Hydrogen and oxygen supply method for hydrogen fuel cell emergency power supply vehicle

Abstract

The invention discloses a hydrogen and oxygen supply method for a hydrogen fuel cell emergency power supply vehicle. The method comprises the following steps: 1, providing a hydrogen and oxygen supply system for the hydrogen fuel cell emergency power supply vehicle; 2, information acquisition; 3, data processing; the FL-LADRC algorithm disclosed by the invention has the characteristics of simple structure, low calculation burden and low implementation cost on the premise of ensuring the disturbance suppression capability, and is particularly suitable for being used by an embedded platform or resource-constrained equipment. According to the method, the problem that the cathode pressure cannot be measured is solved by introducing the sliding mode differential observer, and the dynamic coupling relation between the oxygen flow and the cathode pressure is compensated in real time in combination with the improved expansion state observer, so that the operation risks such as oxygen starvation and flooding are effectively eliminated; in addition, compared with a traditional PID control algorithm, the fuel cell air supply system adopts the IGA-BP prediction and the NNMPC control algorithm, the performance is remarkably improved, and the urgency and stability requirements of the emergency power supply vehicle in the power protection operation can be better met.

Description

Technical Field

[0001] This invention relates to the field of fuel cells, specifically to a method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle. Background Technology

[0002] With the rapid development of new energy technologies, hydrogen fuel cells are increasingly becoming an ideal power source for emergency power vehicles due to their advantages such as high efficiency, cleanliness, long range, and fast response. Hydrogen fuel cell emergency power vehicles are widely used in scenarios such as communication support, power restoration, and field rescue, which place higher demands on the stability of output power and dynamic response capabilities. As an emergency power source, fuel cells face challenges such as slow dynamic response and difficulty in black start. A key factor affecting dynamic response is the hysteresis, nonlinearity, and strong coupling characteristics of the fuel cell's hydrogen and oxygen supply systems, which pose difficulties for the stable output of fuel cell emergency power. The following are the technical bottlenecks: Hydrogen fuel cell hydrogen and oxygen supply control response lag Traditional systems typically employ proportional-integral-derivative (PID) control strategies to regulate hydrogen and oxygen supply. However, PID is not robust to sudden load changes or system nonlinear modeling errors, and is prone to response lag, oscillations during regulation, or large overshoot, which affects the stability of output power.

[0003] The fuel cell output is heavily coupled with the gas supply system. Under complex operating conditions, the output power of a fuel cell is highly coupled with the supply of hydrogen and oxygen, and the precise amount of hydrogen and oxygen supplied cannot be obtained simply based on the power output. If effective decoupling control cannot be achieved, it can easily lead to asynchrony between gas supply and load demand, resulting in output fluctuations or even system instability.

[0004] The steady-state control accuracy is insufficient, making it difficult to cope with disturbed environments. Existing hydrogen and oxygen supply systems typically consist of gas cylinders, pressure reducing valves, mass flow meters, solenoid valves, and pipeline systems. However, the supporting control systems often suffer from large steady-state errors and insufficient system disturbance immunity when dealing with frequent load changes or complex external environmental disturbances due to a lack of accurate modeling and effective feedforward compensation mechanisms. This makes it difficult to achieve high-precision and robust gas supply control. Summary of the Invention

[0005] The purpose of this invention is to provide a method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle, which can improve the steady-state accuracy and dynamic robustness of the fuel cell air supply system, thereby solving the existing technical defects and unmet technical requirements.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle, comprising the following steps: I. A hydrogen and oxygen supply system for a hydrogen fuel cell emergency power vehicle is provided. 1.1) Hydrogen supply system, including: A hydrogen source, wherein the hydrogen source provides hydrogen supply through a hydrogen source output pipeline; A proportional valve, which is connected in series in the hydrogen source output pipeline, is used to initially regulate the hydrogen flow rate; The ejector is connected to a proportional valve via a hydrogen source output pipe; A hydrogen circulation pump is connected to the bypass of the ejector, and the two are connected in parallel to the anode side of the fuel cell stack. An electric fuel cell stack, wherein the anode receives hydrogen gas and the cathode is supplied with oxygen gas, an electrochemical reaction occurs to generate electrical energy, which powers an externally connected load.

[0007] In this application, the automatic adjustment of the ejector requires stable and high-flow operating conditions, while the hydrogen recirculation pump can maintain the required hydrogen supply to the system through precise control when the flow rate is low. Therefore, the parallel configuration of the hydrogen recirculation pump and the ejector allows the system to maintain normal operation in different working conditions by selecting the most efficient method, thereby enhancing the system's flexibility and robustness. The hydrogen recirculation mode using the parallel connection of the recirculation pump and the ejector combines the advantages of both to adapt to different operating conditions: under low-flow conditions, the hydrogen recirculation pump can provide precise flow control to maintain stable anode inlet pressure; when the flow rate demand is high, the ejector can efficiently handle large amounts of hydrogen, reducing the burden on the hydrogen recirculation pump, thereby improving the overall efficiency of the system.

[0008] 1.2) Oxygen supply system, including: An air filter for filtering air; An air compressor, which is connected to an air filter, is used to pressurize the filtered air; An intercooler, which is connected to an air compressor, is used to cool the pressurized air; A humidifier, which is connected to an intercooler, is used to humidify the cooled air; An electric fuel cell stack, wherein the anode receives hydrogen gas and the cathode is supplied with oxygen gas, an electrochemical reaction occurs to generate electrical energy, and the electrical energy is supplied to an externally connected load. A pressure sensor, wherein the pressure sensor is disposed at the cathode of the fuel cell for gas pressure; A sliding mode differential observer, electrically connected to a pressure sensor, is used to process signals from a cathode pressure sensor in the fuel cell stack to estimate the rate of change of cathode gas pressure. A back pressure valve, which is connected to the cathode of the fuel cell stack, is used to regulate the pressure of the gas generated after the cathode reaction of the fuel cell stack before releasing it. Before entering the air compressor, air is filtered by an air filter to remove particulate matter and harmful sulfides from the fuel cell stack. The air compressor uses high-speed rotating blades to drive the airflow rapidly, increasing the intake volume and pressure, thereby producing a higher oxygen concentration and improving the power density of the fuel cell. An intercooler cools the pressurized, high-temperature air, ensuring the fuel cell stack operates within a suitable temperature range. A humidifier moistens the reactant gas at the stack inlet to maintain adequate humidification of the proton exchange membrane, reducing mass transfer and ohmic losses. The supply line connects the air compressor to the fuel cell stack, after which air enters the cathode cavity, passes through the gas diffusion layer, and reaches the catalyst layer to undergo the oxygen reduction reaction. Gas exiting the cathode flows through the exhaust pipe to the back pressure valve at the end, and is finally released back into the atmosphere.

[0009] II. Information Acquisition 2.1) Real-time acquisition of the fuel cell's operating status; 2.2) Hydrogen supply data signal acquisition; 2.3) Obtain dynamic load current; III. Data Processing 3.1) Analyze the operating status of the fuel cell; 3.2) Data processing on the hydrogen side 3.2.1) Denoise the data on the hydrogen side; In order to better simulate the behavior of the fuel cell system and reduce the interference of noise signals in the subsequent neural network training process, this paper uses a joint algorithm of empirical mode decomposition and wavelet thresholding to denoise all the original experimental data, so as to ensure that the model can fit the true state of the system.

[0010] 3.2.2) Normalize the data after denoising. In this application, the significant differences in the order of magnitude of the units may lead to slow convergence of the neural network and long training time. To eliminate the influence of the units on the model fitting effect, this paper uses a normalization method based on linear transformation algorithm to scale the features to the range of [0,1].

[0011] 3.3) Neural Network Optimization 3.3.1) The processed signal is trained using BP-CNN; Backpropagation (BP) neural networks possess excellent nonlinear mapping and generalization capabilities, making them suitable for handling large amounts of training data. According to Kolmogorov's theorem, a single-hidden-layer neural network can theoretically approximate any continuous function and also reduce the risk of overfitting. Therefore, considering the complexity of fuel cell hydrogen supply systems, this paper uses a three-layer BP neural network to build a PEMFC data-driven model.

[0012] 3.3.2) Update the network using stochastic gradient descent; 3.3.3) Set up the IGA algorithm to optimize the BP neural network; This application presents a BP neural network MPC predictive controller optimized by an immune genetic algorithm. The BP neural network possesses excellent nonlinear mapping capabilities, while the immune genetic algorithm (IGA) can optimize for potential issues such as poor training performance or overfitting. By using the neural network to predict the system state and obtaining the optimal control quantity for the system in the future finite time domain through MPC rolling optimization, the parallel operation of the hydrogen circulation pump and ejector can be better coordinated, exhibiting superior control performance. Finally, for special cases where the measurement process of control targets such as anode pressure is limited, a sliding mode observer for the hydrogen supply system is constructed, and its robustness is verified through noise simulation.

[0013] 3.3.4) Output anode pressure and hydrogen metering ratio; 3.3.5) Obtain the output content from step 3.3.4), predict and control the NNNMPC algorithm through a neural network model, and perform adaptive learning and real-time feedback; 3.3.6) Based on the final anode pressure and optimized hydrogen metering ratio output by NNMPC, dynamically adjust the opening of the proportional valve of the hydrogen supply system to control the hydrogen flow rate input to the fuel cell stack.

[0014] Model predictive control (MPC) is an optimization-based control method that adjusts future control inputs through online optimization to achieve precise regulation of the system output within a predetermined time frame. In MPC control, a process model is used to predict the system output over a future period. Based on this output, the optimizer calculates a series of future control inputs by minimizing the control objective function, which typically includes the error between the predicted output and the expected output, as well as variations in the control input. Therefore, MPC can adjust the control inputs in real time to ensure the system moves along the desired trajectory while satisfying other conditions such as system constraints and terminal constraints. In the control scenario of a fuel cell hydrogen supply system, the MPC algorithm has additional advantages over PID in certain aspects: Firstly, due to the various dynamic processes in the hydrogen supply system, traditional control frameworks often face complex problems such as parameter interactions and couplings when handling MIMO systems. MPC can directly handle various complex interactions in MIMO systems, ensuring that control decisions always consider the dynamic behavior of the overall system without additional decoupling operations. Secondly, the operation of fuel cells often involves a series of constraints, such as pressure and flow ranges, which the MPC controller can directly consider and optimize in real time.

[0015] In summary, for the control scenario of fuel cell hydrogen supply systems, the MPC control algorithm offers greater fault tolerance, but it also increases the accuracy requirements of the system model in MPC. Adopting data-driven machine learning modeling methods instead of traditional mechanistic modeling can leverage the advantages of machine learning models in identifying complex nonlinear relationships, effectively handling and adapting to changes and unknown disturbances that the system may encounter in actual operation, and improving the model's adaptability and robustness.

[0016] 3.4) Data processing on the oxygen side 3.4.1) Calculate the control target based on the load current in step 3.3); 3.4.2) The control target is shaped by a first-order low-pass filter to obtain a stable reference input; 3.4.3) Using a sliding mode differential observer and a soft measurement method, the unmeasurable output variables on the oxygen supply side of the fuel cell are estimated in real time to obtain the deviation between the actual output and the reference input; 3.4.4) The deviation is used as the input of the linear error feedback control law, and feedback compensation calculation is performed to obtain a new virtual control input; 3.4.5) The linear extended state observer estimates the uncertainty of the oxygen supply system, and in combination with these uncertainties, the error of the new virtual control input is compensated and eliminated, and the new virtual control input is converted into a real control quantity; 3.4.6) Apply the actual control quantity to the PEMFC air supply system as the air input quantity of the oxygen supply system.

[0017] Preferably, the hydrogen supply data signal acquisition in step 2.2) includes: hydrogen mass flow rate, anode pressure, hydrogen supply system temperature, fuel cell stack humidity, fuel cell stack output power, circulating pump speed, proportional valve opening degree, and hydrogen concentration.

[0018] Preferably, the specific content of the noise reduction processing of the hydrogen-side data in step 3.2.1) is as follows: 3.2.1.1) Receive the signal acquired in step 2.2); 3.2.1.2) Using the EMD algorithm, the noisy signal is deconstructed sequentially from high frequency to low frequency based on the time scale to obtain different characteristics. The data sequence of the characteristic scale, i.e., the intrinsic mode function; 3.2.1.3) By analyzing IMFs using the sample entropy algorithm, IMF components with higher sample entropy values ​​indicate that they contain more noise.

[0019] 3.2.1.4) Remove modal components with high sample entropy; 3.2.1.5) Perform wavelet threshold denoising on the retained IMFs components; In this application, the EMD and wavelet thresholding joint denoising algorithm can effectively remove noise interference in the original data while preserving the output characteristics.

[0020] 3.2.1.6) Divide the denoised original data into training set, validation set and test set; The specific content of normalizing the denoised data in step 3.2.2) is as follows: using a normalization method based on a linear transformation algorithm, the features are scaled to the range of [0,1]. The transformation function is as follows: In the formula, X represents the original data; Xmin and Xmax represent the maximum and minimum values ​​of the data, respectively.

[0021] Preferably, the specific content of training the processed signal using BP-CNN in step 3.3.1) is as follows: 3.3.1.1) A PEMFC data-driven model is built using a three-layer BP neural network; 3.3.1.2) Define r, s, and u as the number of nodes in the input layer, hidden layer, and output layer, respectively. Let i, j, and k represent the i-th (i=1,2,...,r), j-th (j=1,2,...,u), and k-th (k=1,2,...,u) neurons in the input, hidden, and output layers, respectively. The input layer network input is x. i The network outputs of hidden layer neuron j and output layer neuron k are h, respectively. j and y k The activation function g(x) is the sigmoid function. The activation function graph shows ω. 1ij and ω 2jk These are the weights between nodes i and j and nodes j and k, respectively, and the biases of nodes j and k are b1j and b2k, respectively.

[0022] 3.3.1.3) The training process of a neural network can be divided into two parts: forward propagation and backward propagation. During forward propagation, the outputs of each neuron in the hidden layer and the output layer are as follows: 3.3.1.4) During training, the loss function is used to measure the model's predicted values. The degree of difference between the actual value y and the mean squared error (MSE) is a commonly used method, which calculates... The loss is represented by the mean of the squared differences between y and y. The loss function can be expressed by the following formula: In the formula, yi is the actual value; —Predicted value; N —Sample size.

[0023] The specific content of updating the network using stochastic gradient descent in step 3.3.2) includes: When the error between the output and the expected value does not reach the range required by the algorithm, the backpropagation phase begins. At this point, the weights and biases of the network are adjusted layer by layer from the output layer to the input layer according to the gradient descent method. The weight or bias S update rule according to the stochastic gradient descent method is as follows: In the formula, η is the learning rate; —The gradient of parameter S at time t, denoted as ∇S.

[0024] Preferably, step 3.3.3) involves setting the IGA algorithm to optimize the BP neural network; 3.3.3.1) Assuming the immune system is composed of... m It consists of 10 antibodies (corresponding to network model parameters), each antibody having a length of 1000. n This indicates the number of optimization variables, where the antigen is the training data accuracy target. 3.3.3.2) Generating the initial population pop and the corresponding response function value fitness The population size is N, and the best individual in the population is preserved. best.pop and corresponding optimal value best.f ; 3.3.3.3) Calculate the similarity and aggregation fitness among antibodies in the population, and obtain a new population based on the aggregation fitness function value of the antibodies and the roulette wheel selection mechanism. pop ; 3.3.3.4) Employ pairwise crossover and single-point mutation strategies, and perform crossover and mutation operations on the population pop according to the given crossover and mutation probabilities; 3.3.3.5) Calculate the population similarity of the antibodies. If the population similarity exceeds 0.7, randomly generate new individuals and add them to the group. pop among; 3.3.3.6) Calculation pop The fitness value corresponding to each antibody fit Then update the current best individual. best.pop and corresponding optimal value best.f And begin the next iteration; 3.3.3.7) The algorithm terminates when the number of iterations exceeds the given maximum number of iterations. In the IGA algorithm used, the population size was set to 50, the crossover rate to 0.8, the mutation rate to 0.05, and the number of evolutionary iterations to 100.

[0025] Preferably, in step 3.3.5), obtaining the output content from step 3.3.4) and using the neural network model to predict and control the NNMPC algorithm for adaptive learning and real-time feedback specifically involves the following: Based on the constructed IGA-BP neural network model, the fuel cell control method can be expressed as follows: In the formula, x(k) is the state of the hydrogen supply system at time k, and u(k) is the control input of the system at time k.

[0026] The state vector x(k) is composed of the output parameters in weights 3.3.4: in, This is the anode pressure. This is the hydrogen metering ratio.

[0027] Preferably, the specific content of calculating the control target based on the load current in step 2.3) in step 3.4.1) is as follows: Based on the oxygen supply system, the air compressor voltage and back pressure valve opening range are discretized under different load current conditions. The system over-oxygen ratio and cathode pressure that achieve the maximum net output power of the fuel cell stack under each current group are calculated as the expected control target of the system under the current load current. The net output power is defined as the difference between the output power of the fuel cell stack and the parasitic power consumption of the system auxiliary components.

[0028] The specific content of obtaining a stable reference input by first-order low-pass filtering and shaping the control target in step 3.4.2) is as follows: Since the system output uses the cathode intake flow rate instead of the oxygen ratio, it is necessary to obtain a function representing the desired intake flow rate. :

[0029] in, It's the superoxide ratio, I st Let n be the load current, n be the number of cells in the fuel cell stack, and F be the Faraday constant. The molar mass of oxygen This represents the volume fraction of oxygen in the cathode gas.

[0030] After passing through a first-order low-pass filter, the desired input of the intake airflow is... After filtering and shaping, it becomes y1*, the expected input of the cathode pressure. After filtering and shaping, it becomes y2*.

[0031] Preferably, in step 3.4.3), the sliding mode differential observer is used in a soft measurement mode to estimate the unmeasurable output variables on the oxygen supply side of the fuel cell in real time, and the specific content of the deviation between the actual output and the reference input is as follows: Unknown external disturbances may exist within the system. The unmodeled dynamic behavior of the system is considered as internal disturbance. The uncertainty of the model, consisting of both internal and external disturbances, is considered as the total system disturbance, represented by the disturbance term [d1d2]. T If we express this as an expression, then the system model with a disturbance term after feedback linearization can be represented as: in, E ( x The system matrix A(x) is the decoupling matrix used to achieve accurate linearization of a strongly coupled air supply system; the system matrix A(x) is used in the feedback linearization process to jointly construct the linearized system relationship with the decoupling matrix E(x); u i It controls the input quantity. Assume d i (i=1,2) The amplitude is bounded and differentiable, let φ i For the disturbance d i The rate of change of , modify the above formula, and change the disturbance d i Expand into a new state variable x 3,i : The linearly extended state observer LESO for this system is: In the formula, β j,i (j=1,2,3) are the LESO gain parameters, and the observer tracking error δ is defined. j,i =x j，i -z j，i The dynamic equation of the observer's tracking error can then be expressed as: Rewrite the above equation as a state-space equation: Let the characteristic polynomial of the above expression be the Hurwitz polynomial S. 3 +β 1,i S 2 +β 2,i S+β 3，i =(S+ω oi ) 3 The observer bandwidth ω oi >0. At φ i Under bounded conditions, the observer satisfies bounded-input bounded-output stability; therefore, the parameters to be tuned for the extended-state observer are given by β. j，i Simplified to ω oi : For the observer bandwidth ω oi Tuning is performed to make the observer's state variable z j，i It can accurately estimate the expansion state variables, x, in real time. j，i Therefore, the system model with disturbance terms after feedback linearization can be rewritten as: .

[0032] In this application, d represents the total system disturbance (internal disturbance + external disturbance), ϕi represents the rate of change of the disturbance, x represents the actual state (including the extended state), z represents the observer-estimated state, and β represents the observer gain. This transforms the "unmeasurable and uncertain" disturbance into an "estimateable and compensable" state, allowing the fuel cell control strategy to truly be implemented from an "ideal model" into a "real system."

[0033] Preferably, in step 3.4.4), the deviation is used as the input of the linear error feedback control law to perform feedback compensation calculation and obtain a new virtual control input; The system is simplified and improved using a linear state error feedback control law. After feedback linearization, the system input and output become two independent channels. Therefore, two classic closed-loop PID controls are used to set the virtual control input: in, It is a virtual control input. This is the actual output of the system. It outputs the expected reference value. It is the proportional coefficient, the differential coefficient, and the integral coefficient.

[0034] Define systematic error e i =y i *−y i (i=1,2), the systematic error equation is obtained: .

[0035] Preferably, in step 3.4.5), the linear expansion state observer estimates the uncertainty of the oxygen supply system. Based on these uncertainties, the error of the (new) virtual control input is compensated and eliminated. The specific content of converting the (new) virtual control input into a real control quantity is as follows: Let the characteristic polynomial S 3 +k Di S 2 +k pi S+k Ii =(S+ω ci ) 3 Therefore, the controller gain parameter can be simplified to the controller bandwidth ω. ci Adjustment: Finally, the true control input of the linearized air supply system can be redefined as: .

[0036] Compared with the prior art, the beneficial effects of the present invention are: 1. The FL-LADRC algorithm of this invention, while ensuring disturbance suppression capability, features a simple structure, low computational burden, and low implementation cost, making it particularly suitable for embedded platforms or resource-constrained devices. This method solves the problem of unmeasurable cathode pressure by introducing a sliding mode differential observer (SMDO), and combines it with an improved extended state observer (LESO) to compensate for the dynamic coupling relationship between oxygen flow and cathode pressure in real time. This effectively eliminates operational risks such as "oxygen starvation" and "flooding," improving the steady-state accuracy and dynamic robustness of the fuel cell air supply system. Furthermore, the use of IGA-BP prediction and NNMPC control algorithms significantly improves performance compared to traditional PID control algorithms, better meeting the urgency and stability requirements of emergency power vehicles during power supply operations. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the overall structure of the hydrogen and oxygen supply system of a hydrogen fuel cell emergency power vehicle according to the present invention. Figure 2 This is a schematic diagram of the hydrogen supply system in this invention; Figure 3 Here is a schematic diagram of the oxygen supply system in this invention: Figure 4 This is a model structural diagram of the air compressor in this invention; Figure 5 This refers to the data signals of the hydrogen supply system that need to be collected in this invention; Figure 6 This is a flowchart of the EMD + wavelet threshold joint denoising process in this invention; Figure 7 This is a schematic diagram of the three-layer BP neural network structure in this invention; Figure 8 This is a flowchart of the BP neural network training process in this invention; Figure 9 This is a flowchart of the immune genetic learning process in this invention; Figure 10 This is a diagram of the NNMPC control structure in this invention; Figure 11 This is a structural diagram of the ADRC in this invention; Figure 12 This is a control block diagram of the ADRC (Advanced Control Regulator) of the PEMFC (Pneumatic Air Supply System) in this invention. Figure 13 This is a control block diagram of the PEMFC air supply system FL-LADRC in this invention; Detailed Implementation

[0038] The following will refer to the appendices in the embodiments of the present invention. Figure 1-12The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0039] Please see Figure 1-12 Embodiments of the present invention: Example

[0040] A method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle includes the following steps: I. Provide a hydrogen and oxygen supply system for a hydrogen fuel cell emergency power vehicle (such as...) Figure 1 (As shown) 1.1) Hydrogen supply system (such as...) Figure 2 (as shown), including: A hydrogen source, wherein the hydrogen source provides hydrogen supply through a hydrogen source output pipeline; A proportional valve, which is connected in series in the hydrogen source output pipeline, is used to initially regulate the hydrogen flow rate; The ejector is connected to a proportional valve via a hydrogen source output pipe; A hydrogen circulation pump is connected to the bypass of the ejector, and the two are connected in parallel to the anode side of the fuel cell stack. An electric fuel cell stack, wherein the anode receives hydrogen gas and the cathode is supplied with oxygen gas, an electrochemical reaction occurs to generate electrical energy, which powers an externally connected load.

[0041] 1.2) Oxygen supply system (such as...) Figure 3 (as shown), including: An air filter for filtering air; An air compressor, which is connected to an air filter, is used to pressurize the filtered air; The air compressor model consists of three parts: one part is a voltage-controlled motor drive model; another part is a static MAP diagram of the air compressor, defining the relationship between the compressor's outlet mass flow rate, speed, and pressure ratio (the ratio of outlet pressure to inlet pressure); and the third part is a thermodynamic calculation model of the air temperature and humidity after the air compressor compresses the gas. (e.g.) Figure 4 (As shown) An intercooler, which is connected to an air compressor, is used to cool the pressurized air; The high-temperature compressed air leaving the air compressor can damage the proton exchange membrane. Therefore, an intercooler is designed to cool the high-temperature air before it enters the cathode of the fuel cell stack.

[0042] A humidifier, which is connected to an intercooler, is used to humidify the cooled air; The airflow cooled by the intercooler needs to be humidified by a humidifier to achieve the target humidity and improve the overall performance of the fuel cell.

[0043] An electric fuel cell stack, wherein the anode receives hydrogen gas and the cathode is supplied with oxygen gas, an electrochemical reaction occurs to generate electrical energy, and the electrical energy is supplied to an externally connected load. A pressure sensor, wherein the pressure sensor is disposed at the cathode of the fuel cell for gas pressure; A sliding mode differential observer, electrically connected to a pressure sensor, is used to process signals from a cathode pressure sensor in the fuel cell stack to estimate the rate of change of cathode gas pressure. A back pressure valve, which is connected to the cathode of the fuel cell stack, is used to regulate the pressure of the gas generated after the cathode reaction of the fuel cell stack before releasing it. II. Information Acquisition 2.1) Real-time acquisition of the fuel cell's operating status; 2.2) Hydrogen supply data signal acquisition; The data signals acquired for hydrogen supply include: hydrogen mass flow rate, anode pressure, hydrogen supply system temperature, fuel cell stack humidity, fuel cell stack output power, circulating pump speed, proportional valve opening, and hydrogen concentration. (e.g.) Figure 5 (As shown) This paper describes the design of a hydrogen supply system monitoring platform using LabVIEW based on the CAN communication protocol. The platform transmits the collected data to a host computer monitoring system for analysis and storage, completing the acquisition of hydrogen supply data signals. The hardware for the monitoring system uses the ZLG-CAN communication module and USBCAN-II to implement data communication between the terminals. LabVIEW receives data by calling the ControlCAN.dll driver of the USBCAN module and utilizing the VCI_Receive function.

[0044] 2.3) Obtain dynamic load current; IV. Data Processing 3.1) Analyze the operating status of the fuel cell; 3.2) Data processing on the hydrogen side 3.2.1) Denoise the data on the hydrogen side; (e.g.) Figure 6 (As shown) 3.2.1.1) Receive the signal acquired in step 2.2); 3.2.1.2) Using the EMD algorithm, the noisy signal is deconstructed sequentially from high frequency to low frequency based on the time scale to obtain data sequences of different feature scales, i.e., intrinsic mode functions; 3.2.1.3) By analyzing IMFs using the sample entropy algorithm, IMF components with higher sample entropy values ​​indicate that they contain more noise.

[0045] 3.2.1.4) Remove modal components with high sample entropy; 3.2.1.5) Perform wavelet threshold denoising on the retained IMFs components; Specifically: In wavelet thresholding denoising, a noisy signal x of length N... N It can be represented as the original signal f N With noise signal e N The superposition is shown in the following formula: In the formula: f N —The original signal; e N —Gaussian noise signal.

[0046] 3.2.1.6) Divide the denoised original data into training set, validation set and test set; Specifically, the training objective is to fit the mapping relationship between the input and output, without needing to consider the temporal sequence of the data. Therefore, the denoised original data is divided into a training set, a validation set, and a test set in an 8:1:1 ratio.

[0047] 3.2.2) Normalize the data after denoising. Specifically, the normalization method based on linear transformation algorithm is used to scale the features to the range [0,1]. The transformation function is as follows: In the formula, X represents the original data; Xmin and Xmax represent the maximum and minimum values ​​of the data, respectively.

[0048] 3.3) Neural Network Optimization 3.3.1) Train the processed signal using BP-CNN; (e.g.) Figures 7-8 (As shown) 3.3.1.1) A PEMFC data-driven model is built using a three-layer BP neural network; 3.3.1.2) Define r, s, and u as the number of nodes in the input layer, hidden layer, and output layer, respectively. Let i, j, and k represent the i-th (i=1,2,...,r), j-th (j=1,2,...,u), and k-th (k=1,2,...,u) neurons in the input, hidden, and output layers, respectively. The input layer network input is x. i The network outputs of hidden layer neuron j and output layer neuron k are h, respectively. j and y k The activation function g(x) is the sigmoid function. The activation function graph shows ω.1ij and ω 2jk These are the weights between nodes i and j and nodes j and k, respectively, and the biases of nodes j and k are b1j and b2k, respectively.

[0049] 3.3.1.3) The training process of a neural network can be divided into two parts: forward propagation and backward propagation. During forward propagation, the outputs of each neuron in the hidden layer and the output layer are as follows:

[0050] 3.3.1.4) During training, the loss function is used to measure the model's predicted values. The degree of difference between the actual value y and the mean squared error (MSE) is a commonly used method, which calculates... The loss is represented by the mean of the squared differences between y and y. The loss function can be expressed by the following formula: In the formula, yi is the actual value; —Predicted value; N —Sample size.

[0051] The specific content of updating the network using stochastic gradient descent in step 3.3.2) includes: When the error between the output and the expected value does not reach the range required by the algorithm, the backpropagation phase begins. At this point, the weights and biases of the network are adjusted layer by layer from the output layer to the input layer according to the gradient descent method. The weight or bias S update rule according to the stochastic gradient descent method is as follows: In the formula, η is the learning rate; —The gradient of parameter S at time t, denoted as ∇S.

[0052] In this embodiment, the training process of the single-hidden-layer BP neural network training algorithm is as follows: Initialization: Training dataset: X_train, Y_train; Test dataset: X_test, Y_test; Validation dataset: X_val, Y_val; Normalize X_train to Data_input and Y_train to Data_target, with a range of [0,1]. Step 1: Obtain the number of neurons in the input layer (inputNum) and the number of neurons in the output layer (outputNum), and set the number of neurons in the hidden layer (hiddenNum); Step 2: Create a neural network Net with hiddenNum neurons, set the number of iterations M, the learning rate η, and the target error goal; Step 3: Update the network using gradient descent, with the loss function being L(θ); Step 4: Train the network using the train function in MATLAB; 3.3.2) Update the network using stochastic gradient descent; 3.3.3) Set up the IGA algorithm to optimize the BP neural network; (e.g.) Figure 9 (As shown) The initial parameters of a backpropagation (BP) neural network have a significant impact on its performance. The appropriateness of the initial weights and biases determines the network's learning and training speed and its ability to converge to the global optimum. However, the initial parameters of BP neural networks are mostly set empirically, resulting in considerable uncertainty. Therefore, this paper constructs an IGA-BP neural network by combining an Immune Genetic Algorithm (IGA). By applying the IGA to the optimization of the initial weights and biases of the BP network model, the two algorithms complement each other, enabling the IGA-BP neural network algorithm to possess both self-learning and global optimal parameter search characteristics.

[0053] 3.3.3.1) Assuming the immune system is composed of... m It consists of 10 antibodies (corresponding to network model parameters), each antibody having a length of 1000. n This indicates the number of optimization variables, where the antigen is the training data accuracy target. 3.3.3.2) Generating the initial population pop and the corresponding response function value fitness The population size is N, and the best individual in the population is preserved. best.pop and corresponding optimal value best.f ; 3.3.3.3) Calculate the similarity and aggregation fitness among antibodies in the population, and obtain a new population based on the aggregation fitness function value of the antibodies and the roulette wheel selection mechanism. pop ; 3.3.3.4) Employ pairwise crossover and single-point mutation strategies, and perform crossover and mutation operations on the population pop according to the given crossover and mutation probabilities; 3.3.3.5) Calculate the population similarity of the antibodies. If the population similarity exceeds 0.7, randomly generate new individuals and add them to the group. pop among; 3.3.3.6) Calculation pop The fitness value corresponding to each antibody fit Then update the current best individual. best.pop and corresponding optimal value best.f And begin the next iteration; 3.3.3.7) The algorithm terminates when the number of iterations exceeds the given maximum number of iterations. In the IGA algorithm used, the population size was set to 50, the crossover rate to 0.8, the mutation rate to 0.05, and the number of evolutionary iterations to 100.

[0054] 3.3.4) Output anode pressure and hydrogen metering ratio; 3.3.5) Obtain the output content from step 3.3.4), predict and control the NNNMPC algorithm through a neural network model, and perform adaptive learning and real-time feedback; The Neural Network-Model Predictive Control (NNMPC) algorithm combines the concepts of artificial neural networks and predictive control algorithms. Leveraging the ability of neural networks to fit complex nonlinear functions, it achieves more accurate capture of the dynamic behavior of complex systems. In controller design, the input and output values ​​of the controlled system are used as data sources. The network is trained to learn the dynamic characteristics of the system. By adjusting the parameters of the neural network, a nonlinear mapping relationship of the system is established. The neural network model is then used as the predictive model for MPC, thus constructing the NNMPC controller.

[0055] The NNMPC control process relies on the adaptive learning of the neural network model and the real-time feedback of model predictive control, enabling it to better adapt to dynamic changes in the system. In a hydrogen supply system application scenario where a hydrogen circulation pump and ejector are connected in parallel, the ejector's efficiency is low under low load conditions. In this case, the MPC expresses the input weight relationship as an optimization problem, coordinating the proportional valve opening control signal ufcv and the circulation pump voltage control signal ubl, thereby comprehensively regulating the anode pressure and optimizing the hydrogen supply. Under high load conditions, the ejector's efficiency can meet most flow rate changes. In this case, according to the control objective, the circulation pump control voltage ubl mainly plays a regulatory role. (The NNMPC control structure diagram is shown below.) Figure 10 (As shown) In step 3.3.5), obtaining the output from step 3.3.4) and using the neural network model to predict and control the NNMPC algorithm for adaptive learning and real-time feedback, the specific content is as follows: Based on the constructed IGA-BP neural network model, the fuel cell control method can be expressed as: In the formula, x(k) is the state of the hydrogen supply system at time k, and u(k) is the control input of the system at time k.

[0056] The state vector x(k) is composed of the output parameters in weights 3.3.4: in, This is the anode pressure. This is the hydrogen metering ratio.

[0057] Given the nonlinear characteristics of the model, the Finite Control Set (FCS) online optimization solver can search for the optimal control action within a finite set of control actions instead of directly optimizing the continuous control variables.

[0058] Furthermore, fuel cell hydrogen supply systems face various physical and operational constraints. FCS-MPC can directly consider these constraints during the optimization process and, through a closed-loop constraint-based online optimization control strategy, enables MPC to handle system constraints in real time, adapting to system changes and external disturbances. Through online optimization and closed-loop feedback, MPC provides excellent system output performance while ensuring all system constraints are met.

[0059] Both the air compressor voltage and the back pressure valve opening, when used as control command inputs, simultaneously affect the cathode intake flow rate and cathode pressure. Therefore, for the strongly nonlinear and strongly coupled PEMFC air supply system, a suitable decoupling control strategy needs to be designed. Active disturbance rejection control (ADRC), as a decoupling control method, has advantages such as low computational cost, low requirement for model accuracy, and good robustness. Based on these advantages, ADRC is selected as the decoupling control method for the air supply system to achieve decentralized control of cathode flow rate and pressure. The main concept of ADRC is to actively extract disturbance information from the input / output signals before the disturbance significantly affects the final output of the system, and quickly take control actions to suppress the disturbance, thereby effectively reducing its impact on the controlled variable. The ADRC structure is as follows: Figure 11 As shown, it mainly consists of three parts: Tracking Differentiator (TD), Extended State Observer, and Nonlinear State Error Feedback (NLSEF).

[0060] For a strongly nonlinear and strongly coupled two-input, two-output air supply system, an ADRC controller can be embedded between the air compressor voltage input and the superoxide ratio output channels, and another ADRC controller can be embedded between the back pressure valve throttle opening and the cathode pressure output channels. Under the dual ADRC control strategy, the coupling disturbances of flow and pressure, as well as external disturbances to the system, are converted into the total disturbance estimated by the ESO of each channel, and then compensated and eliminated in their respective NLSEF, thereby achieving dynamic decoupling of the system. The ADRC control block diagram of the PEMFC air supply system is shown below. Figure 12As shown in the diagram. The sliding mode differential observer (SMA) is used to estimate the unmeasurable cathode pressure in real time and calculate the system's oxygen ratio. Based on this, two independent ADRC controllers are introduced into the system flow and pressure channels for separate control. The specific control process is as follows: First, the oxygen ratio and the expected cathode pressure, which vary with the load current, are input into the TD to effectively acquire the differential signal and arrange an appropriate transient process to reduce the large overshoot caused by sudden changes in the system control target. Then, the total system disturbance, including coupled interference, estimated by the ESO, is compensated into the control quantity. Combined with the error feedback of the NLSEF, the two control quantities of the system are finally obtained, weakening the coupling effect between the air supply system's flow and pressure and improving the control effect.

[0061] 3.3.6) Based on the final anode pressure and optimized hydrogen metering ratio output by NNMPC, dynamically adjust the opening of the proportional valve of the hydrogen supply system to control the hydrogen flow rate input to the fuel cell stack.

[0062] 3.4) Data processing on the oxygen side 3.4.1) Calculate the control target based on the load current in step 2.3); The specific content of calculating the control target based on the load current in step 2.3) in step 3.4.1) is as follows: Based on the oxygen supply system, the air compressor voltage and back pressure valve opening range are discretized under different load current conditions. The system over-oxygen ratio and cathode pressure that achieve the maximum net output power of the fuel cell stack under each current group are calculated as the expected control target of the system under the current load current. The net output power is defined as the difference between the output power of the fuel cell stack and the parasitic power consumption of the system auxiliary components.

[0063] 3.4.2) The control target is shaped by a first-order low-pass filter to obtain a stable reference input; The specific content of obtaining a stable reference input by first-order low-pass filtering and shaping the control target in step 3.4.2) is as follows: Since the system output uses the cathode intake flow rate instead of the oxygen ratio, it is necessary to obtain a function representing the desired intake flow rate. :

[0064] Where λ_O_2 is the oxygen overload ratio, Ist is the load current, n is the number of cells in the fuel cell stack, F is the Faraday constant, M_(O_2) is the molar mass of oxygen, and χ_(O_2) is the volume fraction of oxygen in the cathode gas.

[0065] After passing through a first-order low-pass filter, the desired input of the intake airflow is... After filtering and shaping, it becomes y1*, the expected input of the cathode pressure. After filtering and shaping, it becomes y2*.

[0066] 3.4.3) Using a sliding mode differential observer and a soft measurement method, the unmeasurable output variables on the oxygen supply side of the fuel cell are estimated in real time to obtain the deviation between the actual output and the reference input; Unknown external disturbances may exist within the system. The unmodeled dynamic behavior of the system is considered as internal disturbance. The uncertainty of the model, consisting of both internal and external disturbances, is considered as the total system disturbance, represented by the disturbance term [d1d2]. T If we express this as an expression, then the system model with a disturbance term after feedback linearization can be represented as:

[0067] E(x) is the decoupling matrix used to achieve precise linearization of the strongly coupled air supply system; the system matrix A(x) is used in the feedback linearization process to jointly construct the linearized system relationship with the decoupling matrix E(x); ui is the control input, assuming d i (i=1,2) The amplitude is bounded and differentiable, let φ i For the disturbance d i The rate of change of , modify the above formula, and change the disturbance d i Expand into a new state variable x 3,i : The linearly extended state observer LESO for this system is: In the formula, β j,i (j=1,2,3) are the LESO gain parameters, and the observer tracking error δ is defined. j,i =x j，i -z j，i The dynamic equation of the observer's tracking error can then be expressed as: Rewrite the above equation as a state-space equation:

[0068] Let the characteristic polynomial of the above expression be the Hurwitz polynomial S. 3 +β 1,i S 2 +β 2,i S+β 3，i =(S+ω oi ) 3 The observer bandwidth ω oi >0. At φ i Under bounded conditions, the observer satisfies bounded-input bounded-output stability; therefore, the parameters to be tuned for the extended-state observer are given by β. j，i Simplified to ω oi : For the observer bandwidth ωoi Tuning is performed to make the observer's state variable z j，i It can accurately estimate the expansion state variables, x, in real time. j，i Therefore, the system model with disturbance terms after feedback linearization can be rewritten as: .

[0069] 3.4.4) The deviation is used as the input to the linear error feedback control law, and feedback compensation calculation is performed to obtain a new virtual control input; The system is simplified and improved using a linear state error feedback control law. After feedback linearization, the system input and output become two independent channels. Therefore, two classic closed-loop PID controls are used to set the virtual control input: It is a virtual control input. This is the actual output of the system. It outputs the expected reference value. It is a proportionality coefficient. These are differential coefficients. It is the integral coefficient.

[0070] Define systematic error e i =y i *−y i (i=1,2), the systematic error equation is obtained: .

[0071] 3.4.5) The linear extended state observer estimates the uncertainty of the oxygen supply system, and combines these uncertainties to compensate for and eliminate the error of the new virtual control input, converting the new virtual control input into a real control quantity; Let the characteristic polynomial S 3 +k Di S 2 +k pi S+k Ii =(S+ω ci ) 3 Therefore, the controller gain parameter can be simplified to the controller bandwidth ω. ci Adjustment: Finally, the true control input of the linearized air supply system can be redefined as: .

[0072] 3.4.6) Apply the actual control quantity to the PEMFC air supply system.

[0073] The control block diagram of the PEMFC air supply system based on feedback linearization and linear active disturbance rejection is as follows: Figure 13As shown. The system control objective is calculated based on the dynamic load current, and after first-order low-pass filtering and shaping, a smooth system reference input is obtained. The sliding mode differential observer uses a soft-sensor method to estimate the unmeasurable system output variable in real time. The deviation from the parameter input is used as the input of the linear error feedback control law. After feedback compensation calculation, a new virtual control input is obtained, which is then output to the feedback linearization module. Combining the system uncertainty estimated by the linear extended state observer, the feedback linearization module compensates for and eliminates the error and converts the virtual control input into a real control quantity, which is ultimately applied to the PEMFC air supply system to achieve tracking control of the cathode intake flow rate and pressure.

[0074] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0075] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle, characterized by comprising the following steps: I. A hydrogen and oxygen supply system for a hydrogen fuel cell emergency power vehicle is provided. 1.1) Hydrogen supply system, including: A hydrogen source, wherein the hydrogen source provides hydrogen supply through a hydrogen source output pipeline; A proportional valve, which is connected in series in the hydrogen source output pipeline, is used to initially regulate the hydrogen flow rate; The ejector is connected to a proportional valve via a hydrogen source output pipe; A hydrogen circulation pump is connected to the bypass of the ejector, and the two are connected in parallel to the anode side of the fuel cell stack. An electric fuel cell stack, wherein the anode receives hydrogen gas and the cathode is supplied with oxygen gas, an electrochemical reaction occurs to generate electrical energy, and the electrical energy is supplied to an externally connected load. 1.2) Oxygen supply system, including: An air filter for filtering air; An air compressor, which is connected to an air filter, is used to pressurize the filtered air; An intercooler, which is connected to an air compressor, is used to cool the pressurized air; A humidifier, which is connected to an intercooler, is used to humidify the cooled air; An electric fuel cell stack, wherein the anode receives hydrogen gas and the cathode is supplied with oxygen gas, an electrochemical reaction occurs to generate electrical energy, and the electrical energy is supplied to an externally connected load. A pressure sensor, wherein the pressure sensor is disposed at the cathode of the fuel cell for gas pressure; A sliding mode differential observer, electrically connected to a pressure sensor, is used to process signals from a cathode pressure sensor in the fuel cell stack to estimate the rate of change of cathode gas pressure. A back pressure valve, which is connected to the cathode of the fuel cell stack, is used to regulate the pressure of the gas generated after the cathode reaction of the fuel cell stack before releasing it. II. Information Acquisition 2.1) Real-time acquisition of the fuel cell's operating status; 2.2) Hydrogen supply data signal acquisition; 2.3) Obtain dynamic load current; III. Data Processing 3.1) Analyze the operating status of the fuel cell; 3.2) Data processing on the hydrogen side 3.2.1) Denoising the data on the hydrogen side; 3.2.2) Normalize the data after denoising. 3.3) Neural Network Optimization 3.3.1) The processed signal is trained using BP-CNN; 3.3.2) Update the network using stochastic gradient descent; 3.3.3) Set up the IGA algorithm to optimize the BP neural network; 3.3.4) Output anode pressure and hydrogen metering ratio; 3.3.5) Obtain the output content from step 3.3.4), predict and control the NNNMPC algorithm through a neural network model, and perform adaptive learning and real-time feedback; 3.3.6) Based on the final anode pressure and optimized hydrogen metering ratio output by NNMPC, dynamically adjust the opening of the proportional valve of the hydrogen supply system to control the hydrogen flow rate input to the fuel cell stack. 3.4) Data processing on the oxygen side 3.4.1) Calculate the control target based on the load current in step 3.3); 3.4.2) The control target is shaped by a first-order low-pass filter to obtain a stable reference input; 3.4.3) Using a sliding mode differential observer and a soft measurement method, the unmeasurable output variables on the oxygen supply side of the fuel cell are estimated in real time to obtain the deviation between the actual output and the reference input; 3.4.4) The deviation is used as the input to the linear error feedback control law, and feedback compensation calculation is performed to obtain a new virtual control input; 3.4.5) The linear extended state observer estimates the uncertainty of the oxygen supply system, and combines these uncertainties to compensate for and eliminate the error of the new virtual control input, converting the new virtual control input into a real control quantity; 3.4.6) Apply the actual control quantity to the PEMFC air supply system as the air input quantity of the oxygen supply system.

2. The method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle according to claim 1, characterized in that, The hydrogen supply data signals collected in step 2.2) include: hydrogen mass flow rate, anode pressure, hydrogen supply system temperature, fuel cell stack humidity, fuel cell stack output power, circulating pump speed, proportional valve opening degree, and hydrogen concentration.

3. The method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle according to claim 1, characterized in that the specific content of the noise reduction processing of the hydrogen-side data in step 3.2.1) is as follows: 3.2.1.1) Receive the signal acquired in step 3.2); 3.2.1.2) Using the EMD algorithm, the noisy signal is deconstructed sequentially from high frequency to low frequency based on the time scale to obtain data sequences of different feature scales, i.e., intrinsic mode functions; 3.2.1.3) By analyzing IMFs using the sample entropy algorithm, IMF components with higher sample entropy values ​​indicate that they contain more noise; 3.2.1.4) Remove modal components with high sample entropy; 3.2.1.5) Perform wavelet threshold denoising on the retained IMFs components; 3.2.1.6) Divide the denoised original data into training set, validation set and test set; The specific content of normalizing the denoised data in step 3.2.2) is as follows: using a normalization method based on a linear transformation algorithm, the features are scaled to the range of [0,1]. The transformation function is as follows: In the formula, X represents the original data; Xmin and Xmax represent the maximum and minimum values ​​of the data, respectively.

4. The method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle according to claim 3, characterized in that, The specific content of training the processed signal using BP-CNN in step 3.3.1) is as follows: 3.3.1.1) A PEMFC data-driven model is built using a three-layer BP neural network; 3.3.1.2) Define r, s, and u as the number of nodes in the input layer, hidden layer, and output layer, respectively. Let i, j, and k represent the i-th (i=1,2,...,r), j-th (j=1,2,...,u), and k-th (k=1,2,...,u) neurons in the input, hidden, and output layers, respectively. The input layer network input is x. i The network outputs of hidden layer neuron j and output layer neuron k are h, respectively. j and y k The activation function g(x) is the sigmoid function. The activation function graph shows ω. 1ij and ω 2jk These are the weights between nodes i and j and nodes j and k, respectively, and the biases of nodes j and k are b1j and b2k, respectively. 3.3.1.3) The training process of a neural network can be divided into two parts: forward propagation and backward propagation. During forward propagation, the outputs of each neuron in the hidden layer and the output layer are as follows: 3.3.1.4) During training, the loss function is used to measure the model's predicted values. The degree of difference between the actual value y and the mean squared error (MSE) is a commonly used method, which calculates... The loss is represented by the mean of the squared differences between y and y. The loss function can be expressed by the following formula: In the formula, yi is the actual value; —Predicted value; N —Sample size; The specific content of updating the network using stochastic gradient descent in step 3.3.2) includes: When the error between the output and the expected value does not reach the range required by the algorithm, the backpropagation phase begins. At this point, the weights and biases of the network are adjusted layer by layer from the output layer to the input layer according to the gradient descent method. The weight or bias S update rule according to the stochastic gradient descent method is as follows: In the formula, η is the learning rate; —The gradient of parameter S at time t, denoted as ∇S.

5. The method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle according to claim 4, characterized in that, Step 3.3.3) sets up the IGA algorithm to optimize the BP neural network; 3.3.3.1) Assuming the immune system is composed of... m It consists of 10 antibodies (corresponding to network model parameters), each antibody having a length of 1000. n This indicates the number of optimization variables, where the antigen is the training data accuracy target. 3.3.3.2) Generating the initial population pop and the corresponding response function value fitness The population size is N, and the best individual in the population is preserved. best.pop and corresponding optimal value best.f ; 3.3.3.3) Calculate the similarity and aggregation fitness among antibodies in the population, and obtain a new population based on the aggregation fitness function value of the antibodies and the roulette wheel selection mechanism. pop ; 3.3.3.4) Employ pairwise crossover and single-point mutation strategies, and perform crossover and mutation operations on the population pop according to the given crossover and mutation probabilities; 3.3.3.5) Calculate the population similarity of the antibodies. If the population similarity exceeds 0.7, randomly generate new individuals and add them to the group. pop among; 3.3.3.6) Calculation pop The fitness value corresponding to each antibody fit Then update the current best individual. best.pop and corresponding optimal value best.f And begin the next iteration; 3.3.3.7) The algorithm terminates when the number of iterations exceeds the given maximum number of iterations. In the IGA algorithm used, the population size was set to 50, the crossover rate to 0.8, the mutation rate to 0.05, and the number of evolutionary iterations to 100.

6. The method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle according to claim 5, characterized in that, In step 3.3.5), obtaining the output from step 3.3.4) and using the neural network model to predict and control the NNMPC algorithm for adaptive learning and real-time feedback, the specific content is as follows: Based on the constructed IGA-BP neural network model, the fuel cell control method can be expressed as: In the formula, x(k) represents the state of the hydrogen supply system at time k, and u(k) represents the control input of the system at time k. The state vector x(k) is composed of the output parameters in weights 3.3.4: in, For anode pressure, This is the hydrogen metering ratio.

7. A method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle according to claim 6, characterized in that, The specific content of calculating the control target based on the load current in step 3.3) in step 3.4.1) is as follows: Based on the oxygen supply system, the air compressor voltage and back pressure valve opening range are discretized under different load current conditions. The system oxygen ratio and cathode pressure that achieve the maximum net output power of the fuel cell stack under each current group are calculated as the expected control target of the system under the current load current. The net output power is defined as the difference between the output power of the fuel cell stack and the parasitic power consumption of the system auxiliary components. The specific content of obtaining a stable reference input by first-order low-pass filtering and shaping the control target in step 3.4.2) is as follows: Since the system output uses the cathode intake flow rate instead of the oxygen ratio, it is necessary to obtain a function representing the desired intake flow rate. : in, It's the superoxide ratio, I st Let n be the load current, n be the number of cells in the fuel cell stack, and F be the Faraday constant. The molar mass of oxygen The volume fraction of oxygen in the cathode gas; the desired input of the intake gas flow rate after passing through a first-order low-pass filter. After filtering and shaping, it becomes y1*, the expected input of the cathode pressure. After filtering and shaping, it becomes y2*.

8. The method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle according to claim 7, characterized in that, In step 4.3.3), the sliding mode differential observer is used in a soft measurement mode to estimate the unmeasurable output variables on the oxygen supply side of the fuel cell in real time, and the specific content of obtaining the deviation between the actual output and the reference input is as follows: Unknown external disturbances may exist within the system. The unmodeled dynamic behavior of the system is considered as internal disturbance. The uncertainty of the model, consisting of both internal and external disturbances, is considered as the total system disturbance, represented by the disturbance term [d1 d2]. T If we express this as an expression, then the system model with a disturbance term after feedback linearization can be represented as: in, E ( x The system matrix A(x) is the decoupling matrix used to achieve accurate linearization of a strongly coupled air supply system. The system matrix A(x) is used in the feedback linearization process, together with the decoupling matrix E(x), to construct the linearized system relationship. i It controls the input quantity, let's assume d i (i=1,2) The amplitude is bounded and differentiable, let φ i For the disturbance d i The rate of change of , modify the above formula, and change the disturbance d i Expand into a new state variable x 3,i : The linearly extended state observer LESO for this system is: In the formula, β j,i (j=1,2,3) are the LESO gain parameters, and the observer tracking error δ is defined. j,i =x j，i -z j，i The dynamic equation of the observer's tracking error can then be expressed as: Rewrite the above equation as a state-space equation: Let the characteristic polynomial of the above expression be the Hurwitz polynomial S. 3 +β 1,i S 2 +β 2,i S+β 3，i =(S+ω oi ) 3 The observer bandwidth ω oi >0, in φ i Under bounded conditions, the observer satisfies bounded-input bounded-output stability; therefore, the parameters to be tuned for the extended-state observer are given by β. j，i Simplified to ω oi : For the observer bandwidth ω oi Tuning is performed to make the observer's state variable z j，i It can accurately estimate the expansion state variables, x, in real time. j，i Therefore, the system model with disturbance terms after feedback linearization can be rewritten as: .

9. A method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle according to claim 8, characterized in that, In step 3.4.4), the deviation is used as the input of the linear error feedback control law to perform feedback compensation calculation and obtain a new virtual control input. The system is simplified and improved using a linear state error feedback control law. After feedback linearization, the system input and output become two independent channels. Therefore, two classic closed-loop PID controls are used to set the virtual control input: in, It is a virtual control input. This is the actual output of the system. It outputs the expected reference value. It is a proportionality coefficient. These are differential coefficients. It is the integral coefficient; Define systematic error e i =y i *−y i (i=1,2), we obtain the systematic error equation: Define the systematic error e i =y i *−y i (i=1,2), the systematic error equation is obtained: .

10. A method for supplying hydrogen and oxygen to a hydrogen fuel cell emergency power vehicle according to claim 9, characterized in that, In step 3.4.5), the linear expansion state observer estimates the uncertainty of the oxygen supply system. Based on these uncertainties, the error of the (new) virtual control input is compensated and eliminated. The specific content of converting the (new) virtual control input into a real control quantity is as follows: Let the characteristic polynomial S 3 +k Di S 2 +k pi S+k Ii =(S+ω ci ) 3 Therefore, the controller gain parameter can be simplified to the controller bandwidth ω. ci Adjustment: Finally, the true control input of the linearized air supply system can be redefined as: .

Images