Real-time SLMP fuel cell hybrid electric vehicle energy management method

Abstract

The invention discloses a real-time SLMP fuel cell hybrid electric vehicle energy management method. The method comprises the following steps: firstly, collecting vehicle operation data in real time, constructing an operation state vector containing kinetic parameters, and initializing a state transition matrix as agent action input; on the basis, the speed and the acceleration are predicted through a self-learning Markov predictor, and dynamic self-adaptive power distribution of the fuel cell and the power cell is achieved in combination with a multi-objective optimization function and power distribution constraints. The method passes hardware-in-the-loop (HIL) verification, aims at reducing hydrogen consumption and improving efficiency, and exits if two times of circulation are not passed. Compared with a traditional method based on rules or static optimization, the method innovatively embeds a self-learning mechanism, has the advantages of being high in adaptability, high in prediction precision and good in generalization performance, and is suitable for energy management optimization under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of fuel cell hybrid electric vehicle technology, specifically relating to a real-time SLMP energy management method for fuel cell hybrid electric vehicles, aiming to improve the economy and system lifespan of fuel cell hybrid electric vehicles. Background Technology

[0002] With increasing environmental awareness and the depletion of fossil fuel resources, fuel cell hybrid electric vehicles (FCHEVs), as a clean and efficient mode of transportation, have become an important direction for the development of new energy vehicles. Fuel cells use hydrogen as fuel to generate electricity through electrochemical reactions, offering advantages such as zero emissions, high efficiency, and long driving range. The power battery, as an auxiliary energy source, provides peak power and recovers braking energy, achieving energy complementarity. However, in actual operation, fuel cell hybrid vehicles face many challenges in energy management: fuel cells have slow response times and their lifespan is easily affected by frequent power fluctuations; the state of charge (SOC) of the power battery needs to be maintained within a reasonable range to avoid overcharging and over-discharging; and the complex and variable operating conditions of the vehicle lead to high hydrogen consumption and low system efficiency.

[0003] In existing technologies, common energy management strategies include rule-based methods (such as power following strategies or isothermal control strategies) and optimization-based methods (such as dynamic programming or equivalent consumption minimization strategies). Rule-based methods are simple to implement, but they ignore future operating condition information, leading to suboptimal energy allocation, increased hydrogen consumption, and accelerated fuel cell degradation. Optimization-based methods, while capable of globally solving for the optimal solution, are computationally burdensome and unsuitable for real-time applications, especially when vehicle speed prediction is inaccurate, limiting their optimization effectiveness. Furthermore, traditional prediction models often rely on fixed parameters and cannot adapt to the self-learning requirements of driver behavior and road condition changes; simultaneously, existing multi-objective optimization frameworks often fail to adequately consider the degradation costs of fuel cells, resulting in shortened system lifespan.

[0004] To address the aforementioned issues, this invention proposes a real-time self-learning Markov prediction (SLMP) energy management method for fuel cell hybrid vehicles. Through adaptive speed prediction, multi-objective optimization functions, and rolling optimization mechanisms, it achieves efficient energy allocation, reduces hydrogen consumption, and extends fuel cell lifespan. Summary of the Invention

[0005] This invention provides a real-time SLMP-based energy management method for fuel cell hybrid vehicles. This method addresses the energy allocation challenges of fuel cell hybrid vehicles by proposing a forward-looking energy management scheme that combines self-learning prediction and multi-objective optimization.

[0006] The concept originated from the limitations of traditional energy management strategies, such as rule-based methods ignoring future operating conditions, leading to high hydrogen consumption, and global optimization methods being computationally burdensome and unsuitable for real-time applications.

[0007] To this end, this invention introduces a self-learning Markov predictor to capture vehicle speed dynamics, constructs a fuel cell-power battery power flow model, and takes minimizing hydrogen consumption as the core objective, incorporating battery SOC maintenance and fuel cell degradation costs to form a multi-objective optimization framework.

[0008] Rolling optimization is achieved through a sequential quadratic programming algorithm, ensuring both real-time performance and economic efficiency. This scheme uses historical data to adaptively update the predictor, dynamically adjusting the power allocation sequence to achieve closed-loop control, thereby improving vehicle economy and fuel cell lifespan.

[0009] The implementation steps of this method are as follows:

[0010] Step 1: Developers pre-input the vehicle's dynamic parameters (such as the vehicle's mass m). v Air drag coefficient C d The parameters, such as frontal area A, rolling resistance coefficient (f_r), and power system parameters (such as fuel cell rated power (P_{stack}) and power battery capacity), are used to build the basic model of the energy management system, ensuring that the model can accurately reflect the physical and dynamic characteristics of the vehicle.

[0011] Step 2: Based on the above input parameters, this method uses the longitudinal dynamics equation to calculate the vehicle's power demand P at the current moment in real time. req (t), this power serves as the target load for the energy management strategy, guiding subsequent power allocation. The specific form of the longitudinal dynamic equation is:

[0012]

[0013] Where v (t) This indicates the vehicle's current speed (unit: m / s), m v ρ represents the vehicle's mass (in kg), and ρ represents the air density (in kg / m³). 3 ), C d The drag coefficient (A) represents the vehicle's frontal area (unit: m²). 2 ), f r The rolling resistance coefficient is represented by g, and g represents the acceleration due to gravity (unit: m / s²). 2 θ(t) represents the current slope angle (in radians). This equation converts mechanical factors such as vehicle acceleration, air resistance, rolling resistance, and slope resistance into power requirements, ensuring that the calculation results respond in real time to changes in vehicle dynamics.

[0014] Step 3: To provide predictive information about future vehicle operating conditions, this method discretizes vehicle acceleration into n states and constructs and continuously updates a self-learning Markov velocity predictor based on real-time collected historical data. The specific construction and updating process is as follows:

[0015] First, the acceleration a(t) is discretized into n states at equal intervals, forming the state space X. a =a1,a2,…,a n , where each a i This represents the median value of a discrete acceleration interval (unit: m / s²). 2 Secondly, set the prediction window H. p (representing the predicted future time step) and learning rate φ (within the range of (0,1), used to control the update speed); then, initialize the set of state transition matrices. Each T l It is an n×n matrix with initial elements [T_l]. ij = 1 / n, representing the initial transition probability of a uniform distribution; then, at time L, based on the observed historical transition markers... (A flag of 1 indicates a transition from state i to j has occurred; otherwise, it is 0.) Update the state transition probability using the following formula:

[0016]

[0017] in

[0018]

[0019] This represents the cumulative transition frequency from state i to j. This represents the cumulative transition frequency from any state to j; finally, using the updated T... l Calculate the expected acceleration

[0020]

[0021] And the predicted speed is obtained recursively.

[0022]

[0023] Where Δt represents the sampling time interval (unit: s). This predictor continuously adapts to actual operating conditions through a self-learning mechanism, providing a reliable future velocity sequence for forward-looking energy management.

[0024] Step 4: Based on the predicted vehicle speed information above, this method establishes a fuel cell-power battery power flow model to describe the energy flow of the hybrid power system. The specific model is as follows:

[0025] P req (t)=[P b (t)+P dc (t)]η m (t) (7)

[0026] in

[0027] P dc (t)=P fcreq (t)η dc (t)

[0028] P b (t) represents the output power of the power battery (unit: kW, positive value for discharging, negative value for charging), P dc (t) represents the output power of the DC / DC converter (unit: kW), P fcreq (t) represents the requested power of the fuel cell (unit: kW), η m (t) represents the motor efficiency (range 0-1), η dc (t) represents the efficiency of the DC / DC converter (range 0-1). This model ensures that the required power is met by both the fuel cell and the power battery, while also considering efficiency losses to optimize power distribution.

[0029] Step 5: To achieve optimal economic efficiency, this method constructs a multi-objective optimization function with minimizing hydrogen consumption as the core, while also introducing a fuel cell degradation cost model. The specific optimization function is as follows:

[0030]

[0031] in Indicates instantaneous hydrogen consumption rate (unit: g / s), m equ (t) represents the equivalent hydrogen consumption rate (unit: g / s), k p SOC represents the SOC maintenance penalty factor (a coefficient used to penalize SOC deviation), SOC(t) represents the current battery state of charge (range 0-1), and SOC(0) represents the initial state of charge; equivalent hydrogen consumption rate.

[0032]

[0033] in This indicates the lower calorific value of hydrogen (unit: MJ / kg), kJ / kg. EF This represents the battery energy-hydrogen equivalence factor. This function minimizes total hydrogen consumption and equivalent cost through integration, while maintaining stable battery state of charge (SOC).

[0034] Step 6: The above equivalent factor k EF The value is dynamically determined based on the power direction of the power battery:

[0035]

[0036] Where η ebg This indicates the average charging efficiency of the battery charging system. Indicates the average operating efficiency of the DC / DC converter. This indicates the average charging efficiency of the power battery. This represents the average discharge efficiency of the power battery. This value selection rule ensures that the equivalence factor accurately reflects the difference in energy conversion efficiency during charging and discharging.

[0037] Step 7: To further optimize the method, a fuel cell degradation cost model is introduced.

[0038] m deg =m shift +m high +m low (11)

[0039] Where m shift The decay cost caused by power changes (unit: g / s), m high The m represents the degradation cost caused by high power output (unit: g / s). low This represents the degradation cost caused by low power output (unit: g / s); specifically,

[0040]

[0041] Where ΔP fc N represents the change in fuel cell output power between adjacent time steps (unit: kW), Δt represents the time step size (unit: s), and N represents the change in fuel cell output power between adjacent time steps. fc α represents the number of fuel cell stacks. shift U represents the transient loss coefficient (unit: V / kW·s). EoL,fc P represents the voltage drop at the end of life (unit: V). stack This indicates the rated power of the fuel cell stack (unit: kW);

[0042]

[0043] Where T high α represents the duration of high-power output (in seconds). high Indicates the high power loss factor (unit: V / h);

[0044]

[0045] Where T low α represents the duration of low-power output (in seconds). lowThis represents the low power loss coefficient (unit: V / h). The model quantifies the impact of power variations and the duration of high and low power on fuel cell lifespan, converting degradation costs into equivalent hydrogen consumption and incorporating them into the optimization function.

[0046] Step 8: The constraints of the above optimization proposition include:

[0047]

[0048] in and These represent the minimum and maximum output power of the fuel cell (unit: kW), respectively. and These represent the minimum and maximum output power of the power battery (unit: kW), and SOC, respectively. min and SOC max These represent the lower and upper limits of the battery's state of charge, respectively. These constraints ensure that power distribution is within physical and safety limits.

[0049] Step 9: To solve this optimization problem, this method employs the Sequential Quadratic Programming (SQP) algorithm, iteratively obtaining the optimal power allocation sequence that achieves the best economy within the prediction time domain. The specific iterative steps are as follows: First, construct the Lagrange function.

[0050]

[0051] Where (x) represents the optimization variable (e.g., and P b (t) sequence), J(x) represents the objective function, g i (x) and h i (x) represent inequality constraints and equality constraints, respectively, and λ i and μ i Represent the Lagrange multiplier; calculate the gradient of the objective function and constraint functions at the current optimization variable (x). Secondly, at the current iteration point x k Solving the quadratic subproblem:

[0052]

[0053] Limited by:

[0054]

[0055] Where H k Let H represent the approximation of the Hessian matrix, and d represent the search direction; then, update H using the BFGS formula. k ,Right now

[0056]

[0057] Where s k =x k+1 -x k , The subproblem is solved repeatedly until convergence, yielding the optimal control sequence u. * (t)(including) and P b (t)); Finally, the first control command in the sequence is sent to the vehicle power system in real time, and the speed predictor is updated again in the next sampling period to form a rolling optimization closed-loop control process. Attached Figure Description

[0058] 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 of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart of the operator deployment optimization method provided by the present invention; Detailed Implementation

[0060] This invention provides a real-time self-learning Markov prediction (SLMP) energy management method for fuel cell hybrid electric vehicles. It dynamically updates the state transition matrix by collecting vehicle operating data in real time and optimizes power allocation using a sequential quadratic programming (SQP) algorithm. The following describes the implementation steps and key parameters.

[0061] Step A: Data Collection and Initialization

[0062] Developers pre-input real-time collected vehicle dynamics and power system parameters, including vehicle speed v(t) ranging from 0 to 30 m / s and acceleration a(t) ranging from [-3, 3] m / s. 2 The dynamic parameters, including discretization into 61 states and slope angle θ(t) (range -5° to 5°), were initialized, and the basic model of the energy management system was constructed. Key parameters included a vehicle mass m of 1800 kg and an air density ρ of 1.225 kg / m³. 3 Air drag coefficient C d The value is 0.3, and the windward area A is 2.2m². 2 Rolling resistance coefficient f r Initialize the state transition matrix set to 0.01, with a fuel cell rated power of 50kW, etc. Where [T] l ] ij = 1 / 61 (61 states), prediction window H pThe learning time is 5 seconds, and the learning rate φ is 0.01. Historical data and the current state are used to construct the transition frequency. Simultaneously, data validity is assessed, with the standard being whether the acceleration a(t) is within the range of [-3, 3] m / s². 2 Within the specified range, is the data missing rate less than 5%? If the data is valid, continue with subsequent steps; if invalid, the system exits and logs an error.

[0063] Step B: Self-learning Markov predictor update

[0064] The state transition matrix is ​​dynamically updated based on historical observation time-domain data. The transition frequency at time L is determined by the transition flag. The value is used to determine whether a transition has occurred; 1 indicates a transition has taken place, and 0 indicates it has not. Update:

[0065]

[0066] The transition probability is calculated using formulas 3 and 4 in step 3 of the invention description as follows:

[0067]

[0068] Simultaneously, a success check is performed: the calculation time is less than 0.1 seconds and the sum of probabilities for each row equals 1. If the update is successful, continue to step C; if unsuccessful, adjust φ in the range [0.001-0.1] or adjust the prediction window H in the range [3-10]. p Return to step A and reinitialize.

[0069] Step C: Velocity and Acceleration Prediction

[0070] Based on the updated state transition matrix T l Calculate the expected future acceleration and the predicted velocity. The expected acceleration is calculated using formula 5 in step 3 of the invention:

[0071]

[0072] Speed ​​recursion is derived from formula 6 in step 3 of the invention:

[0073]

[0074] The sampling interval Δt is 1 second. Simultaneously, the prediction accuracy is judged, with the standard being a deviation from the actual value of less than 10% or a confidence interval of 95%. If accurate, proceed to step D; if inaccurate, resample the observation time-domain data and return to step B to update the matrix.

[0075] Step D: Optimization of Energy Management Strategies

[0076] The predicted velocity is input into the multi-objective optimization framework, and the optimization function is constructed using formula 8 in step 5 of the invention:

[0077]

[0078] in

[0079]

[0080] in The hydrogen consumption rate (based on a baseline value of 0.1 g / s) is... Hydrogen has a low calorific value. The SOC penalty factor is 120 MJ / kg. p The value is 1000, and the SOC(t) ranges from 0.3 to 0.9. The power flow model is derived from formula 7 in step 4 of the invention:

[0081] P req (t)=[P b (t)+P dc (t)]η m (t)

[0082]

[0083] The constraint is achieved through formula 13 in step 8 of the invention content:

[0084]

[0085] The SQP algorithm is used to solve the problem. Simultaneously, convergence is assessed based on the objective function changing by less than 0.01% or the number of iterations being less than 100. If convergence is achieved, proceed to step E; otherwise, adjust the constraints (e.g., ...). or weight k p The range is [500-2000]. Return to this step and optimize again.

[0086] Step E: Closed-loop verification and adjustment

[0087] The optimization results are verified through hardware-in-the-loop experiments, including UDDS and CATC operating conditions, with the goal of reducing hydrogen consumption by 5-10% and improving efficiency by 3-5%. If the verification is successful, the optimization results are output and the loop continues to run; if minor issues are encountered (such as performance deviation <5%), return to step D for re-optimization. If the loop fails to verify after two runs, exit the system and log the error.

Claims

1. A real-time SLMP (Single-Level Multiprocessing) method for energy management of fuel cell hybrid vehicles, characterized in that, Includes the following steps: Developers pre-input the vehicle's dynamic parameters and powertrain parameters to establish a basic model for the energy management system; Based on the input parameters and longitudinal dynamic equations, the vehicle power demand Preq(t) at the current moment is calculated in real time and used as the target load for the energy management strategy. The vehicle acceleration is discretized into n states. Based on real-time collected historical data, a self-learning Markov velocity predictor is constructed and continuously updated to provide predictive information on vehicle operating conditions in the future, providing a basis for forward-looking energy management. Based on the predicted vehicle speed information, a fuel cell-power battery power flow model is established, and a multi-objective optimization function is constructed with the minimum hydrogen consumption as the core. At the same time, a fuel cell degradation cost model is introduced to form a mathematical proposition aimed at optimizing energy distribution. The Sequential Quadratic Programming (SQP) algorithm is used to solve this optimization problem and obtain the optimal power allocation sequence that achieves the best economy in the prediction time domain. The first control command in the optimal power allocation sequence is sent to the vehicle powertrain system in real time to execute the current energy management decision; In the next sampling cycle, the process returns to building and updating the self-learning Markov velocity predictor, forming a rolling optimization closed-loop control process that can continuously adjust the energy allocation strategy according to new operating conditions.

2. The method according to claim 1, wherein the longitudinal dynamic equation is: In the formula: v(t) is the vehicle speed; mv is the total vehicle mass; ρ is the air density; C d Where A is the air drag coefficient; A is the frontal area; f r θ(t) is the rolling resistance coefficient; θ(t) is the slope angle.

3. The method according to claim 1 or 2, wherein step c) of constructing and updating the self-learning Markov velocity predictor includes: S1 divides the acceleration a(t) into n states at equal intervals, forming the state space X. a =(a1,a2,…,a n ); S2 sets the prediction window H p With the learning rate φ∈(0,1); S3 Initialize the set of state transition matrices Where [Tl]ij = 1 / n; S4 at time L is based on the observed historical transition markers. Update state transition probability: S5 uses the updated Tl to calculate the expected acceleration. The predicted speed is then obtained recursively.

4. The method according to claim 1, wherein the fuel cell-power battery power flow model is: P req (t)=[P b (t)+P dc (t)]η m (t) P dc (t)=P fc_req (t)η dc (t) Where P b (t) represents the power of the battery, P dc (t) represents the DC / DC output, P fc_req (t is the requested power of the fuel cell, η) m (t), η dc (t) represents the motor efficiency and DC / DC efficiency, respectively.

5. The method according to claim 1, wherein the multi-objective optimization function in step 1 is: in The instantaneous hydrogen consumption rate is given by kp; the SOC maintenance penalty factor is given by SOC(0); and the initial state of charge is given by SOC(0). LHV H2 The lower calorific value of hydrogen; k EF The equivalent factor for battery electrical energy to hydrogen is 1.

6. The method of claim 5, wherein k EF Values ​​are determined according to the following rules: η ebg This indicates the charging power of the battery charging system. This refers to the average operating efficiency of the DC-DC power converter. With η chg These represent the average discharge power and charging power of the power battery, respectively.

7. The method according to claim 1, wherein the fuel cell degradation cost model is: m deg =m shift +m high +m low Where ΔP fc N represents the change in fuel cell output power between adjacent step sizes. fc T represents the number of fuel cell stacks. high and T low These represent the high-power output time limit and low-power output time limit of the fuel cell, respectively; high power loss α is used. high and low power loss α low To characterize the power regulation rate, and through the transient loss α of the output power. shift To reflect the number of starts and stops; P stack Rated power of the fuel cell stack; U EoL,fc The voltage drop at the end of life; T high T high For high and low power durations.

8. The method according to claim 1, wherein the constraints of the optimization proposition include: P fc_min ≤P fc_req (t)≤P fc_max P b_min ≤P b (t)≤P b_max SOC min ≤SOC(t)≤SOC max Where P fc_min P fc_max These represent the minimum and maximum output power of the fuel cell, respectively; P b_min P b_max These represent the minimum and maximum output power of the power battery, respectively; SOC min &、SOC max These are the lower and upper limits of the battery's state of charge, respectively.

9. The method according to claim 1, wherein the sequential quadratic programming algorithm solves for the optimal control sequence u*(t) through the following iterative steps: S1 constructs the Lagrange function Where g i and h i These are equality constraints and inequality constraints, respectively; calculate the gradients of the objective function and constraint functions at the current optimization variable x. S2 solves the quadratic subproblem at the current iteration point xk: S3 updates Hk using the BFGS formula and repeats (h-2) until convergence, obtaining the optimal control sequence u(t); S4 executes only the first control variable in the sequence and returns to form a rolling optimization closed loop in the next sampling period.

Images