trained comprehensive energy management model to obtain energy control commands, and adjusts the power distribution between the fuel cell and the power battery in real time based on the energy control commands.

[0012] Furthermore, the state space includes vehicle speed, acceleration, fuel cell power, and power battery state of charge; wherein, vehicle speed and acceleration are used to reflect the current operating conditions of the vehicle, and fuel cell power and power battery state of charge are used to reflect the current power distribution state of the vehicle.

[0013] Furthermore, the fuel cell power is obtained by subtracting the power of the power battery from the total vehicle power requirement;

[0014] The formula for calculating the required power of the whole vehicle is as follows:

[0015] ;

[0016] in, Power required for the entire vehicle; For the mechanical efficiency of the transmission system; For motor efficiency; Let g be the mass of the vehicle; g is the gravitational constant. This is the rolling resistance coefficient; For vehicle speed; This is the rotational mass conversion factor; To accelerate the vehicle; This refers to the air drag coefficient; The vehicle's frontal area; This refers to the road slope.

[0017] Furthermore, the formula for calculating the state of charge of the power battery is as follows:

[0018] ;

[0019] ;

[0020] in, This refers to the open-circuit voltage of the power battery. Internal resistance during discharge / charge of the power battery; Power of the battery; The initial state of charge; For power battery capacity; This refers to the current of the power battery. This indicates the current state of charge of the power battery.

[0021] Furthermore, in S2, a genetic algorithm is used to fit the degradation parameters of the current and voltage data of the fuel cell stack under different aging conditions. The calculation formula is as follows:

[0022] ;

[0023] in, For ohmic resistance; i st Current density; For current; This is the maximum current of the fuel cell stack; and The parameters to be identified; This represents the polarization characteristic curve of a fuel cell;

[0024] The ratio of net output power to hydrogen energy at each moment is calculated based on the polarization curve of the fuel cell, and then connected in series according to the life process to obtain the efficiency curve of the fuel cell system throughout its entire life cycle.

[0025] Furthermore, in S2, during the offline training of the policy network using the proximal policy optimization algorithm in the reinforcement learning interactive environment, the objective function of the policy network is... The calculation formula is as follows:

[0026] ;

[0027] in, This is the cutting factor; Indicates network parameters; The dominant function; This represents the state value at time t; This represents the action value at time t; The parameter is The old strategy network; The parameter is New strategy network; Indicates the pruning mechanism; Represents the mathematical expectation; This represents the boundary of the trajectory distribution state under the guidance of the strategy; This indicates taking the minimum value.

[0028] Furthermore, in S2, the objective function calculation formula for the nonlinear model predictive control solver is as follows:

[0029] ;

[0030] in, The cost of hydrogen consumption for fuel cells; Cost of fuel cell degradation; The equivalent hydrogen consumption cost for power batteries; Costs associated with battery degradation; This is the SOC adjustment factor; Indicates the current state of charge of the power battery; This indicates the target SOC. On the other hand, a self-regulating energy management system for fuel cell heavy-duty trucks throughout their entire lifecycle includes:

[0031] The data acquisition module is used to acquire historical vehicle operation data, which includes current and voltage data of fuel cell stack under different aging conditions, state of charge and open circuit voltage data of power battery under different aging conditions, and WTVC operating condition data.

[0032] The model training module is used to construct and train a comprehensive energy management model. This comprehensive energy management model includes a system characteristic identification module, a reinforcement learning decision-making module, and a rolling optimization and correction module. The system characteristic identification module uses a genetic algorithm to perform degradation parameter fitting on the current and voltage data of the fuel cell stack under different aging states, outputting the full lifecycle fuel cell system efficiency curve. It also uses the least squares method to perform polynomial fitting on the state of charge and open-circuit voltage data of the power battery under different aging states, outputting the full lifecycle power battery open-circuit voltage characteristic curve. The reinforcement learning decision-making module, based on the full lifecycle fuel cell system efficiency curve and the full lifecycle power battery open-circuit voltage characteristic curve... A reinforcement learning interactive environment is constructed using the pressure characteristic curve. WTVC operating data is used as the state space input. In this environment, a proximal policy optimization algorithm is employed to train the policy network offline. The trained policy network is then output, and transfer training is performed to obtain a transfer proximal policy network oriented towards the entire lifecycle. This transfer proximal policy network is used to calculate the fuel cell power change rate based on the WTVC operating data. The rolling optimization correction module uses the fuel cell power change rate as the initial value of the nonlinear model predictive control solver through the transfer proximal policy network oriented towards the entire lifecycle, and generates energy control commands based on the constrained fuel cell power change rate.

[0033] The power distribution module is used to input real-time collected vehicle operation data into a trained comprehensive energy management model to obtain energy control commands, and adjust the power distribution between the fuel cell and the power battery in real time based on the energy control commands.

[0034] The present invention adopts the above technical solution and has the following beneficial effects:

[0035] (1) The present invention fits the full life cycle fuel cell efficiency curve and the power battery open circuit voltage characteristic curve through the system characteristic identification module, which can accurately characterize the aging process of the stack and battery, so that the energy management strategy can adapt to different life stages and improve the accuracy of full life cycle control.

[0036] (2) The present invention obtains a transfer network by offline training and transfer using a near-end strategy optimization algorithm, combined with rolling optimization correction of nonlinear model predictive control, taking into account both global optimal decision and real-time constraint processing, thereby reducing the online computational burden and ensuring control accuracy under dynamic conditions.

[0037] (3) This invention achieves multi-objective collaborative optimization by comprehensively balancing hydrogen consumption cost, fuel cell and power battery degradation cost and state of charge adjustment term in objective function, effectively extending the life of key components and reducing the total life cycle operating cost while ensuring the power performance of the vehicle. Attached Figure Description

[0038] Figure 1 This is an overall architecture diagram of a fuel cell heavy-duty truck self-regulating energy management method for the entire life cycle, as described in an embodiment of the present invention.

[0039] Figure 2 This is a flowchart of the self-regulating energy management method for fuel cell heavy-duty trucks throughout the entire life cycle, according to an embodiment of the present invention.

[0040] Figure 3 This is the result of identifying the degradation characteristics of the fuel cell system in the embodiments of the present invention;

[0041] Figure 4 This is the identification result of the degradation characteristics of the power battery in the embodiment of the present invention;

[0042] Figure 5 This is the training result of the PPO policy for transfer reinforcement learning oriented towards the entire life cycle in the embodiments of the present invention;

[0043] Figure 6 This is the energy management result under the simulated operating condition of 50% remaining lifespan of the power system in the embodiments of the present invention;

[0044] Figure 7 This is a diagram of a self-regulating energy management system for fuel cell heavy-duty trucks oriented towards the entire life cycle, according to an embodiment of the present invention. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0046] Specifically, Figure 1This paper presents the overall technical framework for a self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle. The core process includes: First, collecting historical operating data and identifying fuel cell stack degradation parameters and open-circuit voltage characteristics of the power battery using genetic algorithms and least squares methods to fit the full lifecycle efficiency curve (i.e., calculating the ratio of net output power to hydrogen energy at each moment based on the fuel cell polarization curve and connecting them in series according to the lifecycle process to obtain the full lifecycle fuel cell system efficiency curve); Second, constructing a reinforcement learning virtual environment, training and transferring the policy network using a near-end policy optimization algorithm to generate a fuel cell power change rate action space (i.e., solving for the output action, i.e., the fuel cell power change rate, by setting the state space, action space, objective function, and reward function of reinforcement learning); Finally, using this action space as the initial solution for nonlinear model predictive control, and continuously optimizing the objective function under the constraints of system equality and inequality, outputting real-time energy control commands, thus forming a closed-loop management mechanism of data-driven, offline training, and online correction (using Sequential Quadratic Programming). The SQP algorithm solves the optimization problem of the objective function of a nonlinear model predictive control. The SQP algorithm transforms the complex nonlinear programming problem into a simple quadratic programming problem, i.e., an optimization problem where the objective function is a quadratic function and the constraint functions are linear functions. Combining the linearization of the nonlinear constraints, the quadratic programming subproblem can be obtained, as shown in the following equation:

[0047] ;

[0048] Among them, H k For the positive definite approximation of the Hessian matrix of the Lagrange function, f N d represents the characteristic function of the nonlinear system; d represents the disturbance variable. This represents the control variable during the k-th iteration.

[0049] like Figure 2 As shown, the present invention relates to a self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle, comprising:

[0050] S1, acquire historical vehicle operation data; the historical vehicle operation data includes current and voltage data of fuel cell stack under different aging states, state of charge and open circuit voltage data of power battery under different aging states, and WTVC (World Transient Vehicle Cycle) operating condition data.

[0051] S2, Construct and train a comprehensive energy management model; the comprehensive energy management model includes a system characteristic identification module, a reinforcement learning decision-making module, and a rolling optimization and correction module;

[0052] The system characteristic identification module uses a genetic algorithm to perform degradation parameter fitting on the current and voltage data of the fuel cell stack under different aging states, and outputs the full life cycle fuel cell system efficiency curve. It also uses the least squares method to perform polynomial fitting on the state of charge and open circuit voltage data of the power battery under different aging states, and outputs the full life cycle power battery open circuit voltage characteristic curve.

[0053] The reinforcement learning decision-making module constructs a reinforcement learning interactive environment based on the efficiency curve of the fuel cell system and the open-circuit voltage characteristic curve of the power battery throughout its entire life cycle. (Reinforcement learning (RL) learns the mapping relationship from the state space to the optimal action through a virtual environment to maximize global reward. The trained agent needs to continuously interact with state and action data in the virtual environment and optimize the mapping relationship from state to action through the feedback reward given by the virtual environment, so that the evaluation and policy neural network results converge.) Using WTVC operating condition data as the state space input, the policy network is trained offline in the reinforcement learning interactive environment using a proximal policy optimization algorithm. The trained policy network is then transferred to obtain a proximal policy network for the entire life cycle. The proximal policy network is used to calculate the fuel cell power change rate based on the WTVC operating condition data. (Specifically, the WTVC operating conditions, such as vehicle speed, acceleration, the current fuel cell system output power, and the power battery state of charge, are used as the state space, input into the proximal policy optimization network, and then the objective function and reward function are set to output the fuel cell power change rate.)

[0054] The rolling optimization correction module uses a life-cycle-oriented migration proximal policy network to constrain the fuel cell power change rate action space as the initial value for the nonlinear model predictive control solver. Based on the constrained fuel cell power change rate, energy control commands are generated. (The proximal policy optimization network outputs the fuel cell power change rate, which is then directly used as the initial value for nonlinear model predictive control to improve the slow solution speed problem in traditional nonlinear model predictive control. We then selected constrained nonlinear model predictive control, whose solution process is based on the set constraints.)

[0055] The state space includes vehicle speed, acceleration, fuel cell power, and power battery state of charge; wherein, vehicle speed and acceleration are used to reflect the current operating conditions of the vehicle, and fuel cell power and power battery state of charge are used to reflect the current power distribution state of the vehicle.

[0056] Specifically, the fuel cell power is obtained by subtracting the power of the power battery from the total vehicle power requirement;

[0057] The formula for calculating the required power of the whole vehicle is as follows:

[0058] ;

[0059] in, Power required for the entire vehicle; For the mechanical efficiency of the transmission system; For motor efficiency; Let g be the mass of the vehicle; g is the gravitational constant. This is the rolling resistance coefficient; For vehicle speed; This is the rotational mass conversion factor; To accelerate the vehicle; This refers to the air drag coefficient; The frontal area of ​​the vehicle; This refers to the road slope.

[0060] Specifically, the formula for calculating the state of charge of the power battery is as follows:

[0061] ;

[0062] ;

[0063] in, This refers to the open-circuit voltage of the power battery. The internal resistance of the power battery during discharge / charge; Power of the battery; The initial state of charge; For power battery capacity; This refers to the current of the power battery. This indicates the current state of charge of the power battery.

[0064] Specifically, a genetic algorithm is used to fit degradation parameters to the current and voltage data of the fuel cell stack under different aging conditions. The calculation formula is as follows:

[0065] ;

[0066] in, For ohmic resistance; i st Current density; For current; This is the maximum current of the fuel cell stack; and The parameters to be identified; This represents the polarization characteristic curve of a fuel cell;

[0067] The ratio of net output power to hydrogen energy at each moment is calculated based on the polarization curve of the fuel cell, and then connected in series according to the life process to obtain the efficiency curve of the fuel cell system throughout its entire life cycle.

[0068] Specifically, in the offline training of the policy network using the proximal policy optimization algorithm in a reinforcement learning interactive environment, the objective function of the policy network is... The calculation formula is as follows:

[0069] ;

[0070] in, This is the cutting factor; Indicates network parameters; The dominant function; This represents the state value at time t; This represents the action value at time t; The parameter is The old strategy network; The parameter is New strategy network; Indicates the pruning mechanism; Represents the mathematical expectation; This represents the boundary of the trajectory distribution state under the guidance of the strategy; This indicates taking the minimum value.

[0071] Specifically, the objective function calculation formula for the nonlinear model predictive control solver is as follows:

[0072] ;

[0073] in, The cost of hydrogen consumption for fuel cells; Cost of fuel cell degradation; The equivalent hydrogen consumption cost for power batteries; Costs associated with battery degradation; This is the SOC adjustment factor; Indicates the current state of charge of the power battery; Indicates the target SOC.

[0074] S3 inputs real-time collected vehicle operation data into a trained comprehensive energy management model to obtain energy control commands, and adjusts the power distribution between the fuel cell and the power battery in real time based on the energy control commands.

[0075] Specifically, in this embodiment, the comprehensive energy management model can be further subdivided into a longitudinal dynamics model of the fuel cell heavy-duty truck, an output characteristic model and durability model of the fuel cell system, and an output characteristic model and durability model of the power battery. Based on the above subdivided models, the process of this method can be summarized as follows:

[0076] Specifically, the ideal power requirement of the vehicle is calculated based on the rolling resistance, air resistance, acceleration resistance, and gradient resistance overcome during vehicle operation.

[0077] (1)

[0078] In the formula, P req For the required power of the whole vehicle, η m For the mechanical efficiency of the transmission system, η mot For motor efficiency, m v Let g be the vehicle mass, f be the gravitational constant, and f be the rolling resistance coefficient. v Let be the vehicle speed, δ be the rotational mass conversion factor, and a be the speed. v For vehicle acceleration, C D A is the air drag coefficient. v β represents the vehicle's frontal area, and β represents the road slope.

[0079] Specifically, in the output characteristic model of a fuel cell system, the voltage of a single cell is determined by activation loss, ohmic loss, and concentration loss.

[0080] (2)

[0081] In the formula, V cell For the voltage of a single unit, E Nerst Let V be the Nernst voltage. act To activate voltage loss, V ohm For ohmic voltage loss, V con For concentration voltage loss, T st P is the temperature of the fuel cell stack. H2 and P O2 I represents the partial pressure of hydrogen and oxygen within the fuel cell stack. st It is the fuel cell current.

[0082] Specifically, the hydrogen consumption cost of a fuel cell system is calculated from the amount of hydrogen consumed and the price of hydrogen:

[0083] (3)

[0084] In the formula, η fc For the efficiency of the fuel cell system, P fc For the output power of the fuel cell system, LHV H2 It is the lower heating value of hydrogen, p H2 That is the price per kilogram of hydrogen.

[0085] Specifically, in the fuel cell system durability model, the voltage degradation rate is calculated as follows:

[0086] (4)

[0087] In the formula, κlow T represents the decay rate under low power conditions. low For the duration of low power conditions, κ high T represents the decay rate under high power conditions. high It is the duration of high-power conditions, κ cha Let ΔP be the decay rate under variable load conditions. fc It is the slope of the fuel cell system's output power.

[0088] A fuel cell system is considered to have reached its lifespan limit when its voltage at rated power drops by more than 10% compared to its factory specifications. The cost of fuel cell system performance degradation is calculated as follows:

[0089] (5)

[0090] In the formula, k fc V is the lifespan correction factor for the fuel cell system. fc,end P is the voltage drop at the end of the system's lifespan. fc,rate p is the rated power of the fuel cell system. fc This is the price per kilowatt of fuel cell system.

[0091] Specifically, in the power battery output characteristic model, the dynamic change of SOC is described as follows:

[0092] (6)

[0093] (7)

[0094] In the formula, V opc R is the open-circuit voltage of the power battery. d / c P is the internal resistance of the power battery during discharge / charge. bat For power battery power, SOC ini The initial state of charge (SOC) is Q. bat For the power battery capacity, I bat This refers to the current of the power battery.

[0095] Specifically, the calculation of power battery capacity degradation is as follows:

[0096] (8)

[0097] (9)

[0098] In the formula, ΔQ is the percentage of capacity decay relative to the initial capacity, and a c α is the capacity degradation coefficient of the power battery. c β c γ cκ, ε, z, and S0 are dimensionless constant coefficients of the model, and E ac To activate energy, R g Let T be the ideal gas constant, T be the battery temperature, Ah be the cumulative ampere-hour throughput, and M be the Mn. bat The operating mode ratio of the power battery, SOC min This is the minimum SOC value when the power battery is working.

[0099] A 20% decrease in battery capacity is considered the end of the battery's lifespan. The cost of battery system performance degradation is calculated as follows:

[0100] (10)

[0101] In the formula, N bat C represents the total number of cycles of the power battery. bat For the power battery capacity, p bat This refers to the unit price per kilowatt-hour of the power battery.

[0102] The degradation characteristics parameters of the polarization curve of the fuel cell system are identified based on the genetic algorithm, and the degradation characteristics parameters of the open-circuit voltage of the power battery are identified based on the least squares method. The efficiency curve of the fuel cell system and the open-circuit voltage characteristic curve of the power battery are updated throughout the entire life cycle.

[0103] Specifically, in the actual identification process, combining the Tafel empirical formula, Taylor expansion, and empirical parameters for fuel cell system control, the fuel cell cell voltage is simplified as follows:

[0104] (11)

[0105] In the formula, R ohm For ohmic resistance, i st It is the current density, I max ϵ1 and ϵ2 are the maximum current of the fuel cell stack, and ϵ1 and ϵ2 are the parameters to be identified.

[0106] A genetic algorithm is used to identify the aging parameters (ϵ1, ϵ2, R) of polarization characteristic data. ohm and I max The polarization curve of the fuel cell system is fitted. Based on the identified polarization curve model of the fuel cell system, the system efficiency curve under aging conditions is calculated.

[0107] (12)

[0108] In the formula, P st P represents the output power of the fuel cell stack. aux Parasitic power of auxiliary equipment.

[0109] Specifically, the least squares method is used to identify the open-circuit voltage degradation characteristic parameters of the power battery, and the open-circuit voltage model parameters are identified based on polynomial equations:

[0110] (13)

[0111] Where a0 to a5 are the polynomial coefficients to be identified.

[0112] Specifically, the internal resistance estimation model for power batteries is as follows:

[0113] (14)

[0114] (15)

[0115] In the formula, ΔR int a represents the increase in the internal resistance of the power battery. R α is the internal resistance degradation coefficient of the power battery. R β R γ R d R e R C0 represents the dimensionless constant coefficients of the model, and E aR To activate energy, C rate This refers to the discharge rate of the power battery.

[0116] A deep reinforcement learning energy management strategy is constructed based on the PPO algorithm. The state space is defined by vehicle speed, acceleration, fuel cell system output power, and power battery state of charge. The action space is defined by the slope of the fuel cell system output power. The offline training is performed with an instantaneous reward function that includes comprehensive energy cost, system degradation cost, and SOC maintenance penalty.

[0117] Specifically, the state space and action space of the PPO algorithm are as follows:

[0118] (16)

[0119] (17)

[0120] Specifically, the objective function of the policy network in the PPO algorithm is:

[0121] (18)

[0122] In the formula, ε is the cutting coefficient, and A θ′ This is the dominant function.

[0123] Specifically, the network loss function is as follows:

[0124] (19)

[0125] Specifically, the instant reward function is as follows:

[0126] (20)

[0127] In the formula, C obj This represents the comprehensive cost of the powertrain system, including energy cost and degradation cost. λ is the comprehensive cost weighting coefficient, ω is the SOC maintenance penalty weighting coefficient, and σ is the SOC final value penalty weighting coefficient. ref This is the SOC reference value.

[0128] The fuel cell system actions under the trained PPO strategy are used as the initial values ​​for the nonlinear MPC solver. The actions are then optimized and adjusted in combination with the actual system constraints to output the slope of the fuel cell system output power.

[0129] Specifically, the nonlinear MPC uses the slope of the fuel cell system output power as the control variable u, the power battery SOC and the fuel cell system output power as the state variable x, and the vehicle demand power as the disturbance variable d.

[0130] Specifically, the objective function of nonlinear MPC is:

[0131] (twenty one)

[0132] The constraints are:

[0133] ;

[0134] In the formula, C fc,H2 For the cost of hydrogen consumption in fuel cells, C fc,degr For fuel cell degradation costs, C bat,H2 For the equivalent hydrogen consumption cost of the power battery, C bat,degr λsoc represents the degradation cost of the power battery, and SOC adjustment coefficient.

[0135] Based on historical data from real vehicle operation, the periodic degradation characteristics of fuel cell systems and power batteries are identified, and the system response prediction model parameters and optimization constraints in the nonlinear MPC strategy are updated.

[0136] Specifically, to maintain the accuracy of the full lifecycle prediction, the fuel cell efficiency curve, open-circuit voltage of the power battery, and charge / discharge internal resistance in the MPC model need to be updated based on the identification results. Furthermore, the reduction in rated power and the degradation of power battery capacity require corresponding adjustments to the constraints of the objective function to prevent over-discharge and ensure system safety and performance.

[0137] Specifically, the nonlinear MPC objective function after the full lifecycle parameter update is:

[0138] (twenty two)

[0139] The constraints are:

[0140] ;

[0141] In the formula, all variables with the superscript 'a' are variables that affect the parameter updates throughout the entire lifecycle and need to be updated accordingly based on the identification characteristics.

[0142] Based on the identification results of system decay characteristics, transfer reinforcement learning training is carried out on the evaluation network and policy network of the initial PPO policy to obtain a transfer PPO policy for the entire life cycle. The action space generated by the L-PPO policy is used as the initial value of the nonlinear MPC solver for secondary optimization to obtain the L-DRL-MPC energy management policy for the entire life cycle.

[0143] Specifically, this study focuses on updating system characteristics throughout the entire lifecycle, primarily reconstructing the environment based on the results of system characteristic identification. System characteristics change gradually over time, with very few abrupt changes; therefore, only periodic (5-10 times) transfer training is needed throughout the entire lifecycle to effectively control the total computational cost. Transferable experiential knowledge includes the evaluation network of pre-trained policies, policy networks, and practical experience; different parameters need to be transferred for different reconstruction scenarios.

[0144] Specifically, in the policy update process, only the network parameters of the previously updated policy are initialized in the new network, and transfer training is carried out in combination with the reconstructed environment to obtain an adaptively updated energy management policy in fewer rounds.

[0145] Specifically, the evaluation network and policy network of the PPO algorithm consist of fully connected layers, including three hidden layers. The first two hidden layers contain 128 nodes, and the third hidden layer contains 64 nodes.

[0146] Specifically, the fuel cell system selected in this embodiment has a rated power of 150kW and a rated capacity of 120.53kWh for the power battery;

[0147] Specifically, the hyperparameters of the PPO algorithm are set as follows: the policy network learning rate is 1×10⁻ 4 The network learning rate was evaluated as 1×10⁻ 4 The discount factor is 0.99, and the experience pool capacity is 10. 6 The minimum batch size is 64, and the cutting factor is 0.2.

[0148] Specifically, Figure 3The degradation characteristics identification results of the fuel cell system in this embodiment of the invention are shown, illustrating the polarization curves of the fuel cell at different operating times, i.e., the characteristics of single-cell voltage changing with current density. The black scatter dots in the figure represent experimental measurement data, and the solid line represents the fitting results of the identification model. The two are in high agreement, indicating that the model can accurately describe the electrochemical behavior of the battery. As the operating time increases from 50 hours to 200 hours, the overall polarization curve shows a slight downward trend, especially in the higher current density region, reflecting a slight decline in battery performance with aging; however, in the low current density region, the voltage remains relatively stable, indicating that the effect of activation polarization is small. Overall, this figure accurately describes the performance evolution of the fuel cell during long-term operation and verifies the effectiveness and reliability of the established identification model.

[0149] Specifically, Figure 4 The figure illustrates the degradation characteristics identification results of the power battery at different aging stages in this embodiment of the invention. The figure depicts the characteristic curve of the battery open-circuit voltage (OCV) as a function of state of charge (SOC), where the black dots represent experimentally measured data, and the solid line represents the output of the identification model proposed in this invention. It can be seen that as the number of cycles increases, the battery gradually ages and degrades, and its OCV-SOC curve shape (especially in the low SOC range) exhibits corresponding evolutionary characteristics. The identification model curve and the experimental data maintain a high degree of consistency across the entire SOC range, accurately tracking the changes in the electrochemical characteristics of the battery at different degrees of degradation.

[0150] Specifically, Figure 5 The performance changes of the policy network under three aging states—factory delivery, 50% remaining lifespan, and 20% remaining lifespan—are shown in different training rounds. The network loss value converges rapidly to zero as training progresses, indicating model stability. The average reward and single-cycle operating cost gradually increase, while the final SOC value of the power battery gradually decreases, reflecting that the policy network can learn effectively at different lifespan stages. However, there is a trade-off between the cost and the final SOC state in the optimization objective as the training process progresses.

[0151] Specifically, Figure 6 The presentation compares the power output curves of four energy management strategies (PPO, L-PPO, RL-MPC, and L-RL-MPC) under typical driving conditions (CHTC) as a function of time. Overall, the power output of all four strategies gradually increases with the demand of the driving conditions. However, RL-MPC and L-RL-MPC actively reduce the power output in the later stages of operation (after about 1400 seconds) (from 100kW to about 50kW), while PPO and L-PPO maintain a high output (100kW). This indicates that the rolling optimization correction module can dynamically adjust the power allocation according to the objective function and constraints in the prediction time domain, avoiding long-term high-load operation of the fuel cell.

[0152] like Figure 7 As shown, this embodiment also discloses a self-regulating energy management system for fuel cell heavy-duty trucks throughout their entire lifecycle, including:

[0153] The data acquisition module 71 is used to acquire historical vehicle operation data; the historical vehicle operation data includes current and voltage data of fuel cell stack under different aging conditions, state of charge and open circuit voltage data of power battery under different aging conditions, and WTVC operating condition data.

[0154] The model training module 72 is used to construct and train a comprehensive energy management model; the comprehensive energy management model includes a system characteristic identification module, a reinforcement learning decision-making module, and a rolling optimization and correction module.

[0155] The system characteristic identification module uses a genetic algorithm to perform degradation parameter fitting on the current and voltage data of the fuel cell stack under different aging states, outputting the full life cycle fuel cell system efficiency curve. It also uses the least squares method to perform polynomial fitting on the state of charge and open-circuit voltage data of the power battery under different aging states, outputting the full life cycle power battery open-circuit voltage characteristic curve. The reinforcement learning decision module constructs a reinforcement learning interactive environment based on the full life cycle fuel cell system efficiency curve and the full life cycle power battery open-circuit voltage characteristic curve. Using WTVC operating data as the state space input, it uses a proximal policy optimization algorithm to train the policy network offline within the reinforcement learning interactive environment, outputting the trained policy network. The trained policy network is then transferred to obtain a full life cycle-oriented transfer proximal policy network. This transfer proximal policy network is used to calculate the fuel cell power change rate based on the WTVC operating data. The rolling optimization correction module uses the full life cycle-oriented transfer proximal policy network to constrain the output action space as the initial value of the nonlinear model predictive control solver, generating energy control commands based on the constrained fuel cell power change rate.

[0156] The power distribution module 73 is used to input the real-time collected vehicle operation data into the trained comprehensive energy management model to obtain energy control commands, and adjust the power distribution between the fuel cell and the power battery in real time based on the energy control commands.

[0157] The specific implementation of a self-regulating energy management system for fuel cell heavy-duty trucks throughout their entire life cycle is the same as the self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire life cycle, and will not be described again in this embodiment.

[0158] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art will understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle, characterized in that: Includes the following steps: S1, acquire historical vehicle operation data; the historical vehicle operation data includes current and voltage data of fuel cell stack under different aging conditions, state of charge and open circuit voltage data of power battery under different aging conditions, and WTVC operating condition data. S2, Construct and train a comprehensive energy management model; the comprehensive energy management model includes a system characteristic identification module, a reinforcement learning decision-making module, and a rolling optimization and correction module; The system characteristic identification module uses a genetic algorithm to perform degradation parameter fitting on the current and voltage data of the fuel cell stack under different aging states, and outputs the full life cycle fuel cell system efficiency curve. It also uses the least squares method to perform polynomial fitting on the state of charge and open circuit voltage data of the power battery under different aging states, and outputs the full life cycle power battery open circuit voltage characteristic curve. The reinforcement learning decision module constructs a reinforcement learning interactive environment based on the efficiency curve of the fuel cell system and the open-circuit voltage characteristic curve of the power battery throughout the entire life cycle. The WTVC operating data is used as the state space input. The policy network is trained offline in the reinforcement learning interactive environment using the proximal policy optimization algorithm. The trained policy network is output and the trained policy network is transferred to obtain a transfer proximal policy network for the entire life cycle. The migration near-end strategy network is used to calculate the fuel cell power change rate based on WTVC operating data. The rolling optimization correction module uses the fuel cell power change rate as the initial value of the nonlinear model predictive control solver through a migration near-end policy network oriented towards the entire life cycle, and generates energy control commands based on the constrained fuel cell power change rate. S3 inputs real-time collected vehicle operation data into a trained comprehensive energy management model to obtain energy control commands, and adjusts the power distribution between the fuel cell and the power battery in real time based on the energy control commands.

2. The self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle, as described in claim 1, is characterized in that... In S2, the state space includes vehicle speed, acceleration, fuel cell power, and power battery state of charge; wherein, vehicle speed and acceleration are used to reflect the current vehicle operating conditions, and fuel cell power and power battery state of charge are used to reflect the current power distribution state of the vehicle.

3. The self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle, as described in claim 2, is characterized in that... The fuel cell power is obtained by subtracting the power battery power from the total vehicle power requirement. The formula for calculating the required power of the whole vehicle is as follows: ； in, Power required for the entire vehicle; For the mechanical efficiency of the transmission system; For motor efficiency; Let g be the mass of the vehicle; g is the gravitational constant. This is the rolling resistance coefficient; For vehicle speed; This is the rotational mass conversion factor; To accelerate the vehicle; This refers to the air drag coefficient; The frontal area of ​​the vehicle; This refers to the road slope.

4. The self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle, as described in claim 2, is characterized in that... The formula for calculating the state of charge of the power battery is as follows: ； ； in, This refers to the open-circuit voltage of the power battery. The internal resistance of the power battery during discharge / charge; Power of the battery; The initial state of charge; For power battery capacity; This refers to the current of the power battery. This indicates the current state of charge of the power battery.

5. The self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle, as described in claim 1, is characterized in that... In S2, a genetic algorithm is used to fit the degradation parameters of the current and voltage data of the fuel cell stack under different aging conditions. The calculation formula is as follows: ； in, For ohmic resistance; i st Current density; For current; This represents the maximum current of the fuel cell stack. and The parameters to be identified; This represents the polarization characteristic curve of a fuel cell; The ratio of net output power to hydrogen energy at each moment is calculated based on the polarization curve of the fuel cell, and then connected in series according to the life process to obtain the efficiency curve of the fuel cell system throughout its entire life cycle.

6. The self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle, as described in claim 1, is characterized in that... In S2, the policy network is trained offline using the proximal policy optimization algorithm in a reinforcement learning interactive environment. The objective function of the policy network is... The calculation formula is as follows: ； in, This is the cutting factor; Indicates network parameters; The dominant function; This represents the state value at time t; This represents the action value at time t; The parameter is The old strategy network; The parameter is New strategy network; Indicates the pruning mechanism; Represents the mathematical expectation; This represents the boundary of the trajectory distribution state under the guidance of the strategy; This indicates taking the minimum value.

7. The self-regulating energy management method for fuel cell heavy-duty trucks throughout their entire lifecycle, as described in claim 1, is characterized in that... In S2, the objective function calculation formula for the nonlinear model predictive control solver is as follows: ； in, The cost of hydrogen consumption for fuel cells; Cost of fuel cell degradation; The equivalent hydrogen consumption cost for power batteries; Costs associated with battery degradation; This is the SOC adjustment factor; Indicates the current state of charge of the power battery; Indicates the target SOC.

8. A self-regulating energy management system for fuel cell heavy-duty trucks throughout their entire lifecycle, characterized in that: include: The data acquisition module is used to acquire historical vehicle operation data, which includes current and voltage data of fuel cell stack under different aging conditions, state of charge and open circuit voltage data of power battery under different aging conditions, and WTVC operating condition data. The model training module is used to construct and train a comprehensive energy management model. The comprehensive energy management model includes a system characteristic identification module, a reinforcement learning decision module, and a rolling optimization correction module. The system characteristic identification module uses a genetic algorithm to perform degradation parameter fitting on the current and voltage data of the fuel cell stack under different aging states, and outputs the full life cycle fuel cell system efficiency curve. It also uses the least squares method to perform polynomial fitting on the state of charge and open circuit voltage data of the power battery under different aging states, and outputs the full life cycle power battery open circuit voltage characteristic curve. The reinforcement learning decision module constructs a reinforcement learning interactive environment based on the efficiency curve of the fuel cell system and the open-circuit voltage characteristic curve of the power battery throughout the entire life cycle. The WTVC operating data is used as the state space input. The policy network is trained offline in the reinforcement learning interactive environment using the proximal policy optimization algorithm. The trained policy network is output and the trained policy network is transferred to obtain a transfer proximal policy network for the entire life cycle. The migration near-end strategy network is used to calculate the fuel cell power change rate based on WTVC operating data. The rolling optimization correction module uses the fuel cell power change rate action as the initial value of the nonlinear model predictive control solver through a migration near-end policy network oriented towards the entire life cycle, and generates energy control commands based on the constrained fuel cell power change rate. The power distribution module is used to input real-time collected vehicle operation data into a trained comprehensive energy management model to obtain energy control commands, and adjust the power distribution between the fuel cell and the power battery in real time based on the energy control commands.

Images