A method for calculating dynamic energy demand boundary of fuel cell electric vehicle

Abstract

A kind of dynamic energy demand boundary calculation method of fuel cell electric vehicle.The present application relates to the control technical field of fuel cell vehicle, and specifically relates to a kind of dynamic energy demand boundary calculation method of fuel cell electric vehicle.By establishing vehicle traction power model, off-line generation demand boundary library and fluctuation index library, and in online stage, combining residual compensation, confidence tightening and hardware power feasible region intersection, the upper and lower bounds of dynamic power demand in current control period are quickly output, to provide executable, updateable, robust front feed input for subsequent fuel cell vehicle energy management.The method comprises the following steps: establishing vehicle traction power model, off-line generation demand boundary library and fluctuation index library, and in online stage, combining residual compensation, confidence tightening and hardware power feasible region intersection, the upper and lower bounds of dynamic power demand in current control period are quickly output.

Description

Technical Field

[0001] This invention relates to the field of fuel cell vehicle control technology, and specifically to a method for calculating the dynamic energy demand boundary of a fuel cell electric vehicle. Background Technology

[0002] With increasingly stringent requirements for energy conservation and emission reduction in the automotive industry, fuel cell vehicles have attracted widespread attention due to their high energy conversion efficiency, low emissions, and promising driving range. Most existing fuel cell vehicle energy management technologies directly use the current demand power or single-point predicted power as the control input. However, in engineering implementation, this approach has several shortcomings: First, vehicle operating conditions exhibit significant time-varying and randomness, making it difficult for single-point power to reflect future short-term load changes, resulting in insufficient controller adaptability to sudden loads and rapid fluctuations. Second, static thresholds, operating condition identification tags, or single predicted values ​​cannot form an executable power range directly matching the controller's constraint structure. Third, both fuel cells and power batteries are constrained by upper and lower limits of output power, efficiency ranges, SOC boundaries, and dynamic response capabilities. If the upper layer cannot provide physically feasible demand boundaries, the lower layer's power allocation is prone to becoming conservative, lagging, or even infeasible. Traditional operating condition identification methods, primarily based on semantic classification, struggle to meet the real-time control requirements for continuous, quantifiable, and embeddable constraint-oriented control-oriented descriptions.

[0003] To address the aforementioned issues, the industry's initial approach was to directly embed model predictive control (MPC) into the vehicle controller, solving the constrained power allocation optimization problem in real time within each control step, aiming to simultaneously cover the dual requirements of multi-time-domain demand forecasting and boundary constraints. Theoretically, this approach can directly incorporate future power demand changes within a finite time domain into the optimization objective.

[0004] However, the purely online MPC approach faces insurmountable real-time bottlenecks in engineering practice. For fuel cell vehicle energy management, the system state dimension is high, and the prediction time domain typically needs to cover several seconds to tens of seconds, resulting in the need to solve a high-dimensional non-convex optimization problem for each control step. On a typical onboard controller, this solution time often exceeds the control cycle, failing to meet real-time requirements. Furthermore, the stochastic nature of power demand makes it difficult for online MPC to form a structured description of the future power range; its output is a single optimal solution rather than a power boundary interval, which cannot be directly used for downstream constraint setting and strategy switching.

[0005] To address the issue of excessive online computation, another widely explored approach is the combination of offline global optimization and operating condition identification labels. The basic idea of ​​this approach is to offline run a global optimization algorithm (such as dynamic programming) for several typical operating conditions (e.g., WLTC, HWFET, urban congestion), storing the optimization results in a rule table or lookup table. In the online phase, the operating condition identification module categorizes the current driving state into a specific typical operating condition category and then calls the corresponding offline rule table to output power commands.

[0006] The core flaw of this approach lies in its reliance on semantic labels for describing operating conditions, failing to provide continuous and quantifiable control-oriented power boundaries. First, operating condition classification is inherently lossy compression; when actual driving conditions are near category boundaries, classification results are unstable, leading to policy jumps and power oscillations during category switching. Second, the offline global optimal solution corresponds to the entire driving cycle under specific operating condition assumptions, while the future short-term power demand range of actual driving continuously changes with real-time states; the static rule table cannot reflect the demand boundary corresponding to the current state point in real time. Third, the operating condition identification labels cannot carry upper and lower bounds and fluctuation intensity information of power demand; downstream controllers still need to handle constraint boundaries themselves, essentially shifting the difficulty of boundary calculation downstream rather than truly solving it. In summary, this approach is deemed incapable of providing power boundary information that matches the controller's constraint structure.

[0007] With the development of machine learning technology, deep learning-based driving behavior prediction methods have also been introduced into demand power prediction. This approach uses sequence models such as LSTM and Transformer to model historical driving data, directly predicting the power point value sequence at several future moments in an end-to-end manner. It attempts to bypass the complexity of vehicle dynamics modeling in a data-driven way and capture the nonlinear patterns of driver behavior.

[0008] However, this approach suffers from three fundamental drawbacks in control-oriented applications: First, the deep learning point-value prediction model outputs an expected estimate of future power, rather than the potential range of future power, making it impossible for the controller to set safety constraints based on this output. Second, the prediction model has limited generalization ability in operating conditions outside the training set (such as special slopes, extreme weather, and unfamiliar road sections), and the prediction error may be drastically amplified under extreme conditions. Furthermore, the model itself cannot perceive its own prediction reliability and cannot proactively trigger safety protection when predictions fail. Third, large sequence models have high inference computation requirements, making them difficult to deploy on resource-constrained vehicle controllers. Moreover, the model output lacks physical interpretability, making it difficult to integrate with the hardware constraints of fuel cells and power batteries. Therefore, this approach also cannot meet the power boundary calculation requirements of control constraint-oriented applications.

[0009] In addition, a simpler and more direct approach was also explored in the project: based on the single-point predicted power, an empirical fixed safety margin was superimposed to construct the power boundary. This approach is easy to implement, requires no additional optimization calculations, and can provide the controller with a certain degree of power range reference.

[0010] However, the inherent flaw of the fixed safety margin scheme lies in its decoupling from actual operating conditions. Under stable cruise conditions, the fixed margin is often too large, widening the power boundary beyond actual demand, thus compressing the energy management strategy's optimization space and leading to a conservative power allocation between the fuel cell and the power battery. Under highly dynamic conditions such as rapid acceleration and steep inclines, the fixed margin is often insufficient, potentially exceeding the preset boundary and causing boundary penetration, rendering the controller constraints ineffective. The fixed margin value is essentially a static guess of the operating condition intensity, unable to adaptively adjust to the current state, the degree of operating condition fluctuation, and the level of prediction error. Therefore, it is difficult to simultaneously meet the dual objectives of safety and economy under diverse actual operating conditions. Consequently, this scheme also fails to meet engineering requirements. Summary of the Invention

[0011] To address the problem that existing demand forecasting methods struggle to provide structured demand boundaries that match controller constraints, this invention proposes a dynamic energy demand boundary calculation method for fuel cell electric vehicles. By establishing a vehicle traction power model, generating an offline demand boundary library and a fluctuation index library, and combining residual compensation, confidence tightening, and hardware power feasible region intersection in the online phase, the method quickly outputs the upper and lower bounds of the dynamic power demand within the current control cycle. This provides an executable, updatable, and robust feedforward input for subsequent fuel cell vehicle energy management.

[0012] The method includes the following steps: S1. Obtain the real-time status of the vehicle during operation and calculate the vehicle's traction power requirement. ; S2. Combining the fuel cell output capacity, the power battery charging and discharging capacity, and the electric drive system losses, the feasible power domain of the whole vehicle hardware is constructed. S3, Definition of the The state vector at each control moment , used to calculate dynamic energy demand over a finite time domain in the future; S4, Based on vehicle traction power requirements Build an offline requirement boundary library; S5. Constructing offline short-time domain fluctuation indicators ; S6, according to The corresponding offline boundary value is retrieved from the offline demand boundary library, and multidimensional interpolation and residual compensation calculations are performed on the offline boundary value to obtain the dynamic initial energy demand boundary in the future finite time domain. S7, according to The confidence level of the state estimation is tightened to improve the confidence level of the dynamic initial energy demand boundary within the future finite time domain. Then, the intersection of the tightened demand boundary and the feasible power domain of the vehicle hardware is obtained to obtain the dynamic energy demand boundary in the future finite time domain. and ; S8, according to and renew Online short-time domain fluctuation index was obtained. .

[0013] Furthermore, the real-time status of the vehicle during operation includes: vehicle mass. longitudinal acceleration of vehicles Vehicle speed and vehicle frontal area ; Vehicle traction power requirements The formula for calculation is: ,in, The equivalent efficiency of the transmission system. This refers to the traction force on the wheel side.

[0014] Furthermore, the feasible domain of vehicle hardware power includes: the maximum power of vehicle hardware. and the minimum power of the whole vehicle hardware ; , ,in, and These represent the maximum and minimum stable power output of the fuel cell, respectively. and These refer to the maximum allowable discharge power and maximum charging power of the power battery, respectively. This represents the equivalent losses of the electric drive system.

[0015] Furthermore, the first The state vector at each control moment ,in, For the first Vehicle speed at a given moment in time. For the first Vehicle acceleration at a control moment For the first Road gradient at a given time point For the first Battery state of charge at each control moment For the first The output power of the fuel cell system at each control moment. For the first The power output of the battery at each control moment.

[0016] Furthermore, in step S4, for any state grid point... Given a set of reachable accelerations and slope set Under the conditions, based on Find the upper boundary of the vehicle dynamics model respectively and the lower boundary of the vehicle dynamics model ; , ; against and Introducing a safety buffer Obtain the robust upper bound. and robust lower boundary ; The offline demand boundary library includes: offline demand upper boundary. and the lower boundary of offline requirements ; , ,in, and These are the empirical upper and lower boundaries, obtained by performing quantile statistics on historical demand power within the state neighborhood.

[0017] Furthermore, offline short-term fluctuation indicators The relation is: in, and All are weighting coefficients. This represents the total number of time steps within the short-time domain. For arbitrary state grid points Obtaining the future from the nominal evolutionary trajectory A power sequence with a step size, for The mean, for In the future The maximum value of each step size. for In the future The minimum value of each step size.

[0018] Furthermore, in step S6, according to Retrieve the corresponding offline boundary value from the offline requirement boundary library. and ; against and Perform multidimensional interpolation to obtain the upper boundary of the initial demand in the future finite time domain. and the lower boundary of initial demand within the finite time domain in the future ; against and By performing residual compensation, the upper boundary of the initial dynamic energy demand in the future finite time domain is obtained. and the lower boundary of the dynamic initial energy requirement within the finite time domain in the future , , ,in, For the first The disturbance estimate at the first control moment is obtained through the first control moment. The actual power demand of the vehicle at the first control moment and the first control moment Calculation of the deviation of vehicle traction power demand at each control moment.

[0019] Furthermore, in step S7, the confidence tightening quantity is constructed. , ,in, For adjustment coefficients, For the first Reliability of state estimation at each control moment , For the first The statistical standard deviation of residuals at each control time point; Then according to right and By tightening the confidence level, we obtain the upper bound of the dynamic energy demand over a finite time domain in the future. and the lower boundary of dynamic energy demand in the future finite time domain ; , .

[0020] Furthermore, the dynamic energy demand boundary within the finite time domain includes: the upper boundary of the dynamic energy demand within the finite time domain. and the upper boundary of dynamic energy demand within a finite future time domain ; , ,in, and These represent the maximum and minimum power of the vehicle hardware within a finite future time domain, respectively.

[0021] Furthermore, in step S8, The formula for calculation is:

[0022] in, and All are weighting coefficients.

[0023] The beneficial effects of the method described in this invention are as follows: (1) Unlike existing methods that only predict the single-point power demand at a certain moment, this invention uses the upper and lower bounds of the power demand over a finite future time domain as the prediction object. This design transforms the energy management strategy from a passive, single-point feedback correction mode that responds to the current power shortage to a feedforward optimization under boundary constraints within the allowable power range in the future. The upper and lower bounds of the demand can be directly used as time-varying constraints for optimization algorithms such as model predictive control, enabling the controller to adjust the power sharing ratio of the fuel cell and the power battery in advance, reducing inefficient output and frequent power adjustments in high-fluctuation intervals.

[0024] (2) Based on residual compensation, this invention further introduces a boundary tightening mechanism jointly driven by the reliability of state estimation and the standard deviation of residual statistics. When the reliability decreases or the residual fluctuation increases, the demand boundary is actively contracted inward to reserve a safety margin for the controller; when the state estimation is reliable, the boundary is automatically widened to make full use of the energy-saving space, thus achieving the technical effect of pre-emptive defense against uncertainty with extremely low computational cost.

[0025] (3) This invention intersects the predicted contracted power demand boundary with the feasible power range of the vehicle hardware, so that the final output dynamic energy demand boundary always falls within the physical power range that the fuel cell and power battery can execute. This mechanism completely eliminates the risk of the upper-level optimizer outputting power commands that the lower-level hardware cannot execute from the algorithm level, and ensures the safe and reliable deployment of the control strategy in the real physical system.

[0026] (4) This invention constructs an offline mapping library between vehicle state and demand boundary. In the online stage, only table lookup, interpolation, and lightweight compensation are needed to quickly output the complete time-domain power boundary. This architecture, which combines offline global optimization with online local correction, enables this invention to complete boundary updates on the vehicle controller in milliseconds without large-scale online optimization at each step. It provides an engineering solution for real-time robust control under complex operating conditions that has both a global optimization perspective and meets edge computing constraints. Attached Figure Description

[0027] Figure 1 This is a flowchart of the method described in this invention; Figure 2 This is a schematic diagram of the target vehicle configuration of the present invention; Figure 3 This is a boundary tracking result diagram of the dynamic energy demand boundary described in this invention under WLTC conditions; Figure 4 This is a boundary tracking result diagram of the dynamic energy demand boundary described in this invention under HWFET operating conditions; Figure 5 This invention compares the effect of dynamic energy demand boundary on vehicle energy consumption optimization under 5 WLTC cycle conditions. Figure 6 This invention compares the effect of dynamic energy demand boundary on vehicle energy consumption optimization under 14 HWFET operating cycles. Detailed Implementation

[0028] The technical solution of the present invention will now be clearly and completely described 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.

[0029] Example 1 This embodiment provides a method for calculating the dynamic energy demand boundary of a fuel cell electric vehicle. It employs a combined offline-online approach to achieve real-time calculation of the dynamic energy demand boundary. In the offline phase, based on the vehicle dynamics model, road gradient variation range, driving condition samples, and power system operating constraints, a short-time domain demand boundary library and a fluctuation index library corresponding to different state points are constructed. In the online phase, based on the vehicle's current speed, acceleration, road gradient, power battery SOC, fuel cell output power, and power battery output power, the offline boundary is quickly queried, and the final dynamic energy demand boundary is obtained by combining residual compensation, boundary tightening, and hardware intersection.

[0030] The flowchart of the method is as follows: Figure 1 As shown, the method includes the following steps: S1. Obtain the real-time status of the vehicle during driving, construct a longitudinal dynamics model of the vehicle, and calculate the vehicle's traction power requirement. ; The relevant operations in step S1 will be introduced with specific examples: The target vehicle in this embodiment is a fuel cell vehicle, and the schematic diagram of the target vehicle configuration is shown below. Figure 2 As shown. The total vehicle weight of the target vehicle. The longitudinal acceleration of the vehicle is 1860 kg. Vehicle speed was obtained through actual measurement using an accelerometer. The vehicle's frontal area was obtained through actual measurements using a vehicle speed sensor. 2.2m 2The tire rolling radius is 0.35 m; the maximum speed of the front motor is 14,000 rpm and the maximum torque is 137 Nm; the maximum speed of the rear motor is 10,000 rpm and the maximum torque is 97 Nm; the nominal voltage of the power battery is 362 V and the battery capacity is 40 Ah; the minimum stable output power of the fuel cell is 2 kW, the optimal output power is 20 kW, and the maximum output power is 50 kW. In this embodiment, the road slope is... The rolling resistance coefficient is 0 degrees. The air density is 0.012. It is 1.225 kg / m 2 air drag coefficient With a value of 0.3, the above parameters can be directly used for vehicle dynamics modeling, hardware power boundary calculation, and online correction of dynamic energy demand boundaries.

[0031] The power system of a fuel cell vehicle mainly includes a hydrogen energy storage system, a fuel cell system, a power battery system, and a motor and transmission system. To calculate the traction power requirements of the vehicle under different operating conditions, a longitudinal dynamics and traction power requirement model of the vehicle is first established.

[0032] During operation, a vehicle must overcome acceleration resistance, rolling resistance, gradient resistance, and air resistance; therefore, the traction force on the wheel sides... The following vehicle longitudinal dynamics model relationship is satisfied: ; Based on wheel-side traction It can obtain the vehicle's traction power requirements. Specifically: ,in, The equivalent efficiency of the transmission system can be obtained by combining the efficiency graphs of the motor and reducer, as well as the results of bench calibration. If a partial efficiency graph is unavailable, it can also be selected within the range of 0.85 to 0.95 based on the calibration results of the entire vehicle transmission system. This embodiment uses... =0.92. The above vehicle traction power requirements. The relational expression is used to convert real-time driving status into a unified expression of traction power demand, providing a basic input for subsequent dynamic energy demand boundary calculations.

[0033] S2. Combining the fuel cell output capacity, the power battery charging and discharging capacity, and the electric drive system losses, the feasible power domain of the whole vehicle hardware is constructed. The relevant operations in step S2 will be introduced with specific examples: To reflect the hydrogen consumption characteristics of a fuel cell system, a hydrogen consumption model for the fuel cell system is established. Theoretically, the hydrogen consumption of a fuel cell system can be expressed as: ; Based on the polarization curve of the fuel cell, the hydrogen consumption of the fuel cell system can also be expressed as:

[0034] in, For fuel cell output power, The value is acquired in real time through the CAN signal of the fuel cell controller. The lower heating value of hydrogen is taken as 120 MJ / kg. For fuel cell system efficiency, This represents the hydrogen consumption rate function. The above model is used to quantitatively reflect the hydrogen consumption level of fuel cells under different power outputs, thus providing a calculation basis for the economic indicators in the objective function.

[0035] The power of the power battery satisfies the following relationship:

[0036] in, For the power battery side power, in this embodiment, when At that time, the power battery is in a discharging state. At this time, the power battery is in a charging or energy recovery state.

[0037] The SOC (State of Charge) of a power battery over a continuous time period t can be expressed as:

[0038] The efficiency of a power battery is defined as the ratio of battery power to total power, that is:

[0039] Its internal resistance satisfies:

[0040] in, The initial state of charge, For battery charging and discharging efficiency, For power battery capacity, For the power battery current, The internal resistance of the power battery, In charging state. It is in a discharge state. This is due to battery power loss.

[0041] The above model is used to describe the relationship between voltage, current and SOC changes of a power battery under charging and discharging conditions.

[0042] In real-time operation, the vehicle's traction power demand is ultimately shared by the fuel cell system and the power battery system. Therefore, this embodiment combines the fuel cell output capacity, the power battery charging and discharging capacity, and the electric drive system losses to construct the vehicle's hardware power feasible domain, providing hardware-executable conditions for determining the subsequent dynamic energy demand boundary. The feasible power range for the vehicle hardware includes: the maximum power of the vehicle hardware. and the minimum power of the whole vehicle hardware ; , ,in, and These represent the maximum and minimum stable power output of the fuel cell, respectively. It is 50kW. It is 2kW. and These are the maximum allowable discharge power and maximum charging power of the power battery, respectively, provided by the vehicle's battery management system based on SOC, temperature, and SOH data from a lookup table. The equivalent losses of the electric drive system are provided by the battery supplier.

[0043] The feasible domain of the vehicle's hardware power will be used to intersect with the demand boundary in the subsequent online dynamic boundary correction stage, thereby obtaining the final boundary with hardware executability.

[0044] S3, Definition of the The state vector at each control moment , used to calculate dynamic energy demand over a finite time domain in the future; The relevant operations in step S3 will be introduced with specific examples: Unlike existing methods that directly predict the power demand at a single point in time, this invention uses the demand boundary within a finite future time domain as the prediction object.

[0045] First, define the... The state vector at each control moment , used to calculate dynamic energy demand over a finite time domain in the future; ,in, For the first Vehicle speed at a given moment in time. For the first Vehicle acceleration at a control moment For the first Road gradient at a control time For the first Battery state of charge at each control moment For the first The output power of the fuel cell system at each control moment. For the first The power output of the battery at each control moment.

[0046] The power demand range within the next N control steps (finite time domain) is used as the core output of the dynamic energy demand boundary: ,in, For the first The lower bound of the power demand within a prediction step. For the first The upper bound of the required power within a prediction step size. In addition to the upper and lower bounds, this invention also defines a short-time domain fluctuation index. It is used to quantify the degree of fluctuation in power demand over a finite time domain in the future.

[0047] S4, Based on vehicle traction power requirements Build an offline requirement boundary library; The relevant operations in step S4 will be introduced with specific examples: To ensure online computation speed, this invention first constructs a requirement boundary library based on the discretized state grid during the offline phase. For any state grid point... Given a set of reachable accelerations and slope set Under the condition of vehicle traction power demand Find the upper boundary of the vehicle dynamics model respectively and the lower boundary of the vehicle dynamics model ; , ; Taking into account model errors, interpolation errors, and environmental disturbances, this invention also addresses... and Introducing a safety buffer Obtain the robust upper bound. and robust lower boundary ; Consistent with the power dimension, it can be obtained from the statistical quantiles of the absolute values ​​of historical residuals, empirical calibration results, or the Lipschitz upper bound. This embodiment uses the 95th quantile of the absolute values ​​of historical residuals for determination. A minimum buffer of 0.5 kW and a maximum buffer of 5 kW are set, meaning that a basic safety margin is maintained when historical residuals are small, while avoiding excessively widening the boundary when residuals abnormally increase. If the quantile is set to 90%, the boundary is narrower and the energy-saving space is larger, but robustness is reduced; if it is set to 99%, the boundary coverage capability is enhanced, but the space for controller optimization is reduced. In this embodiment, the 95th quantile is used to achieve a trade-off between boundary coverage and energy saving.

[0048] .

[0049] To further improve the coverage of actual working conditions by the boundary, this invention also uses historical multi-working-condition sample sets. Data-driven calibration was performed, and the multi-condition sample set D was acquired through real vehicle road tests.

[0050] In the state neighborhood Within this framework, quantile statistics are performed on historical power demand to obtain the empirical boundary:

[0051] in, The set of sample power requirements within the state neighborhood. express Quantile function.

[0052] Finally, the model boundary and data boundary are conservatively merged to obtain the offline requirement boundary library: , .

[0053] S5. Constructing offline short-time domain fluctuation indicators ; The relevant operations in step S5 will be introduced with specific examples: To describe the drastic changes in power demand in the short-term time domain, this invention simultaneously establishes a fluctuation index model during the offline phase. The nominal evolution trajectory refers to the trajectory from a certain state grid point. Starting from this point, without superimposed random disturbances, residual compensation, or confidence tightening, the acceleration sequence is set according to a preset nominal acceleration sequence. and nominal slope sequence The future state trajectory obtained by forward recursion. In this embodiment, and The values ​​are preferentially obtained from the mean of historical samples within the state's neighborhood; if the number of samples in the state's neighborhood is insufficient, the current acceleration and current slope are kept unchanged. For a given state point... The power sequence for the next N steps is obtained under the nominal evolution trajectory. Therefore, the offline volatility indicator is defined as follows:

[0054] in, and These are all weighting coefficients used to adjust the relative contributions of the power standard deviation term and the power amplitude term to the fluctuation index. They can be satisfied through offline cross-validation, engineering calibration, or normalization. The value is determined by the method, and the range is 0 to 1. In this embodiment, the value is... , N is the total number of time steps in the short time domain, ranging from 10 to 30. In this embodiment, N=30, corresponding to a 3-second prediction time domain with a 0.1-second sampling interval. For arbitrary state grid points The future number obtained under the nominal evolutionary trajectory Power requirements per step Let be the mean of this short-time power sequence. and These represent the maximum and minimum values ​​of the sequence over the next N steps, respectively. Offline fluctuation indicators can characterize the load fluctuation intensity corresponding to different state points in advance without increasing the burden of online optimization. This allows the subsequent controller to use narrower boundaries to improve economy under stable operating conditions, and to increase safety margins and adjust control costs in advance under highly dynamic operating conditions, thereby reducing the risks of frequent power changes in fuel cells and unreasonable charging and discharging of power batteries.

[0055] S6, according to The corresponding offline boundary value is retrieved from the offline demand boundary library, and multidimensional interpolation and residual compensation calculations are performed on the offline boundary value to obtain the dynamic initial energy demand boundary in the future finite time domain. The relevant operations in step S6 will be introduced with specific examples: During the online operation phase, the first step is to determine the current vehicle status. Retrieve the corresponding offline boundary value from the offline boundary database. and ,against and Perform multidimensional interpolation to obtain the upper boundary of the initial demand in the future finite time domain. and the lower boundary of initial demand within the finite time domain in the future ; Due to model mismatch, external disturbances, and state observation errors during actual driving, this invention further introduces an online residual compensation mechanism. The deviation between the current actual requirements and the model requirements is defined as: The deviation between the current actual demand and the model demand is defined as: ,in For the first The actual power demand of the vehicle at each control moment. For the first The vehicle traction power requirement at each control moment.

[0056] The disturbance estimate is updated using an exponential weighting method. , ,in, For the first The disturbance estimate at each control moment. It is a forgetting factor. The smaller the current residual The higher the weight, the more suitable it is for strong transient conditions such as rapid acceleration and sudden changes in gradient; The larger the value, the more historical residuals are retained, which is suitable for stable operating conditions such as high-speed cruise and can suppress boundary jitter caused by measurement noise. This embodiment takes... It is used to balance the response speed to sudden loads and the smoothness of the boundary.

[0057] The demand boundary after the above online residual supplementation includes: the upper boundary of the dynamic initial energy demand in the future finite time domain. and the lower boundary of the dynamic initial energy requirement within the finite time domain in the future ; , .

[0058] The demand boundary, after the aforementioned online residual compensation, can be quickly shifted and updated according to the current operating condition deviation, enhancing its adaptability to sudden load changes and prediction errors. Specifically, Reflecting the instantaneous deviation between actual power demand and model traction power, the exponentially weighted update transforms this deviation into a continuous disturbance estimate. When the vehicle enters a working condition involving changes in gradient, load, or sensor system deviation, This allows for synchronous translation of the upper and lower boundaries, avoiding reliance solely on offline boundaries that could lead to actual needs falling outside the boundaries; as disturbances gradually disappear, the forgetting factor... make Smooth decay prevents a single anomalous sampling point from having a long-term impact on subsequent boundaries. Thus, the online residual compensation mechanism achieves real-time correction of model bias and environmental disturbances with relatively low computational cost.

[0059] S7, according to The confidence level of the state estimation is tightened to improve the confidence level of the dynamic initial energy demand boundary within the future finite time domain. Then, the intersection of the tightened demand boundary and the feasible power domain of the vehicle hardware is obtained to obtain the dynamic energy demand boundary in the future finite time domain. and ; and It can be used for controller calls.

[0060] The relevant operations in step S7 will be introduced with specific examples: After completing residual compensation, this invention further tightens the boundary online based on the confidence level estimated by the current state. Define the first... The reliability of the state estimation at each control time step is: The standard deviation of the residuals is .in, It can be derived from the residual sequence of the most recent M control periods. The calculated variance can also be updated online using exponential weighted variance. This embodiment uses a sliding window statistical method with M=100, according to... calculate, The mean of the residuals within the sliding window. This is the index value of the residual sequence.

[0061] according to and Construct confidence tightening quantity , Wherein, γ is an adjustment coefficient, which can be determined by offline calibration or boundary coverage target, and its value ranges from 0.5 to 3.0. In this embodiment, γ = 1.5 is used, and... The limit is no more than 40% of the compensated boundary width. When the size is small, the boundary tightening is weak, the controller has a larger optimization space but its resistance to uncertainty is reduced; when A larger value results in a more conservative boundary, which can improve security under low-confidence conditions, but may sacrifice some energy-saving potential. The value in this embodiment is used to achieve a balance between state estimation error, residual fluctuation, and energy consumption optimization space.

[0062] according to right and By tightening the confidence level, we obtain the upper bound of the dynamic energy demand over a finite time domain in the future. and the lower boundary of dynamic energy demand in the future finite time domain ; , ; Subsequently, the boundary, after being adjusted by the residual compensation mechanism and confidence level tightening, is intersected with the feasible domain of the vehicle's hardware power within the future finite time domain to obtain the final dynamic energy demand boundary: , ; The final dynamic energy demand boundary ensures that the output boundary not only updates in line with actual load changes but also always falls within the power range executable by the fuel cell system and the power battery system. Specifically, the residual compensation mechanism corrects systematic biases in model predictions, confidence tightening actively narrows the boundary to reserve a safety margin when state estimation is unreliable or residual fluctuations are large, and hardware power feasible region intersection restricts the algorithm boundary to a range executable by the fuel cell, power battery, and electric drive system. The synergistic effect of these three mechanisms enables the final boundary to both track the actual load and prevent the lower-level controller from generating power commands beyond its hardware capabilities.

[0063] S8, according to and renew Online short-time domain fluctuation index was obtained. .

[0064] The relevant operations in step S8 will be introduced with specific examples: After obtaining the final dynamic energy demand boundary ( and Following this, the present invention further updates the online volatility index based on the boundary sequence. Let the boundary center trajectory be: Let the trajectory of the boundary center be:

[0065] according to , and The update process of the online short-term fluctuation index can be obtained as follows: and The value can be determined through offline operating condition verification or controller sensitivity calibration, and the range is 0 to 1; when more attention is paid to boundary uncertainties, the value can be appropriately increased. When more attention is paid to the rate and smoothness of power change, the power consumption can be appropriately increased. This embodiment takes , It is used to simultaneously evaluate changes in boundary width and boundary center.

[0066] Finally, in online applications, this invention outputs the following three signals: (1) Lower boundary of demand power corresponding to each prediction step size in the future finite time domain ; (2) Upper boundary of demand power corresponding to each prediction step size in the future finite time domain ; (3) Short-term fluctuation index corresponding to the current moment .

[0067] The above results can be directly provided to the energy management controller of fuel cell vehicles for setting control constraints, adjusting cost functions, or selecting control strategies.

[0068] The boundary tracking result diagram of the dynamic energy demand boundary under WLTC conditions obtained by the method described in Example 1 is shown below. Figure 3 As shown. By Figure 3It is evident that, under conditions of frequent acceleration and deceleration and significant fluctuations in power demand during WLTC operation, the upper and lower boundaries output by the method described in this invention can change synchronously with the peak and valley values ​​of the vehicle's traction power. When the vehicle accelerates rapidly or accelerates again, the upper boundary rises in advance to cover the impending high power demand; when the vehicle decelerates or operates under low load, the lower boundary decreases accordingly while maintaining a reasonable margin. This result demonstrates that the dynamic energy demand boundary can transform future short-term power demand changes into time-varying constraints that the controller can directly invoke, thereby providing a basis for subsequent energy management to adjust the power sharing between the fuel cell and the power battery in advance.

[0069] The boundary tracking results of the dynamic energy demand boundary under HWFET operating conditions obtained by the method described in Example 1 are shown in the figure. Figure 4 As shown, by Figure 4 It can be seen that under HWFET high-speed cruising conditions, the vehicle power demand changes relatively smoothly. The overall width of the dynamic energy demand boundary output by this invention is smaller than that under WLTC conditions, and the boundary center changes slowly with cruising power, only expanding moderately during vehicle speed adjustment. This result indicates that the method can adaptively adjust the boundary range according to the intensity of operating condition fluctuations, reducing unnecessary conservative margins under stable operating conditions, and maintaining coverage of actual demand during power change phases.

[0070] Example 2 This embodiment further defines Embodiment 1. It provides comparative experiments under five WLTC conditions, comparing the results on termination SOC, hydrogen fuel consumption, equivalent hydrogen fuel consumption, and percentage improvement in energy efficiency based on rule-based strategies, rule-based strategies combined with dynamic energy demand boundaries, MPC (Model Predictive Control) strategies, and MPC strategies combined with dynamic energy demand boundaries. The rule-based strategy refers to a common rule-based energy management strategy in existing fuel cell vehicle engineering applications. This strategy determines the fuel cell output power and power cell compensation power based on the power battery SOC threshold, the fuel cell's efficient operating range, and the required power threshold. For example, it increases the fuel cell output when the SOC is below a set lower limit and decreases the fuel cell output when the SOC is above a set upper limit or the required power is low. This type of strategy does not perform finite-domain predictive optimization. The experimental results are as follows: Figure 5 As shown, by Figure 5It can be seen that under five WLTC operating conditions, the equivalent hydrogen fuel consumption based on the rule-based strategy is 1726.098g, which is reduced to 1661.285g after introducing the dynamic energy demand boundary, representing an energy saving optimization of 3.755%. In contrast, under the MPC strategy, the equivalent hydrogen fuel consumption of MPC alone is 1580.101g, which is further reduced to 1496.731g after introducing the dynamic energy demand boundary, representing an overall energy saving of 13.288% compared to the rule-based strategy. This indicates that in operating conditions like WLTC, where acceleration and deceleration are frequent and power demand fluctuates significantly, the DEDB (the dynamic energy demand boundary obtained by the method described in Example 1) can effectively characterize the range of power demand changes in the short time domain in advance. This allows the controller to no longer passively allocate power based solely on the current single-point power, but to adjust the power sharing ratio between the fuel cell and the power battery in advance according to the upper and lower bounds of future power demand, thereby reducing the inefficient output and frequent power adjustments of the fuel cell system in the high-fluctuation range. Especially when combined with MPC, DEDB (the dynamic energy demand boundary obtained by the method described in Example 1) provides a more reasonable feedforward boundary constraint for prediction optimization, making the optimization space of MPC more in line with the actual load change trend. Therefore, the equivalent hydrogen consumption is further reduced, showing a stronger comprehensive energy saving effect.

[0071] Example 3 This embodiment further defines Embodiment 1. This embodiment provides comparative experiments under 14 HWFET operating conditions, based on rule-based strategies, rule-based strategies combined with dynamic energy demand boundaries, MPC (Model Predictive Control) strategies, and MPC strategies combined with dynamic energy demand boundaries, regarding the termination SOC, hydrogen fuel consumption, equivalent hydrogen fuel consumption, and percentage improvement in energy saving effect. The experimental results are as follows: Figure 6 As shown, by Figure 6It can be seen that under 14 HWFET operating conditions, the equivalent hydrogen fuel consumption based on the rule-based strategy is 2665.608 g, which is reduced to 2633.061 g after introducing DEDB, with an energy saving optimization of 1.221%. The equivalent hydrogen fuel consumption of the MPC strategy is 2511.587 g, while MPC+DEDB further reduces it to 2502.888 g, with a total energy saving of 6.104% compared to the rule-based strategy. Compared with WLTC, HWFET is a highway operating condition with relatively stable vehicle speed and less power demand fluctuation than the urban cycle operating condition. Therefore, the boundary prediction and fluctuation constraint benefits brought by DEDB are relatively weakened, but further optimization can still be achieved at the equivalent hydrogen consumption level. It is particularly noteworthy that the direct hydrogen fuel consumption of MPC+DEDB in HWFET is 2366.356 g, which is close to that of MPC (2363.039 g). However, its equivalent hydrogen fuel consumption is lower, indicating that DEDB does not simply pursue the lowest instantaneous hydrogen consumption. Instead, it achieves a more balanced coordination between fuel cell output, power battery SOC changes, and future demand boundaries, making the energy consumption evaluation results more favorable after considering battery energy changes.

[0072] Based on the experimental results of Examples 2 and 3, the advantages of DEDB are mainly reflected in three aspects. First, it can transform the uncertainty of future power demand into upper and lower boundaries that can be directly invoked by the controller, so that the energy management strategy changes from a single-point response to feedforward optimization under boundary constraints.

[0073] Secondly, it can significantly suppress the blind distribution of power in highly dynamic operating conditions such as WLTC, reduce the inefficient operation of fuel cells and unreasonable charging and discharging of power batteries, thus making the energy saving improvement more obvious.

[0074] Third, in relatively stable operating conditions such as HWFET, although power fluctuations are small and the marginal benefit of DEDB decreases, the system can still maintain better overall equivalent energy consumption during long-term cruise by adjusting the power boundary and constraining the energy distribution range. This shows that dynamic energy demand boundary can not only improve energy saving under complex operating conditions, but also enhance the adaptability and stability of energy management strategies under different driving cycles.

[0075] Example 4 This embodiment further defines Embodiment 1, providing an ablation test on a dynamic initial energy demand boundary tightening strategy. Under five WLTC conditions, using the equivalent hydrogen fuel consumption of 1580.101 g for the standalone MPC strategy as a control, three configurations were set: "residual compensation only," "confidence tightening only," and "this scheme (residual compensation + confidence tightening + hardware intersection)." The termination SOC, equivalent hydrogen fuel consumption, equivalent energy saving rate relative to the standalone MPC strategy, boundary coverage, and average boundary width were statistically analyzed. The results are shown in Table 1. Table 1. Results of ablation experiments using a dynamic energy demand boundary tightening strategy.

[0076] As shown in Table 1, when only residual compensation is used, the boundary can be shifted as a whole with the model deviation, thus reducing the equivalent hydrogen consumption by 3.814% compared to MPC alone. However, residual compensation only controls the position of the boundary center, not the width of the boundary. When the residual fluctuates drastically or the reliability of the state estimation decreases, the shift operation may be contaminated by noise, while the boundary width remains unchanged from the original width of the offline library. This fails to provide an additional safety margin for such uncertainties, resulting in a boundary coverage rate of only 96.72%. When using only residual compensation, the boundary coverage is relatively low. Confidence tightening can address this issue specifically. Confidence tightening constructs a tightening amount by jointly using the state estimation confidence level and the residual statistical standard deviation. When uncertainty increases, it proactively shifts the lower bound upward and the upper bound downward, shrinking the boundary inward to increase the safety buffer. When confidence is high and residuals are stable, the tightening amount approaches zero, and the boundary basically maintains its original width, maximizing energy-saving space. When confidence decreases or residual fluctuations intensify, the tightening amount increases accordingly, and the boundary automatically narrows to cope with uncertainty.

[0077] When only confidence tightening is used, the boundary coverage rate increases to 97.85%, but it can only change the boundary width and cannot eliminate the systematic deviation between model requirements and actual requirements. Therefore, the energy saving rate is 3.333%.

[0078] However, when residual compensation and confidence tightening are used in combination, the boundary may exceed the physical capabilities of the underlying hardware under certain extreme conditions due to excessive contraction or drastic translation. This embodiment further introduces a hardware intersection mechanism, which takes the intersection of the tightened demand boundary and the feasible power range of the vehicle hardware, and uses the common part of the overlap between the two as the final executable boundary of the output. This "intersection" operation ensures that the power command received by the controller will never exceed the physical limits of the fuel cell and the power battery.

[0079] This solution combines residual compensation, confidence tightening, and hardware intersection, resulting in an equivalent hydrogen consumption of 1496.731g, a 5.276% energy saving compared to MPC alone, an increased boundary coverage of 99.41%, and a reduction in average boundary width to 10.35 kW.

[0080] The above results demonstrate that the synergy of these three elements forms a complete, in-depth security defense: residual compensation, which quickly approximates the actual load through boundary shifting; confidence tightening, which proactively reserves a safety buffer when the state is uncertain; and hardware intersection, which forcibly truncates at physical limits, ensuring that all optimizations remain within the hardware's executable range. In the ablation experiment, the joint configuration achieved an optimal overall performance of 5.276% energy saving, 99.41% coverage, and a boundary width of 10.35 kW, proving the effectiveness of this progressive collaborative architecture.

Claims

1. A method for calculating the dynamic energy demand boundary of a fuel cell electric vehicle, characterized in that, The method includes the following steps: S1. Obtain the real-time status of the vehicle during operation and calculate the vehicle's traction power requirement. ; S2. Combining the fuel cell output capacity, the power battery charging and discharging capacity, and the electric drive system losses, the feasible power domain of the whole vehicle hardware is constructed. S3, Definition of the The state vector at each control moment , used to calculate dynamic energy demand over a finite time domain in the future; S4, Based on vehicle traction power requirements Build an offline requirement boundary library; S5. Constructing offline short-time domain fluctuation indicators ; S6, according to The corresponding offline boundary value is retrieved from the offline demand boundary library, and multidimensional interpolation and residual compensation calculations are performed on the offline boundary value to obtain the dynamic initial energy demand boundary in the future finite time domain. In step S6, according to Retrieve the corresponding offline boundary value from the offline requirement boundary library. and ; against and Perform multidimensional interpolation to obtain the upper boundary of the initial demand in the future finite time domain. and the lower boundary of initial demand within the finite time domain in the future ; against and By performing residual compensation, the upper boundary of the initial dynamic energy demand in the future finite time domain is obtained. and the lower boundary of the dynamic initial energy requirement within the finite time domain in the future , , ,in, For the first The disturbance estimate at the first control moment is obtained through the first control moment. The actual power demand of the vehicle at the first control moment and the first control moment Calculation of the deviation of vehicle traction power demand at each control moment; S7, according to The confidence level of the state estimation is tightened to improve the confidence level of the dynamic initial energy demand boundary within the future finite time domain. Then, the intersection of the tightened demand boundary and the feasible power domain of the vehicle hardware is obtained to obtain the dynamic energy demand boundary in the future finite time domain. and ; In step S7, the confidence tightening quantity is constructed. , ,in, For adjustment coefficients, For the first Reliability of state estimation at each control moment , For the first The statistical standard deviation of residuals at each control time point; Then according to right and By tightening the confidence level, we obtain the upper bound of the dynamic energy demand over a finite time domain in the future. and the lower boundary of dynamic energy demand in the future finite time domain ; ， ； The dynamic energy demand boundary within the finite time domain includes: the upper boundary of the dynamic energy demand within the finite time domain. and the lower boundary of dynamic energy demand in the future finite time domain ; ， ,in, and These represent the maximum and minimum power of the vehicle hardware within a finite future time domain, respectively. S8, according to and renew Online short-time domain fluctuation index was obtained. ; In step S8, The formula for calculation is: in, and All are weighting coefficients.

2. The method for calculating the dynamic energy demand boundary of a fuel cell electric vehicle according to claim 1, characterized in that, The real-time status of the vehicle during operation includes: vehicle mass. longitudinal acceleration of vehicles Vehicle speed and vehicle frontal area ; Vehicle traction power requirements The formula for calculation is: ,in, The equivalent efficiency of the transmission system. This refers to the traction force on the wheel side.

3. The method for calculating the dynamic energy demand boundary of a fuel cell electric vehicle according to claim 2, characterized in that, The feasible range of vehicle hardware power includes: the maximum power of vehicle hardware. and the minimum power of the whole vehicle hardware ; , ,in, and These represent the maximum and minimum stable power output of the fuel cell, respectively. and These refer to the maximum allowable discharge power and maximum charging power of the power battery, respectively. This represents the equivalent losses of the electric drive system.

4. The method for calculating the dynamic energy demand boundary of a fuel cell electric vehicle according to claim 3, characterized in that, No. The state vector at each control moment ,in, For the first Vehicle speed at a given moment in time. For the first Vehicle acceleration at a control moment For the first Road gradient at a given time point For the first Battery state of charge at each control moment For the first The output power of the fuel cell system at each control moment. For the first The power output of the battery at each control moment.

5. The method for calculating the dynamic energy demand boundary of a fuel cell electric vehicle according to claim 4, characterized in that, In step S4, for any state grid point Given a set of reachable accelerations and slope set Under the conditions, based on Find the upper boundary of the vehicle dynamics model respectively and the lower boundary of the vehicle dynamics model ; ， ； against and Introducing a safety buffer Obtain the robust upper bound. and robust lower boundary ; The offline demand boundary library includes: offline demand upper boundary. and the lower boundary of offline requirements ; , ,in, and These are the empirical upper and lower boundaries, obtained by performing quantile statistics on historical demand power within the state neighborhood.

6. The method for calculating the dynamic energy demand boundary of a fuel cell electric vehicle according to claim 5, characterized in that, Offline short-term fluctuation index The relation is: in, and All are weighting coefficients. This represents the total number of time steps within the short-time domain. For arbitrary state grid points Obtaining the future from the nominal evolutionary trajectory A power sequence with a step size, for The mean, for In the future The maximum value of each step size. for In the future The minimum value of each step size.

Images