Adaptive virtual impedance low voltage ride through method for hydrogen fuel cell systems

Abstract

The application provides a kind of adaptive virtual impedance low voltage ride-through method suitable for hydrogen fuel cell system, it is related to hydrogen fuel cell power generation control technical field.The application obtains the current tolerance and reactive power support boundary in future time window accurately by using dynamic capability prediction model after triggering low voltage ride-through mode by real-time monitoring grid-connected point voltage and fuel cell body state, provides accurate constraint for adaptive adjustment of virtual impedance.Based on this boundary and voltage drop characteristics, a multi-objective optimization model is constructed, and the optimal virtual impedance value and power command change trend are obtained by solving, overcoming the mismatch problem of fixed virtual impedance.Further combined with the dynamic adjustment of converter and fuel cell controller, the virtual impedance can track the system operating condition changes in real time, while considering the reactive power support demand to suppress fault current impact and avoid oscillation or off-grid.The application improves the ride-through success rate and operation stability of hydrogen fuel cell power generation system under low voltage fault.

Description

Technical Field

[0001] This invention relates to the field of hydrogen fuel cell power generation control technology, and in particular to an adaptive virtual impedance low voltage ride-through method suitable for hydrogen fuel cell systems. Background Technology

[0002] When a hydrogen fuel cell power generation system is connected to the grid, a voltage dip in the grid can cause an overcurrent surge in the converter, threatening the safety of the components. Existing technologies often use the virtual impedance method to suppress fault current, which involves inserting a fixed resistor or inductive component in series in the control loop to increase the equivalent impedance.

[0003] However, the output impedance of a hydrogen fuel cell system exhibits strong nonlinear time-varying characteristics, being affected in real time by operating conditions such as stack temperature, output power, and anode hydrogen pressure. A virtual impedance with fixed parameters cannot adapt to this dynamic change across the entire operating range. When the impedance parameter is too small, fault current suppression is insufficient, easily leading to device damage or protective grid disconnection; when the parameter is too large, excessive current limiting severely weakens the system's reactive power support capability to the grid, and may even trigger system oscillations.

[0004] Therefore, the current virtual impedance technology with fixed parameters has poor compatibility with hydrogen fuel cell systems, affecting the grid connection reliability and safety of hydrogen fuel cell power generation systems. Summary of the Invention

[0005] This invention provides an adaptive virtual impedance low voltage ride-through method suitable for hydrogen fuel cell systems, which solves the problem of poor compatibility between fixed parameter virtual impedance technology and hydrogen fuel cell systems.

[0006] In a first aspect, the present invention provides an adaptive virtual impedance low-voltage ride-through method suitable for hydrogen fuel cell systems. The method includes: real-time monitoring of the grid connection point voltage and key state parameters of the hydrogen fuel cell system; triggering a low-voltage ride-through mode when the grid connection point voltage drops to a preset threshold; predicting the dynamic capability boundary of the hydrogen fuel cell system within a future time window based on the key state parameters using a dynamic capability prediction model, the dynamic capability boundary including a current tolerance boundary and a reactive power support boundary; establishing a multi-objective optimization model with fault suppression and reactive power support as objectives based on the dynamic capability boundary and the voltage drop characteristics of the power grid, and solving for the optimal virtual impedance value and power command change trend; and dynamically adjusting the converter and the hydrogen fuel cell system controller based on the optimal virtual impedance value and power command change trend to achieve adaptive matching of the virtual impedance.

[0007] Secondly, embodiments of the present invention provide an adaptive virtual impedance low-voltage ride-through device suitable for hydrogen fuel cell systems. This device includes a communication module and a processing module. The communication module is used to monitor the grid connection point voltage and key state parameters of the hydrogen fuel cell system in real time. The processing module is used to trigger a low-voltage ride-through mode when the grid connection point voltage drops to a preset threshold. Based on the key state parameters, it predicts the dynamic capability boundary of the hydrogen fuel cell system within a future time window using a dynamic capability prediction model. The dynamic capability boundary includes a current tolerance boundary and a reactive power support boundary. Based on the dynamic capability boundary and the voltage drop characteristics of the power grid, a multi-objective optimization model is established with fault suppression and reactive power support as objectives, and the optimal virtual impedance value and power command change trend are obtained by solving the model. Based on the optimal virtual impedance value and power command change trend, dynamic adjustments are made to the converter and the controller of the hydrogen fuel cell system, respectively, to achieve adaptive matching of the virtual impedance.

[0008] Thirdly, embodiments of the present invention provide an electronic device including a memory and a processor. The memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0009] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0010] This invention provides an adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems. By real-time monitoring of the grid connection point voltage and the fuel cell's internal state, after triggering the low-voltage ride-through mode, this invention utilizes a dynamic capability prediction model to accurately obtain the current tolerance and reactive power support boundaries within the future time window, providing accurate constraints for the adaptive adjustment of the virtual impedance. Based on these boundaries and voltage drop characteristics, a multi-objective optimization model is constructed to solve for the optimal virtual impedance value and power command change trend, overcoming the mismatch problem caused by the inability of a fixed virtual impedance to adapt to the nonlinear output impedance of the fuel cell. Furthermore, dynamic adjustment is achieved by combining the converter and fuel cell controller, enabling the virtual impedance to track changes in system operating conditions in real time. This suppresses fault current surges while simultaneously considering reactive power support requirements, avoiding oscillations or grid disconnection caused by impedance mismatch. This invention improves the ride-through success rate and operational stability of hydrogen fuel cell power generation systems under low-voltage faults, enhancing grid connection reliability and safety. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic flowchart of an adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0014] like Figure 1 As shown, the present invention provides an adaptive virtual impedance low-voltage ride-through method suitable for hydrogen fuel cell systems. The method includes steps S101-S104.

[0015] S101. Real-time monitoring of the grid connection point voltage and key status parameters of the hydrogen fuel cell system.

[0016] In some embodiments, key state parameters of the fuel cell system include stack output power, stack average temperature, anode hydrogen pressure change rate, proton exchange membrane high-frequency impedance, and air compressor speed.

[0017] For example, in this embodiment of the invention, a three-phase voltage transformer can be used to collect the line voltage at the grid connection point. After A / D conversion, a digital signal is obtained, with a sampling frequency of 10kHz, synchronized with the converter control cycle. The sampled voltage is low-pass filtered (cutoff frequency 1kHz) to remove high-frequency harmonic interference from the power grid; abnormal sampled values ​​are eliminated using the 3σ criterion, and abnormal values ​​are replaced by the average of the first three valid sampled values. The effective value of the grid connection point voltage is calculated in real time, with a calculation window of 20ms (one power frequency cycle), and updated on a sliding basis.

[0018] For example, the fuel cell stack output power reflects the current load level of the stack and directly determines its current withstand capability during faults. Higher power means lower allowable overload current. The stack output current and voltage are collected by DC-side current and voltage sensors, and the power (P) is calculated in real time. stack =I dc ×U dc ; For example, the average temperature of the fuel cell stack affects the conductivity and mechanical strength of the proton exchange membrane. When the temperature exceeds 85°C, the current withstand threshold decreases rapidly. PT100 temperature sensors are placed at five locations: the end plates at the beginning and end of the fuel cell stack and the three bipolar plates in the middle. The arithmetic mean of the five measurement points is taken.

[0019] For example, the anode hydrogen pressure change rate reflects the dynamic response capability of the hydrogen supply. When the change rate exceeds ±5 kPa / s, the power fluctuation of the fuel cell stack will increase significantly. A pressure sensor is used to collect the anode inlet hydrogen pressure, and the pressure change within adjacent 10 ms is calculated. .

[0020] For example, the high-frequency impedance of a proton exchange membrane indirectly reflects the water content of the membrane; an impedance exceeding 100 mΩ. At a diameter of cm², membrane dehydration is severe, leading to a decrease in reactive power support capacity. Online measurement was performed using a 1kHz AC injection method, with the injection current amplitude ≤50mA to avoid affecting the normal operation of the fuel cell stack.

[0021] For example, the air compressor speed determines the air supply. When the speed is below 60% of the rated speed, it cannot meet the oxygen demand of high-power dynamic adjustment. A Hall effect speed sensor is used to collect the data, with a resolution of 1 rpm.

[0022] S102. When the grid connection point voltage drops to a preset threshold, the low voltage ride-through mode is triggered. Based on the key state parameters of the system, the dynamic capability boundary of the hydrogen fuel cell system within the future time window is predicted through the dynamic capability prediction model.

[0023] In some embodiments, the preset trigger threshold is 0.9 times the rated voltage. When the effective value of the grid connection point voltage is lower than this threshold for three consecutive power frequency cycles, the low voltage ride-through mode is formally triggered to avoid false triggering caused by instantaneous voltage fluctuations. After triggering, the normal grid connection control parameters are immediately frozen, and the control process under fault conditions is started.

[0024] In some embodiments, the dynamic capability boundary includes a current tolerance boundary and a reactive power support boundary.

[0025] For example, the current withstand boundary is the maximum instantaneous peak current that the fuel cell stack and converter are allowed to output within the next 100ms without permanent damage.

[0026] For example, the reactive power support boundary is the maximum reactive power output value that the fuel cell stack and auxiliary systems can provide within the next 100ms while maintaining stable operation.

[0027] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1023.

[0028] S1021. Based on the key state parameters of the ontology, feature extraction is performed to generate a state time series sequence.

[0029] S1022. Input the state time sequence into the pre-trained dynamic capability prediction model, and output the predicted values ​​of current withstand boundary and reactive power support boundary within the future time window through forward calculation.

[0030] In some embodiments, the pre-trained dynamic capability prediction model is a 2-layer stacked LSTM network with 64 neurons in the hidden layer, 32 neurons in the fully connected layer, and 2 neurons in the output layer (corresponding to the current tolerance boundary and the reactive power support boundary, respectively).

[0031] For example, the forward computation process is as follows: The input layer inputs a 10×5 time series matrix into the first LSTM layer and outputs a hidden state sequence with a dimension of 10×64; the second LSTM layer performs further feature extraction on the hidden state sequence and outputs the hidden state at the last time step (with a dimension of 64); the fully connected layer maps the 64-dimensional features into a 2-dimensional output, namely the predicted values ​​of the current withstand boundary and reactive power support boundary within the next 100ms.

[0032] S1023. Based on the predicted values ​​of current withstand boundary and reactive power support boundary within the future time window, determine the dynamic capability boundary.

[0033] For example, in this embodiment of the invention, the lower limit of the 95% confidence interval of the model output prediction value can be taken as the final dynamic capability boundary, that is: ;in, , The standard deviation of the model's prediction error is obtained through statistics on the training set, ensuring that the boundary has a sufficient safety margin.

[0034] S103. Based on the dynamic capability boundary and the voltage drop characteristics of the power grid, a multi-objective optimization model with fault suppression and reactive power support as objectives is established, and the optimal virtual impedance value and power command change trend are obtained by solving the model.

[0035] In some embodiments, the voltage sag characteristics of the power grid include sag depth and sag rate.

[0036] For example, the depth of the fall: Among them, U rated U is the rated voltage of the power grid. grid This represents the effective value of the current grid connection point voltage. For example, the rate of fall: Where Δt=10ms is the calculation step size.

[0037] As one possible implementation, step S103 can be specifically implemented as steps S1031-S1036.

[0038] S1031. Based on the voltage drop characteristics of the power grid, calculate the reference value of reactive power support required by the power grid.

[0039] For example, when the drop depth d > 0.2, the converter needs to provide reactive power support, and the reference value calculation formula is as follows: Among them, S rated Q represents the rated apparent power of the hydrogen fuel cell system, expressed in kVA; when d ≤ 0.2, Q ref =0, no additional reactive power support is required.

[0040] S1032. With the dual objectives of minimizing fault current suppression deviation and minimizing the deviation between actual reactive power output and reactive power support reference value, and with the current tolerance boundary and reactive power support boundary in the dynamic capability boundary as constraints, a multi-objective optimization model is constructed.

[0041] In some embodiments, the dual objective function is defined as follows: Objective 1 (fault current suppression): Minimize the relative deviation between the actual fault current and the current withstand boundary; Objective 2 (Reactive Power Support): Minimize the relative deviation between the actual reactive power output and the grid reference value; Take the maximum value in the denominator to avoid Q. ref A division by zero error occurs when the value is 0.

[0042] In some embodiments, hard constraints ;in, , For virtual resistance and virtual inductance, , This represents the predicted dynamic capability boundary.

[0043] S1033. Determine the fault severity index based on the voltage drop depth and drop rate in the voltage drop characteristics.

[0044] For example, the calculation of the fault severity index. ; where r max =20V / ms is the preset maximum drop rate, and the value of S is in the range of [0,1]. The larger the value, the more serious the fault.

[0045] S1034. Based on the fault severity index and dynamic capability boundary, dynamically adjust the first weight coefficient and the second weight coefficient.

[0046] In some embodiments, the first weighting coefficient is the weighting coefficient of the fault current suppression deviation, and the second weighting coefficient is the weighting coefficient of the deviation between the actual reactive power output and the reactive power support reference value; the first weighting coefficient and the second weighting coefficient are used to dynamically balance the priorities of fault current suppression and reactive power support.

[0047] For example, the adaptive weight adjustment rule is as follows: First weight coefficient (fault current suppression): w1 = 0.4 + 0.4 × S. Second weight coefficient (reactive power support): w2 = 1. w1. When S≥0.8 (serious fault), w1=0.9 and w2=0.1 are forced to prioritize system safety; when S≤0.2 (minor fault), w1=0.4 and w2=0.6 to prioritize meeting the reactive power demand of the power grid.

[0048] S1035. Based on the dynamically adjusted first and second weight coefficients and the multi-objective optimization model, adaptive adjustment is performed to obtain a multi-objective optimization model with adaptive weights.

[0049] In some embodiments, single-objective transformation involves weighted fusion of the two objective functions into a comprehensive cost function. .

[0050] S1036. Using the model predictive control algorithm, solve the multi-objective optimization model with adaptive weights to obtain the optimal virtual impedance value and the power command change trend.

[0051] In some embodiments, the present invention may employ a rolling time-domain optimized model predictive control algorithm, with a prediction time domain of 20 control cycles (200ms) and a control time domain of 1 control cycle (10ms). In each control cycle, only the optimal solution of the first cycle is executed, and the solution is recalculated in the next cycle.

[0052] For example, step S1036 can be specifically implemented as steps A1-A5.

[0053] A1. In each control cycle, the virtual impedance value of the previous control cycle is used as the reference value. Within a preset range, the virtual resistance value and virtual inductance value in the virtual impedance value are discretized to obtain multiple sets of candidate impedance values.

[0054] In some embodiments, if the current control cycle is the first control cycle after triggering the low voltage ride-through mode, the reference value is zero.

[0055] For example, if the fault severity index of the current control cycle is lower than the preset lower limit, the preset range is determined as the benchmark value.

[0056] For example, the preset lower limit of the fault severity index is 0.3, corresponding to a minor fault scenario with a drop depth ≤30% and a drop rate ≤6V / ms. Each control cycle (10ms) synchronously reads the current fault severity index S, compares it with the preset lower limit, and triggers the corresponding branch logic.

[0057] If S < 0.3, it is judged as a minor fault. At this time, the system current surge is small and the reactive power demand is low, so there is no need to expand the virtual impedance adjustment range. The virtual impedance preset range is directly locked to the basic preset range: virtual resistance Rv∈[0,0.5]Ω, virtual inductance Lv∈[0,5]mH, reducing the number of candidate impedances and compressing the single-cycle solution time to within 0.5ms.

[0058] Another example is that if the fault severity index of the current control cycle is higher than or equal to the preset lower limit, then the scaling factor is determined according to the fault severity index of the current control cycle, wherein the larger the fault severity index, the larger the scaling factor; the upper and lower limits of the basic preset range are multiplied by the scaling factor respectively to obtain the dynamically scaled preset range.

[0059] For example, if S ≥ 0.3, it is judged as a moderate to severe fault. In this case, the adjustment range needs to be expanded to balance current suppression and reactive power support. First, calculate the scaling factor k: The value of k is limited to [1.0, 2.0] to avoid excessively large ranges leading to a surge in computation or excessively small ranges failing to meet optimization requirements. The upper and lower limits of the basic preset range are multiplied by the scaling factor k to obtain the dynamically scaled preset ranges: virtual resistance dynamic range: Rv∈[0, 0.5×k]Ω; virtual inductance dynamic range: Lv∈[0, 5×k]mH. For example, when S=1.0 (most severe fault), k=2.0, the virtual resistance range expands to [0, 1.0]Ω, and the virtual inductance range expands to [0, 10]mH, providing sufficient adjustment margin for extreme fault scenarios.

[0060] The basic preset range is as follows: virtual resistance Rv∈[0,0.5]Ω, virtual inductance Lv∈[0,5]mH; the virtual impedance value is always 0 during normal operation, and the reference value is 0 in the first control cycle after low voltage ride-through trigger.

[0061] For example, the discretization rules are as follows: virtual resistance discretization step length: 0.05Ω, with a total of 11 candidate values; virtual inductance discretization step length: 0.5mH, with a total of 11 candidate values; combined to generate 11×11=121 sets of candidate impedance values, balancing the solution accuracy and calculation speed (single cycle solution time ≤0.8ms).

[0062] Based on the dynamically scaled preset range, while keeping the step size constant (virtual resistance 0.05Ω, virtual inductance 0.5mH), candidate impedance combinations are regenerated. For example, when k=2.0, 21 candidate values ​​are generated for the virtual resistance and 21 candidate values ​​are generated for the virtual inductance, for a total of 441 candidate combinations. The single-cycle solution time is still controlled within 1.2ms, meeting the real-time requirements.

[0063] A2. Based on the grid connection point voltage, output current, and multiple candidate impedance values ​​of the current control cycle, as well as the pre-established converter output current prediction model, the peak fault current and reactive power output value corresponding to each candidate impedance value are predicted.

[0064] For example, the converter output current prediction model input and output: Input: dq-axis component u of the grid-connected voltage of the current control cycle. d u q The dq-axis component of the converter output current i d i q Candidate virtual impedance values ​​(Rv, Lv); Output: Peak fault current for the next control cycle, Ifault,peak = and reactive power output value .

[0065] A3. Based on the objective function of the multi-objective optimization model with adaptive weights, and the peak fault current and reactive power output corresponding to each group of candidate impedance values, calculate the comprehensive cost corresponding to each group of candidate impedance values.

[0066] For example, the comprehensive cost calculation is as follows: Substitute the predicted output of each candidate impedance into the comprehensive cost function J to calculate 121 cost values.

[0067] A4. The set of candidate impedance values ​​with the lowest overall cost is taken as the optimal virtual impedance value for the current control cycle.

[0068] For example, optimal impedance selection: Candidate impedances that satisfy all hard constraints are screened, and the set with the lowest overall cost is selected as the optimal virtual impedance value (R) for the current control cycle. v ,L v If none of the candidate impedances meet the constraints, then the group with the smallest current peak value is selected.

[0069] A5. Calculate the difference between the optimal virtual impedance value and the reference value. Determine the adjustment direction and magnitude of active and reactive power based on the direction and magnitude of the difference, as a trend of power command change.

[0070] For example, the mapping relationship between power adjustment and virtual impedance change: ; where k p =200kW / Ω, k q =100kVar / mH is the proportionality coefficient, which matches the rated power of the system; the negative sign indicates that when the virtual impedance increases, the active and reactive power outputs decrease accordingly.

[0071] For example, power command limits: the adjustment range of active power shall not exceed 30% of the current active power, and the adjustment range of reactive power shall not exceed 50% of the current reactive power support boundary, so as to avoid system instability caused by sudden power changes.

[0072] S104. Based on the optimal virtual impedance value and the power command change trend, dynamic adjustments are made in conjunction with the controllers of the converter and the hydrogen fuel cell system to achieve adaptive matching of the virtual impedance.

[0073] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1046.

[0074] S1041. Embed the virtual resistance and virtual inductance values ​​from the optimal virtual impedance values ​​into the voltage outer loop control circuit of the converter, and generate a compensation voltage signal based on the virtual resistance value, virtual inductance value and the current output current of the converter.

[0075] In some embodiments, the embedding location is precisely between the output of the outer voltage loop PI regulator and the input of the inner current loop, rather than at the output of the inner current loop in the prior art, to avoid affecting the dynamic response speed of the current loop. The accurate formula for the compensation voltage (strictly aligned with the grid-synchronized rotating coordinate system) is as follows: ; Where ω = 314 rad / s is the rated angular frequency of the power grid, i d i q The output current dq-axis component after low-pass filtering (cutoff frequency 500Hz).

[0076] S1042. Superimpose the compensation voltage signal onto the original current reference command generated by the voltage outer loop control loop to generate the corrected current inner loop reference command.

[0077] In some embodiments, the superposition logic involves directly adding the compensated voltage to the original voltage reference command output by the outer loop PI regulator. Limiting protection: The corrected voltage reference command limit is... This avoids current distortion caused by overmodulation.

[0078] S1043. Based on the corrected inner current loop reference command, control the duty cycle of the converter switching transistor to achieve real-time adjustment of the output current.

[0079] In some embodiments, the modulation method is space vector pulse width modulation (SVPWM), which achieves a DC bus voltage utilization rate 15.47% higher than SPWM. The switching frequency is 10kHz, synchronized with the sampling frequency, and the dead time is set to 2μs to prevent bridge arm shoot-through. Current loop control uses a PI regulator with a bandwidth of 1kHz to ensure current tracking error ≤2%.

[0080] S1044. The direction and magnitude of active power adjustment and reactive power adjustment in the power command change trend are synchronously fed forward to the auxiliary system controller of the hydrogen fuel cell system.

[0081] In some embodiments, the feedforward time advance is 50ms, precisely matching the inherent response time of the hydrogen fuel cell auxiliary system (hydrogen injection valve response time 30ms, air compressor response time 40ms). The feedforward signal format uses CAN2.0B bus transmission at a baud rate of 500kbps, with a transmission delay ≤1ms, and includes three parameters: active power adjustment, reactive power adjustment, and adjustment rate.

[0082] S1045, The auxiliary system controller performs linkage adjustment based on the feedforward signal.

[0083] Among them, the opening degree and opening degree change rate of the hydrogen injection valve are pre-adjusted to match the change rate of active power, the speed of the air compressor is pre-increased to ensure the supply of oxygen for the electrochemical reaction required for dynamic adjustment of active power, and the flow rate of the cooling water pump is pre-adjusted to suppress the temperature fluctuation of the fuel cell stack.

[0084] In some embodiments, the hydrogen injection valve is adjusted proportionally according to the active power adjustment amount ΔP. ;in The opening change rate is limited to ±5% / s to avoid drastic fluctuations in hydrogen pressure.

[0085] Air compressor adjustment: Adjust the air compressor speed according to the active power adjustment amount ΔP. Where kcomp = 10 rpm / kW; the rate of change of speed is limited to ±1000 rpm / s to prevent surge.

[0086] Cooling water pump regulation: Adjust the water pump flow rate according to the rate of change of fuel cell stack temperature ΔT / Δt. ;in .

[0087] S1046. During low voltage ride-through, continuously monitor the changes in stack temperature and anode hydrogen pressure. If the temperature exceeds the limit or the hydrogen pressure fluctuation exceeds the safe range, generate a correction signal and feed it back to the auxiliary system controller to perform closed-loop correction of the linkage adjustment amount until the low voltage ride-through mode ends.

[0088] In some embodiments, the safety thresholds are: upper limit of stack temperature 85°C, normal range of anode hydrogen pressure 150~200kPa, and allowable range of hydrogen pressure fluctuation ±10kPa. Correction logic: When the stack temperature exceeds 80°C, the dynamic capability boundary is tightened by reducing the current withstand boundary by 5% for every 1°C increase; when the hydrogen pressure fluctuation exceeds ±10kPa, the active power adjustment rate is immediately reduced to 50% of the original rate.

[0089] This invention provides an adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems. By real-time monitoring of the grid connection point voltage and the fuel cell's own state, after triggering the low-voltage ride-through mode, a dynamic capability prediction model is used to accurately obtain the current tolerance and reactive power support boundaries within the future time window, providing accurate constraints for the adaptive adjustment of the virtual impedance. Based on these boundaries and voltage drop characteristics, a multi-objective optimization model is constructed to solve for the optimal virtual impedance value and power command change trend, overcoming the mismatch problem caused by the inability of a fixed virtual impedance to adapt to the nonlinear output impedance of the fuel cell. Furthermore, dynamic adjustment is achieved by combining the converter and fuel cell controller, enabling the virtual impedance to track changes in system operating conditions in real time. This suppresses fault current surges while also considering reactive power support requirements, avoiding oscillations or grid disconnection caused by impedance mismatch. This invention improves the ride-through success rate and operational stability of hydrogen fuel cell power generation systems under low-voltage faults, enhancing grid connection reliability and safety.

[0090] Optionally, the adaptive virtual impedance low voltage ride-through method for hydrogen fuel cell systems provided in this embodiment of the invention further includes steps S201-S204 before step S102.

[0091] S201. Collect operational data of the hydrogen fuel cell system during the entire fault ride-through process under various operating conditions.

[0092] In some embodiments, the operating conditions include fault ride-through scenarios with different output power, different stack temperature, different anode hydrogen pressure change rate, and different voltage drop depths.

[0093] In some embodiments, the output power covers 10% to 100% of the rated power, with test points set at 10% intervals, and each power point is tested three times; the stack temperature covers 40℃ to 90℃ (the normal operating temperature range of the stack), with test points set at 10℃ intervals; the anode hydrogen pressure change rate covers -10kPa / s to +10kPa / s, including extreme conditions such as insufficient or excessive hydrogen supply; the voltage drop depth covers 10% to 90% of the rated voltage drop, with test points set at 10% intervals, including symmetrical and asymmetrical drops (single-phase and two-phase drops); and the drop duration covers 100ms to 3s.

[0094] S202. Preprocess the operating data under various working conditions to construct a training sample set.

[0095] The input for each sample is a time sequence of state over N consecutive historical moments, and the output is the actual peak current and reactive power output capacity within the future time window corresponding to each historical moment.

[0096] For example, the sliding window parameters are: window length N=10 (corresponding to 10ms of historical data), sliding step size=1, that is, one sample is generated at each time moment; input sample: each sample input consists of 5 key state parameters of the ontology for 10 consecutive historical time moments, forming a time series matrix with a dimension of 10×5; output label: the output of each sample is the actual maximum current peak value and the actual maximum output reactive power value within the next 100ms (20 control cycles) corresponding to that time moment, where: actual current peak value: the maximum value of the DC current output by the stack within the next 100ms; actual reactive power output capability value: the maximum reactive power output value of the converter within the next 100ms without triggering overcurrent protection.

[0097] S203. Construct a long short-term memory network model.

[0098] In some embodiments, the Long Short-Term Memory (LSTM) network model includes an input layer, one or more LSTM hidden layers, a fully connected layer, and an output layer.

[0099] In some embodiments, this invention employs a two-layer stacked LSTM network as the core architecture of the dynamic capability prediction model. The overall forward propagation process is as follows: input time series sequence → first layer LSTM (returns the complete sequence) → first layer Dropout → second layer LSTM (only returns the hidden state at the last moment) → second layer Dropout → first fully connected layer (ReLU activation) → output layer (linear activation). The following sections fully explain the connection relationships and computational logic of each module through formulas from four dimensions: computation within a single LSTM unit, inter-layer connections in stacked LSTMs, Dropout layer processing, and fully connected layer mapping.

[0100] I. Internal Computational Logic of a Single LSTM Unit. The LSTM unit controls the flow and storage of information through four core components: the forget gate, input gate, cell candidate, and output gate, thus solving the gradient vanishing problem of traditional RNNs. For any time t, the input consists of the current feature vector xt and the previous hidden state ht. 1. Cell state ct at the previous moment 1. The output is the current hidden state ht and the current cell state ct.

[0101] 1. Forget Gate: Controls the degree to which the old cell state is preserved. The forget gate outputs a value between 0 and 1, where 1 indicates that the old cell state is completely preserved, and 0 indicates that it is completely discarded. The calculation formula is: Where: σ is the Sigmoid activation function, ; W f Here is the forget gate weight matrix, with dimension (d) input +d hidden ,d hidden ), d input d is the input feature dimension. hidden is the number of LSTM hidden units; bf is the forget gate bias vector with dimension (d hidden ,);[h t 1,x t This indicates that the hidden state from the previous time step is concatenated with the input from the current time step along the feature dimension.

[0102] 2. Input Gate: Controls the degree to which new information is written. The input gate consists of two parts: first, determining which new information needs to be updated; and second, generating candidate cell states. The calculation formula is: ; Where: it is the input gate output, dimension (d) hidden ,); The candidate cell state is generated by the tanh activation function, with a value range of [-1, 1]. , For the input gate weights and biases, , The candidate cell state weights and biases are defined, with dimensions consistent with the parameters corresponding to the forget gate.

[0103] 3. Cell State Update: Fusing old state with new information. The cell state is the "memory carrier" of LSTM, and is updated through the combined action of the forget gate and the input gate. ; where ⊙ represents the Hadamard product (element-wise multiplication), ensuring that the cell state of each dimension is updated independently.

[0104] 4. Output Gate: Controls the output of the hidden state. The output gate determines which information from the cell state will be output as the current hidden state: ; ; where: o t For output gate output, dimension (d) hidden ,);h t The hidden state at the current time step serves as the input for the next time step and the output for the current layer.

[0105] II. Inter-layer connections in a 2-layer stacked LSTM. In this model, the first LSTM layer returns the hidden state sequence for all time steps, while the second LSTM layer only returns the hidden state for the last time step, achieving layer-by-layer extraction and compression of temporal features.

[0106] 1. Connection from the input layer to the first LSTM layer. The input layer receives data in dimensions (batch_size, seq_len, d). input The time series matrix is ​​given by: batch_size (set to 32 during training); seq_len=10 (time step length, corresponding to 10ms of historical data); d input =5 represents the number of key state parameters of the ontology.

[0107] The number of hidden units d in the first LSTM layer hidden1 =64. For each time step t∈{0,1,...,9}, perform the above calculation for a single LSTM unit to obtain the hidden state h for that time step. t (1) and cell state c t (1) Finally, the first LSTM layer outputs the complete hidden state sequence. The dimension is (batch_size, seq_len, d). hidden1 = (batch_size, 10, 64).

[0108] 2. Connection from the first LSTM layer to the second LSTM layer. The hidden state sequence H output by the first layer. (1) It is directly used as the input to the second LSTM layer. The number of hidden units d in the second LSTM layer hidden2 =32, and LSTM unit computation is performed for each time step t, but only the hidden state h9 of the last time step (t=9) is retained. (2) As output, the states at intermediate time steps are discarded. The final output dimension of the second LSTM layer is (batch_size, d...). hidden2 The feature vector of (batch_size, 32) is used to complete the global compression of temporal features.

[0109] III. Calculation Logic of the Dropout Layer. To prevent overfitting, a Dropout layer is added after each LSTM layer. During training, a portion of the neuron outputs are randomly discarded; during testing, the outputs are not discarded but are scaled for compensation. The calculation formula is: h′=h⊙m; where: h is the input to the Dropout layer (i.e., the output of the LSTM layer); m is the mask vector, following a Bernoulli distribution. Bernoulli (1 p), where p=0.2 is the dropout probability; during testing, the Dropout layer output is h′=h×(1 p), to ensure that the expected outputs are consistent during the training and testing phases.

[0110] In this model, the first layer of Dropout acts on H. (1)The output dimension remains (batch_size, 10, 64); the second-level Dropout operates on h9. (2) The output dimension remains (batch_size, 32).

[0111] IV. Feature Mapping and Output of Fully Connected Layers. The temporal features extracted by two layers of LSTM and Dropout are mapped to the final dynamic capability boundary prediction values ​​through two fully connected layers.

[0112] 1. First fully connected layer (feature fusion). Input is the output h of the second Dropout layer. ′(2) (Dimension 32), number of neurons in the fully connected layer dfc1=16, using the ReLU activation function: ;W fc1 b is the weight matrix of the first fully connected layer, with dimensions (32, 16); fc1 The bias vector has dimensions (16,); ReLU(z) = max(0,z) is the ReLU activation function, introducing a nonlinear transformation. The output a1 has dimensions (batch_size, 16).

[0113] 2. Output Layer (Regression Prediction). The input is the output a1 of the first fully connected layer. The output layer has 2 neurons, corresponding to the normalized current tolerance boundary and reactive power support boundary prediction values, respectively. A linear activation function is used: Where: Wout is the output layer weight matrix with dimension (16,2); bout is the bias vector with dimension (2,); the output y has dimension (batch_size,2), where y[:,0] is the predicted current withstand boundary value, and y[:,1] is the predicted reactive power support boundary value. Finally, the output layer results are denormalized (using the min-max parameters saved during training) to obtain the actual physical dynamic capability boundary values.

[0114] S204. Using mean squared error as the loss function, the long short-term memory network model is iteratively trained using the time backpropagation algorithm until the loss function converges.

[0115] In some embodiments, the training hyperparameter settings are as follows: batch size: 32; initial learning rate: 0.001, using cosine annealing to decay the learning rate every 10 epochs, minimum learning rate = 1e-6; optimizer: Adam optimizer, β1 = 0.9, β2 = 0.999, ε = 1e-8; maximum number of iterations (epochs): 100; early stopping mechanism: when the validation set loss no longer decreases for 10 consecutive epochs, training is terminated early to save the optimal model parameters.

[0116] For example, loss function calculation: using mean squared error (MSE) as the loss function, the formula is: ; Where N is the batch size. , To predict and accurately measure the actual peak current, , For predicted and actual reactive power output values.

[0117] For example, the training process is as follows: each epoch iterates through the training set once, calculates the training loss and updates the model parameters; after each epoch, the validation loss is calculated on the validation set and the optimal model is recorded; the gradient norm is monitored in real time during training, and gradient clipping is performed when the gradient norm exceeds 5 to prevent gradient explosion.

[0118] S205. Validate the trained Long Short-Term Memory network model to ensure that the prediction error is below the preset threshold, and then obtain the dynamic capability prediction model.

[0119] For example, the validation metrics used are root mean square error (RMSE) and mean absolute percentage error (MAPE): ; For example, the verification criteria are: on an independent test set, the MAPE of the current withstand boundary prediction is ≤5%, and the RMSE is ≤10A; the MAPE of the reactive power support boundary prediction is ≤8%, and the RMSE is ≤5kVar; under extreme conditions (drop depth >70%, power >90% of rated power), the prediction error does not exceed 1.5 times the above criteria.

[0120] Thus, this invention constructs a sample set by collecting fault ride-through data under full boundary conditions, and trains it using a 2-layer stacked LSTM network combined with early stop and dropout mechanisms. Validated by an independent test set, the current prediction MAPE is ≤5% and the reactive power prediction MAPE is ≤8%, which can accurately predict the dynamic capability boundary of the system and improve the robustness of low voltage ride-through.

[0121] Optionally, the adaptive virtual impedance low voltage ride-through method for hydrogen fuel cell systems provided in this embodiment of the invention further includes steps S301-S306.

[0122] S301: Collect step response data of the converter under different virtual impedance values, different grid connection point voltages, and different output current conditions.

[0123] In some embodiments, the test conditions cover: collecting step response data of the converter under the following full operating conditions: virtual resistance: 0~1.0Ω, step size 0.1Ω; virtual inductance: 0~10mH, step size 1mH; grid connection point voltage: 0.2~1.0 times the rated voltage, step size 0.1 times; output current: 0.2~1.2 times the rated current, step size 0.2 times; DC bus voltage: 400~600V (typical DC bus voltage range for hydrogen fuel cell systems), step size 50V.

[0124] S302. Based on the step response data of the converter under different virtual impedance values, different grid connection point voltages, and different output current conditions, generate an input-output sample set.

[0125] The inputs of the input-output sample set include virtual resistance value, virtual inductance value, grid connection point voltage and output current, and the outputs of the input-output sample set include fault current peak value and reactive power output value.

[0126] In some embodiments, the input features for each sample include the virtual resistance value Rv, the virtual inductance value Lv, and the dq-axis component u of the grid connection point voltage for the current period. d u q The output current of the converter along the dq axis i d i q DC bus voltage U dc There are a total of 7 features. Output label: The output of each sample is the dq-axis component of the converter output current in the next control cycle (10ms later). dnext i qnext The peak fault current and reactive power output can be calculated from this result. Dataset partitioning: The dataset is randomly divided into a training set, a validation set, and an independent test set in a 7:2:1 ratio, with a total sample size of ≥50,000.

[0127] S303. The discrete state-space model of the converter is used as the basic form of the prediction model. The state variables include the output current and the virtual impedance state, and the input variables include the virtual impedance regulation and the grid voltage disturbance.

[0128] In some embodiments, a continuous state-space model of the converter is used as the basic form of the prediction model, and the standard form is: Where: state variable x = [i d i q U dc ] T The dimension is 3×1, representing the output current dq-axis components and the DC bus voltage, respectively; the input variable u=[ΔRv,ΔLv,u d ,u q]T, with dimensions of 4×1, represents the virtual resistance adjustment, virtual inductance adjustment, and the dq-axis component of the grid voltage, respectively; the output variable y=[i d i q ]T, with a dimension of 2×1, represents the dq-axis component of the converter output current; A, B, C, and D are the system matrix, input matrix, output matrix, and direct transfer matrix to be identified, with dimensions of 3×3, 3×4, 2×3, and 2×4, respectively.

[0129] S304. Using the least squares identification algorithm, the system matrix in the discrete state-space model of the converter is estimated by using the input and output sample sets to obtain a linearized prediction model.

[0130] In some embodiments, the model discretization preprocessing involves discretizing the continuous state-space model using the zero-order preservation method, with a sampling time Ts = 10 ms, to obtain the discrete form: ; Parameter identification: Batch least squares algorithm is used to analyze discrete matrix A using the input and output data of the training set. d B d C d D d Perform parameter estimation to minimize the mean square error between the predicted and actual outputs.

[0131] Continuous model reconstruction: The continuous system matrices A, B, C, and D are obtained through the inverse transformation of the discrete matrix, which are then used for flexible discretization of different control cycles.

[0132] S305. Discretize the linearized prediction model to obtain a prediction model suitable for discrete control cycles.

[0133] The inputs to the prediction model are the virtual impedance value, grid connection point voltage, and output current for the current cycle, and the outputs are the peak fault current and reactive power output for the next cycle.

[0134] In some embodiments, for the 10ms control cycle of this system, the continuous state-space model is discretized again using the zero-order hold method to obtain the final converter output current prediction model. The input of this model is the seven characteristic quantities of the current cycle, and the output is the dq-axis component of the current in the next cycle. The single-step prediction time is ≤0.1ms.

[0135] S306. Verify the prediction accuracy of the prediction model to ensure that the root mean square error between the output of the prediction model and the actual response is lower than the preset threshold. After the verification is passed, the converter output current prediction model is obtained.

[0136] In some embodiments, an independent test set is used to validate the prediction model. The validation criteria are: root mean square error (RMSE) of the predicted dq-axis components of the current ≤ 2A; peak fault current RMSE calculated from the predicted current ≤ 2.5A; reactive power output RMSE ≤ 3kVar; under extreme conditions (voltage drop depth > 70%, output current > 1.0 times rated current), the prediction error does not exceed 1.5 times the above criteria. After successful validation, the model parameters are embedded into the converter controller for candidate impedance response prediction.

[0137] Thus, this invention can construct a high-precision discrete prediction model for converters through multi-condition step response data acquisition and least squares identification, effectively characterizing the dynamic relationship between virtual impedance, grid voltage and output current, controlling the prediction error within the allowable range, providing reliable support for rapid evaluation of candidate impedances, and significantly improving optimization solution efficiency and control real-time performance.

[0138] Optionally, the adaptive virtual impedance low voltage ride-through method for hydrogen fuel cell systems provided in this embodiment of the invention further includes steps S401-S406.

[0139] S401: Real-time monitoring of grid connection point voltage. When the grid connection point voltage recovers to more than 90% of the rated value and the duration exceeds the preset time window, the low voltage ride-through mode is triggered to exit.

[0140] In some embodiments, the voltage determination criterion is that the effective values ​​of all three-phase voltages at the grid connection point must recover to more than 90% of the rated voltage. Stabilization time requirement: After voltage recovery, it must remain stable for 200ms (10 power frequency cycles) to avoid erroneous shutdown caused by instantaneous fluctuations in grid voltage. Three-phase imbalance determination: The three-phase voltage imbalance must be ≤5%; otherwise, the shutdown condition is not met.

[0141] S402. The virtual impedance value of the current cycle is gradually reduced according to the preset attenuation step size. The virtual resistance value and virtual inductance value are reduced by one step size in each control cycle until they both drop to zero.

[0142] In some embodiments, the attenuation step size is calculated as follows: the total attenuation time is set to 500ms, and the attenuation step size is: The decay occurs once per control cycle (10ms), and a total of 50 cycles are required to complete the decay. Attenuation process constraints: The Rv / Lv ratio remains constant during decay to ensure a constant impedance angle for the virtual impedance and avoid sudden changes in reactive power.

[0143] S403. During the attenuation process, if the grid connection point voltage is detected to drop below the preset threshold again, the attenuation will be stopped immediately and the low voltage ride-through mode will be re-entered.

[0144] In some embodiments, the response time is as follows: During the attenuation process, if the grid connection point voltage drops below 90% of the rated voltage again, the attenuation is immediately stopped within one control cycle (10ms), and the current virtual impedance value is locked. Reentry logic: After locking, the low-voltage ride-through control process is immediately re-executed without re-initialization, achieving seamless switching.

[0145] S404. When the virtual impedance value drops to zero, switch the converter's control mode from virtual impedance compensation state back to conventional grid-connected control state.

[0146] In some embodiments, the switching conditions are: when both the virtual resistance and virtual inductance decay to 0, and the grid connection point voltage stabilizes for 100ms. The switching process employs a seamless switching technique to smoothly transition the inner current loop reference command from virtual impedance compensation mode to conventional grid connection mode, with the current surge during switching ≤5% of the rated current. After switching, all fault status flags are reset, the historical data cache of the dynamic capability prediction model is cleared, and preparation is made for the next fault response.

[0147] Thus, this invention establishes a four-in-one low-voltage ride-through exit mechanism integrating judgment, attenuation, self-locking, and switching. It avoids false exit through strict voltage stabilization time judgment, achieves shock-free exit through linear equal impedance angle attenuation, and achieves rapid protection against secondary faults through 10ms-level fault recurrence response. Ultimately, it achieves a seamless and smooth transition between fault state and normal grid-connected state, fully meeting the highest requirements of the power grid for the continuity of distributed power generation grid-connected operation.

[0148] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0149] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 500 includes: a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501. When the processor 501 executes the computer program 503, it implements the steps in the above-described method embodiments. Alternatively, when the processor 501 executes the computer program 503, it implements the functions of each module / unit in the above-described device embodiments.

[0150] For example, the computer program 503 may be divided into one or more modules / units, which are stored in the memory 502 and executed by the processor 501 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 503 in the electronic device 500.

[0151] The processor 501 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0152] The memory 502 can be an internal storage unit of the electronic device 500, such as a hard disk or memory of the electronic device 500. The memory 502 can also be an external storage device of the electronic device 500, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 500. Furthermore, the memory 502 can include both internal and external storage units of the electronic device 500. The memory 502 is used to store the computer program and other programs and data required by the terminal. The memory 502 can also be used to temporarily store data that has been output or will be output.

[0153] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An adaptive virtual impedance low-voltage ride-through method suitable for hydrogen fuel cell systems, characterized in that, include: Real-time monitoring of the grid connection point voltage and key status parameters of the hydrogen fuel cell system; When the grid connection point voltage drops to a preset threshold, the low voltage ride-through mode is triggered. Based on the key state parameters of the system, the dynamic capability boundary of the hydrogen fuel cell system in the future time window is predicted through the dynamic capability prediction model. The dynamic capability boundary includes the current tolerance boundary and the reactive power support boundary. Based on the dynamic capability boundary and the voltage drop characteristics of the power grid, a multi-objective optimization model with fault suppression and reactive power support as objectives is established, and the optimal virtual impedance value and power command change trend are obtained by solving the model. Based on the optimal virtual impedance value and the power command change trend, dynamic adjustments are made in conjunction with the controllers of the converter and the hydrogen fuel cell system to achieve adaptive matching of the virtual impedance.

2. The adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems according to claim 1, characterized in that, The key state parameters of the main body include the stack output power, stack average temperature, anode hydrogen pressure change rate, proton exchange membrane high-frequency impedance, and air compressor speed of the hydrogen fuel cell system. The prediction of the dynamic capability boundary of the hydrogen fuel cell system within a future time window based on the key state parameters of the system and using a dynamic capability prediction model includes: Based on the key state parameters of the ontology, feature extraction is performed to generate a state time sequence. The state time sequence is input into the pre-trained dynamic capability prediction model, and through forward computation, the predicted values ​​of current withstand boundary and reactive power support boundary within the future time window are output. The dynamic capability boundary is determined based on the predicted values ​​of current withstand boundary and reactive power support boundary within the future time window.

3. The adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems according to claim 1, characterized in that, Before predicting the dynamic capability boundary of the hydrogen fuel cell system within a future time window using a dynamic capability prediction model based on the key state parameters, the process also includes: The system collects operational data of the hydrogen fuel cell system during the entire fault ride-through process under various operating conditions, including fault ride-through scenarios with different output power, different stack temperature, different anode hydrogen pressure change rate, and different voltage drop depth. The operating data under various working conditions are preprocessed to construct a training sample set. The input of each sample is a state time series sequence of N consecutive historical moments, and the output is the actual peak current and reactive power output capacity value in the future time window corresponding to each historical moment. Construct a long short-term memory network model, which includes an input layer, one or more LSTM hidden layers, a fully connected layer, and an output layer; Using mean squared error as the loss function, the long short-term memory network model is iteratively trained using the time backpropagation algorithm until the loss function converges; The trained Long Short-Term Memory (LSTM) network model was validated to ensure that the prediction error was below a preset threshold, thus obtaining the dynamic capability prediction model.

4. The adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems according to claim 1, characterized in that, The voltage sag characteristics of the power grid include sag depth and sag rate; Based on the dynamic capability boundary and the voltage sag characteristics of the power grid, a multi-objective optimization model is established with fault suppression and reactive power support as objectives. The optimal virtual impedance value and power command variation trend are obtained by solving the model, including: Based on the voltage drop characteristics of the power grid, calculate the reference value of reactive power support required by the power grid; A multi-objective optimization model is constructed with the dual objectives of minimizing fault current suppression deviation and minimizing the deviation between actual reactive power output and the reactive power support reference value, and with the current tolerance boundary and reactive power support boundary in the dynamic capability boundary as constraints. Based on voltage drop characteristics, drop depth and drop rate are used to determine the fault severity index; Based on the fault severity index and the dynamic capability boundary, the first weighting coefficient and the second weighting coefficient are dynamically adjusted; the first weighting coefficient is the weighting coefficient of the fault current suppression deviation, and the second weighting coefficient is the weighting coefficient of the deviation between the actual reactive power output and the reactive power support reference value; the first weighting coefficient and the second weighting coefficient are used to dynamically balance the priorities of fault current suppression and reactive power support. Based on the dynamically adjusted first and second weight coefficients and the multi-objective optimization model, an adaptive adjustment is performed to obtain a multi-objective optimization model with adaptive weights. The model predictive control algorithm is used to solve the multi-objective optimization model with adaptive weights to obtain the optimal virtual impedance value and the power command change trend.

5. The adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems according to claim 4, characterized in that, The method employs a model predictive control algorithm to solve the adjusted multi-objective optimization model, obtaining the optimal virtual impedance value and the power command change trend, including: In each control cycle, the virtual impedance value of the previous control cycle is used as the reference value. Within a preset range, the virtual resistance value and virtual inductance value in the virtual impedance value are discretized to obtain multiple sets of candidate impedance values. If the current control cycle is the first control cycle after triggering the low voltage ride-through mode, the reference value is zero. Based on the grid connection point voltage, output current, and multiple candidate impedance values ​​of the current control cycle, as well as the pre-established converter output current prediction model, the peak fault current and reactive power output value corresponding to each set of candidate impedance values ​​are predicted. Based on the objective function of the multi-objective optimization model with adaptive weights, and the peak fault current and reactive power output corresponding to each group of candidate impedance values, the comprehensive cost corresponding to each group of candidate impedance values ​​is calculated. The set of candidate impedance values ​​with the lowest overall cost is taken as the optimal virtual impedance value for the current control cycle. Calculate the difference between the optimal virtual impedance value and the reference value, and determine the adjustment direction and magnitude of active and reactive power based on the direction and magnitude of the difference, as a trend of power command change.

6. The adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems according to claim 5, characterized in that, The method further includes: If the fault severity index of the current control cycle is lower than the preset lower limit, then the preset range is determined as the benchmark value; If the fault severity index of the current control cycle is higher than or equal to the preset lower limit, then the scaling factor is determined according to the fault severity index of the current control cycle, wherein the larger the fault severity index, the larger the scaling factor. Multiply the upper and lower limits of the basic preset range by the scaling factor to obtain the dynamically scaled preset range.

7. The adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems according to claim 5, characterized in that, Before establishing a multi-objective optimization model based on dynamic capability boundaries and grid voltage sag characteristics, with fault suppression and reactive power support as objectives, and solving for the optimal virtual impedance value and power command change trend, the following steps are also included: Collect step response data of the converter under different virtual impedance values, different grid connection point voltages, and different output current conditions; Based on the step response data of the converter under different virtual impedance values, different grid connection point voltages, and different output currents, an input-output sample set is generated. The input of the input-output sample set includes virtual resistance value, virtual inductance value, grid connection point voltage, and output current. The output of the input-output sample set includes fault current peak value and reactive power output value. The discrete state-space model of the converter is used as the basic form of the prediction model. The state variables include the output current and the virtual impedance state, and the input variables include the virtual impedance regulation and the grid voltage disturbance. The least squares identification algorithm is used to estimate the parameters of the system matrix in the discrete state-space model of the converter using the input and output sample set, so as to obtain a linearized prediction model. The linearized prediction model is discretized to obtain a prediction model suitable for discrete control cycles. The input of the prediction model is the virtual impedance value, grid connection point voltage and output current of the current cycle, and the output is the fault current peak value and reactive power output value of the next cycle. The prediction accuracy of the prediction model is verified to ensure that the root mean square error between the output of the prediction model and the actual response is lower than a preset threshold. After the verification is passed, the converter output current prediction model is obtained.

8. The adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems according to claim 1, characterized in that, The process of dynamically adjusting the virtual impedance based on the optimal virtual impedance value and the power command change trend, in conjunction with the controllers of the converter and the hydrogen fuel cell system, to achieve adaptive matching of the virtual impedance includes: The virtual resistance and virtual inductance values ​​in the optimal virtual impedance value are embedded into the voltage outer loop control circuit of the converter, and a compensation voltage signal is generated based on the virtual resistance value, virtual inductance value and the current output current of the converter. The compensation voltage signal is superimposed on the original current reference command generated by the outer voltage loop control loop to generate the corrected inner current loop reference command. Based on the corrected inner current loop reference command, the duty cycle of the converter switching transistors is controlled to achieve real-time adjustment of the output current. The direction and magnitude of active power adjustment and reactive power adjustment in the power command change trend are synchronously fed forward to the auxiliary system controller of the hydrogen fuel cell system. The auxiliary system controller performs linkage adjustment based on the feedforward signal: pre-adjusting the opening degree and opening change rate of the hydrogen injection valve to match the change rate of active power, pre-increasing the speed of the air compressor to ensure the supply of oxygen for the electrochemical reaction required for dynamic adjustment of active power, and pre-adjusting the flow rate of the cooling water pump to suppress the temperature fluctuation of the fuel cell stack. During low voltage ride-through, the stack temperature and anode hydrogen pressure changes are continuously monitored. If the temperature exceeds the limit or the hydrogen pressure fluctuation exceeds the safe range, a correction signal is generated and fed back to the auxiliary system controller to perform closed-loop correction of the linkage adjustment amount until the low voltage ride-through mode ends.

9. The adaptive virtual impedance low-voltage ride-through method for hydrogen fuel cell systems according to claim 1, characterized in that, The method further includes: Real-time monitoring of grid connection point voltage; when the grid connection point voltage recovers to more than 90% of the rated value and the duration exceeds the preset time window, triggering the exit from low voltage ride-through mode; The virtual impedance value of the current cycle is gradually reduced according to the preset decay step size, and the virtual resistance value is reduced by one step size in each control cycle until the virtual resistance value and the virtual inductance value are reduced to zero. During the attenuation process, if the grid connection point voltage is detected to drop below the preset threshold again, the attenuation will be stopped immediately and the low voltage ride-through mode will be re-entered. Once the virtual impedance value drops to zero, the converter's control mode will be switched from virtual impedance compensation state back to conventional grid-connected control state.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor being configured to invoke and run the computer program stored in the memory to perform the method as described in any one of claims 1 to 9.

Images