A multi-port electrical-hydrogen coupling energy router control method, medium and system

Abstract

The application provides a multi-port electric-hydrogen coupling energy router control method, medium and system, belongs to the technical field of multi-port electric-hydrogen coupling energy router control, and is based on an enhanced electrochemical model to correct hydrogen production efficiency of an alkaline electrolytic cell online, dynamically issues a virtual impedance parameter to each port output impedance according to a bus voltage stability evaluation value, inputs the corrected hydrogen production efficiency and a power prediction value into a rolling horizon multi-working condition switching optimization algorithm based on Lyapunov stability constraint to output a power instruction and a working condition switching instruction, issues a running mode instruction and a power instruction to each port converter according to a corresponding control strategy of the working condition, and the converter level executes current inner loop control, PQ calculation and PWM phase shift control, so that the technical problem that the multi-port electric-hydrogen coupling energy router is caused to be unstable due to multi-port power voltage strong coupling oscillation in the multi-working condition switching process is solved.

Description

Technical Field

[0001] This invention belongs to the field of multi-port electro-hydrogen coupling energy router control technology, specifically, it relates to a multi-port electro-hydrogen coupling energy router control method, medium and system. Background Technology

[0002] Multi-port electro-hydrogen coupled energy routers are core equipment for achieving coordinated regulation of hydrogen sources, grid, load, and storage. They collect energy from multiple ports, including photovoltaic, energy storage, alkaline electrolyzers, proton exchange membrane electrolyzers, fuel cells, and AC grids, via a DC bus, and route and distribute power according to operating conditions. In current multi-port energy router systems, the converters at each port operate in parallel on the same DC bus. System-level control typically employs a hierarchical control architecture with fixed parameters, coordinating the operation of each port using preset power commands and operating condition switching rules. Operating condition switching decisions rely on threshold logic tuned based on engineering experience, and power prediction often uses a single time-series model or simple regression methods.

[0003] However, due to the nonlinear impedance characteristics of electrochemical loads such as alkaline electrolyzers and proton exchange membrane electrolyzers, and the negative impedance superposition effect of converter output impedance in a wide frequency band, the DC bus is prone to coupled oscillations when multiple ports are connected in parallel, which can not be dynamically adjusted according to the oscillation risk of each port. This makes it difficult to suppress coupled oscillations and reduces system stability.

[0004] In other words, existing technologies have a technical problem where multi-port electro-hydrogen coupled energy routers experience system instability due to strong coupling oscillations of power and voltage across multiple ports during multi-condition switching. Summary of the Invention

[0005] In view of this, the present invention provides a control method, medium and system for a multi-port electro-hydrogen coupling energy router, which can solve the technical problem in the prior art of system instability caused by strong coupling oscillation of multi-port power voltage during multi-condition switching of multi-port electro-hydrogen coupling energy routers.

[0006] The present invention is implemented as follows: The first aspect of the present invention provides a control method for a multi-port electro-hydrogen coupling energy router, comprising the following steps: The SCADA database of the multi-port electro-hydrogen coupled energy router collects the voltage, current, power, state of charge of the energy storage battery, alkaline concentration, membrane water content, and DC 750V bus voltage of each port. The collected data are input into the cross-modal spatiotemporal graph attention model, and the predicted values ​​of future power demand for each port and the predicted values ​​of DC 750V bus voltage fluctuation are output. The collected port currents and real-time cell voltage feedback values ​​of the alkaline electrolyzer are input into the enhanced electrochemical model to solve the dynamic equation of bubble coverage, output the bubble coverage correction value, and perform online correction on the estimated value of hydrogen production efficiency of the alkaline electrolyzer based on the bubble coverage correction value, and output the corrected estimated value of hydrogen production efficiency. The bus voltage stability assessment value is calculated based on the predicted value of the DC 750V bus voltage fluctuation amplitude. According to the range to which the bus voltage stability assessment value belongs, the corresponding virtual impedance parameters are sent to the converter-level controller to shape the output impedance of each port and suppress multi-port coupled oscillation. The predicted future power demand of each port, the corrected hydrogen production efficiency estimate, and the current operating condition identifier are input into the rolling time-domain multi-operating condition switching optimization algorithm based on Lyapunov stability constraints. Within each control step, the power command and operating condition switching command of each port are output. Based on the operating condition switching command, determine the current operating condition, and issue operating mode commands and power commands to each port converter according to the control strategy corresponding to the operating condition. The converter-level controller receives operating mode commands and power commands, and performs current inner loop control, PQ calculation and PWM phase shift control to maintain the stability of the DC 750V bus voltage and the power balance of each port.

[0007] Specifically, the cross-modal spatiotemporal graph attention model abstracts the system topology of the multi-port electro-hydrogen coupled energy router into a graph structure, with 8 ports serving as graph nodes, power flow between ports as directed edges, and the weight of the directed edges dynamically assigned by the real-time transmission power. The network adopts a three-layer stacked structure: a bottom-layer multi-head graph attention layer, a middle-layer bidirectional gated cyclic unit time encoder, and a top-layer cross-modal fusion layer.

[0008] Specifically, the top-level cross-modal fusion layer encodes the electrical quantity branch and the chemical quantity branch separately through fully connected encoding, and then merges them into a unified graph node feature vector through attention weighting. The electrical quantity branch includes the voltage and current of each port, and the chemical quantity branch includes the alkali concentration, membrane water content and energy storage battery state of charge.

[0009] Specifically, the training of the cross-modal spatiotemporal graph attention model involves adding Kirchhoff's current law residual as a physical constraint loss term to the loss function. The total loss function is composed of a weighted sum of the power prediction mean square error term, the DC 750V bus voltage prediction mean square error term, and the physical constraint loss term. The three weights are determined through grid search experiments on the validation set.

[0010] Specifically, the training dataset for the cross-modal spatiotemporal graph attention model is collected from historical operating data in the SCADA database. The historical data is used as the input window, and the predicted power demand of each port and the predicted voltage fluctuation of the DC 750V bus are used as labels to construct time-series sample pairs. The dataset is divided according to the proportion of the training set, the proportion of the validation set, and the proportion of the test set.

[0011] Specifically, the parameters of the dynamic equation for bubble coverage in the enhanced electrochemical model are obtained by performing bubble generation and desorption experiments within a preset range of current density coverage in the alkaline electrolyzer, and by fitting and analyzing the nonlinear response data of the real-time cell voltage feedback value and the current at each port of the alkaline electrolyzer.

[0012] The bus voltage stability assessment value is based on the partition threshold. and The system is divided into a stable zone, a transition zone, and an oscillation risk zone. When the bus voltage stability assessment value does not exceed... Currently, small virtual impedance parameters are issued. When the bus voltage stability assessment value is located at... and When the intermediate virtual impedance parameter is issued, the bus voltage stability assessment value exceeds Currently, large virtual impedance parameters are being issued.

[0013] Wherein, the partition threshold and Specifically, this is achieved by conducting at least a preset number of experiments under different load disturbance amplitudes and different port switching combinations, recording the time-domain response data of the DC 750V bus voltage deviation and the power deviation of each port, and then determining the stability domain boundary through statistical analysis.

[0014] Specifically, the rolling time-domain multi-condition switching optimization algorithm based on Lyapunov stability constraints models the multi-condition switching problem of a multi-port electro-hydrogen coupled energy router as a mixed integer optimal control problem with switching constraints. The binary variable of the condition switching is used as an integer decision, and the power command of each port is used as a continuous decision, so as to jointly optimize the system operating efficiency.

[0015] The rolling time-domain multi-condition switching optimization algorithm based on Lyapunov stability constraints further includes constructing a Lyapunov function with the DC 750V bus voltage deviation and the power deviation of each port as state components. It requires that the Lyapunov function value strictly decreases monotonically after each control step, and this condition is embedded into the optimization problem as a hard constraint.

[0016] Specifically, the solution of the rolling time-domain multi-condition switching optimization algorithm based on Lyapunov stability constraints involves solving the integer variables using the branch and bound method, applying Lagrange relaxation to the Lyapunov stability constraints, updating the dual variables using the subgradient method, ensuring that the number of inner iterations does not exceed the upper limit of the iteration limit, and ensuring that the total computation time does not exceed the upper limit of the online computation time.

[0017] The current operating conditions include fuel cell 10kV grid-connected power generation, new energy off-grid hydrogen production, and new energy weak grid-connected hydrogen production. Under the fuel cell 10kV grid-connected power generation condition, the AC 10kV grid port operates in V / F mode to support the DC 750V bus voltage, the energy storage port operates in P / Q mode to maintain power balance, and the fuel cell port connects to the DC 750V bus for power generation in P / Q mode.

[0018] In the off-grid hydrogen production process, the energy storage port operates in V / F mode to support a DC 750V bus voltage, the photovoltaic port operates in maximum power point tracking mode, and the alkaline electrolyzer port and proton exchange membrane electrolyzer port operate in P / Q mode. The control system adjusts the electrolyzer operating power according to the photovoltaic power output and the state of charge of the energy storage battery.

[0019] In the case of weak grid connection hydrogen production of new energy, the interaction power between the multi-port electro-hydrogen coupling energy router and the AC 10kV grid is set not to exceed the weak grid connection interaction power ratio of the hydrogen production load power. The AC 10kV grid port operates in V / F mode, the energy storage port maintains power balance in P / Q mode, and the photovoltaic port operates in maximum power point tracking mode.

[0020] The optimal values ​​for the training set ratio, validation set ratio, and test set ratio are 7, 2, and 1, respectively; the optimal input window duration is 60 minutes; the optimal prediction time domain for the predicted future power demand of each port is 15 minutes; the optimal upper limit for the number of inner layer iterations is 20; the optimal upper limit for the online calculation time of the entire process is 30 seconds; the optimal value for the weak grid-connected interactive power ratio is 10%; and the optimal preset range for the alkaline electrolytic cell current density is 0.1. Up to 2.0 .

[0021] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, perform the aforementioned multi-port electro-hydrogen coupling energy router control method.

[0022] A third aspect of the present invention provides a multi-port electro-hydrogen coupling energy router control system, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0023] This invention predicts the fluctuation amplitude of DC 750V bus voltage using a cross-modal spatiotemporal graph attention model, calculates the bus voltage stability assessment value based on the prediction results, and dynamically sends the corresponding virtual impedance parameters to the converter-level controller according to the interval to which the assessment value belongs, so as to shape the output impedance of each port in real time, so that the impedance characteristics of the multi-port parallel converter meet the stability criteria, thereby actively suppressing the coupling oscillation caused by the superposition of negative impedance during the switching of operating conditions.

[0024] Traditional fixed-parameter controllers cannot detect dynamic changes in oscillation risk. This invention, however, utilizes a closed-loop mechanism of prediction-evaluation-parameter adjustment to complete impedance shaping before oscillation occurs, fundamentally eliminating the cause of oscillation due to impedance mismatch. Simultaneously, the rolling time-domain optimization algorithm based on Lyapunov stability constraints embeds system stability conditions as hard constraints, ensuring that operating condition switching decisions do not mathematically compromise system stability. The synergistic effect of these two mechanisms systematically improves oscillation suppression capabilities.

[0025] In summary, this invention solves the technical problem mentioned in the background art of system instability caused by strong coupling oscillation of power voltage across multiple ports during multi-condition switching in multi-port electro-hydrogen coupled energy routers. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 Topology diagram of a multi-port electro-hydrogen coupling energy router system.

[0028] Figure 3 The multi-port electro-hydrogen coupling energy router management system adopts a hierarchical partitioning architecture.

[0029] Figure 4 Topology and control strategy of fuel cell 10kV grid-connected power generation system.

[0030] Figure 5 Topology and control strategy of off-grid hydrogen production system for new energy.

[0031] Figure 6 System topology and control strategy for hydrogen production under weak grid connection conditions in new energy sources. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0033] like Figure 1 The diagram shown is a flowchart of a multi-port electro-hydrogen coupling energy router control method provided by the first aspect of the present invention. This method includes the following steps: S01. Collect the voltage, current, power, state of charge of the energy storage battery, alkaline concentration, membrane water content, and DC 750V bus voltage of each port from the SCADA database of the multi-port electro-hydrogen coupling energy router. Input the voltage, current, power, state of charge of the energy storage battery, alkaline concentration, membrane water content, and DC 750V bus voltage of each port into the cross-modal spatiotemporal graph attention model, and output the predicted power demand of each port for the next 15 minutes and the predicted fluctuation range of the DC 750V bus voltage. S02. Input the current at each port collected in S01 and the real-time cell voltage feedback value of the alkaline electrolyzer into the enhanced electrochemical model, solve the dynamic equation of bubble coverage, output the bubble coverage correction value, and correct the estimated value of hydrogen production efficiency of the alkaline electrolyzer online based on the bubble coverage correction value, and output the corrected estimated value of hydrogen production efficiency. S03. Based on the predicted value of DC 750V bus voltage fluctuation amplitude output by S01, calculate the bus voltage stability assessment value. According to the range to which the bus voltage stability assessment value belongs, send the corresponding virtual impedance parameters to the converter-level controller to shape the output impedance of each port and suppress multi-port coupled oscillation. S04. Input the predicted power demand values ​​for each port in the next 15 minutes from the output of S01, the corrected hydrogen production efficiency estimate from the output of S02, and the current operating condition identifier into the rolling time-domain multi-operating condition switching optimization algorithm based on Lyapunov stability constraints, and output the power command and operating condition switching command for each port in each control step. S05. Based on the operating condition switching command output by S04, determine whether the current operating condition is the 10kV grid-connected power generation of fuel cell, the off-grid hydrogen production of new energy, or the weak grid-connected hydrogen production of new energy, and issue the operating mode command and power command to each port converter according to the control strategy corresponding to the operating condition. S06, the converter-level controller receives the operating mode command and power command issued by S05, executes current inner loop control, PQ calculation and PWM phase shift control, and maintains the stability of the DC 750V bus voltage and the power balance of each port.

[0034] The specific structure of the cross-modal spatiotemporal graph attention model is as follows: the system topology of the multi-port electro-hydrogen coupled energy router is abstracted into a graph structure. Eight ports—photovoltaic port, energy storage port, alkaline electrolyzer port, proton exchange membrane electrolyzer port, fuel cell port, 10kV AC grid port, DC load port, and AC load port—are each used as graph nodes. The power flow between ports is represented as directed edges, and the weights of these directed edges are dynamically assigned based on the real-time transmitted power. The network adopts a three-layer stacked structure: the bottom layer is a multi-head graph attention layer, which assigns attention weights to the neighboring graph nodes of each graph node in the topology graph and calculates the spatial association weighted embedding vector of each graph node. The multi-head mechanism allows the model to simultaneously focus on the spatial association relationships of different subspaces. The middle layer is a bidirectional gated recurrent unit time-series encoder, which extracts the historical time-series features of each graph node bidirectionally (forward and backward) to capture the long-range dependence and short-term fluctuation features of the power sequence. The temporal encoded output of each graph node is concatenated with the weighted embedding vector associated with the lower spatial layer and then fed into the top layer. The top layer is a cross-modal fusion layer, which encodes the electrical quantity branches (voltage and current at each port) and the chemical quantity branches (alkali concentration, membrane water content, and state of charge of the energy storage battery) through fully connected layers, and then fuses them into a unified graph node feature vector through attention-weighted concatenation. The graph convolutional layer performs three iterations of neighborhood aggregation. After each aggregation, the original graph node features are preserved through residual skip connections to avoid excessive smoothing of graph node embedding. During the training phase, the Kirchhoff current law residual is added to the loss function as a physical constraint loss term. During backpropagation, prediction results that violate power conservation are penalized simultaneously, forcing the model output to meet the power conservation constraint. The output layer is divided into two task heads. The first task head outputs the predicted power demand of each port for the next 15 minutes, and the second task head outputs the predicted value of the DC 750V bus voltage fluctuation amplitude.

[0035] The cross-modal spatiotemporal graph attention model performs cross-modal joint modeling of electrical and chemical quantities of the multi-port electro-hydrogen coupled energy router in a unified graph structure. This enables the model to simultaneously perceive the impact of the electrochemical state of the electrolyzer on power demand. The physical constraint loss term ensures the consistency of the prediction results at the global power conservation level. The cascaded structure of the time encoder and the graph attention layer allows the model to extract high-quality features in both spatial topological association and temporal dynamics. This improves the power prediction accuracy in scenarios with strong multi-port coupling and heterogeneous physical quantities, provides reliable input for subsequent optimization decisions, and reduces control malfunctions and incorrect switching of operating conditions caused by prediction deviations.

[0036] The steps for establishing the training dataset for the cross-modal spatiotemporal graph attention model specifically include: collecting historical operating data from the SCADA database of the multi-port electro-hydrogen coupled energy router. The collected fields include the voltage, current, and power of each port, the state of charge of the energy storage battery, the alkaline concentration, the membrane water content, and the DC 750V bus voltage. The time resolution of the data collection is 1 minute, covering complete operating records under three operating conditions: 10kV grid-connected power generation of fuel cells, off-grid hydrogen production of new energy, and weak grid-connected hydrogen production of new energy. The collected data is processed by imputing missing values, removing outliers, and normalizing. Using the historical data of the past 60 minutes as the input window, and using the predicted power demand of each port and the predicted fluctuation amplitude of the DC 750V bus voltage for the next 15 minutes as labels, time-series sample pairs are constructed. The training set, validation set, and test set are divided in a 7:2:1 ratio.

[0037] The specific steps for training the cross-modal spatiotemporal graph attention model include: using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64; the total loss function is composed of a weighted sum of the power prediction mean square error, the DC 750V bus voltage prediction mean square error, and the Kirchhoff current law residual physical constraint term, the weights of which are determined through grid search experiments on the validation set; the training epochs are capped at 200 epochs, and an early stopping mechanism is introduced, terminating training when the validation set loss does not decrease for 20 consecutive epochs; after training, the prediction performance is evaluated on the test set, and if the average absolute error of the predicted power demand values ​​for each port on the test set exceeds a preset error threshold, the model structure or hyperparameters are readjusted and retrained; the preset error threshold is determined through cross-validation experiments on multiple sets of historical data.

[0038] The bus voltage stability assessment value is calculated based on the predicted value of DC 750V bus voltage fluctuation amplitude, and its partition threshold is... and ( The results were obtained by conducting no fewer than 50 sets of experiments under different load disturbance amplitudes and different port switching combinations, recording the time-domain response data of DC 750V bus voltage deviation and power deviation of each port, and then determining the stability domain boundary through statistical analysis.

[0039] Let the bus voltage stability assessment value be... The partitioning adjustment rule for the virtual impedance parameter is as follows: when When the system is in the stable region, it sends small virtual impedance parameters to the converter-level controller to reduce power transmission losses; when At this time, the system is in the transition region, and sends medium virtual impedance parameters to the converter-level controller to balance power transmission losses and suppress coupled oscillations; when At this time, the system is in the oscillation risk zone, and a large virtual impedance parameter is issued to the converter-level controller to prioritize the suppression of multi-port coupled oscillations. The virtual impedance parameters corresponding to each interval are determined through impedance characteristic experiments and oscillation suppression effect analysis under the corresponding operating conditions.

[0040] The parameters of the dynamic equation for bubble coverage in the enhanced electrochemical model are determined by covering the alkaline electrolyzer current density with a value of 0.1. Up to 2.0 Bubble generation-desorption experiments were conducted within the specified range, and the nonlinear response data of the real-time cell voltage feedback value and the current at each port of the alkaline electrolyzer were obtained through fitting analysis.

[0041] The principle and specific implementation of the rolling time-domain multi-condition switching optimization algorithm based on Lyapunov stability constraints are as follows: The multi-condition switching problem of a multi-port electro-hydrogen coupled energy router is modeled as a mixed integer optimal control problem with switching constraints, with a prediction time domain of 15 minutes and a control time domain of 1 minute; the binary variables of the condition switching are used as integer decisions, and the power commands of each port are used as continuous decisions, jointly optimizing the system operating efficiency; a Lyapunov function is constructed with the DC 750V bus voltage deviation and the power deviation of each port as state components, requiring that the Lyapunov function value strictly decreases monotonically after each control step. This condition is embedded as a hard constraint in the optimization problem, thus ensuring that the system does not become unstable during the switching of operating conditions at the solution level. Integer variables are solved using the branch-and-bound method, and Lagrange relaxation is applied to the Lyapunov stability constraints. Dual variables are updated using the subgradient method, with an upper limit of 20 iterations for the inner layer. The entire computation time is controlled within 30 seconds to meet online requirements. The specific form and weight coefficients of the Lyapunov function are determined through iterative analysis of simulation experiments and real-world disturbance experiments under three operating conditions. It is necessary to ensure that the deceleration rate of the Lyapunov function value meets the preset convergence speed requirement under typical disturbances in each operating condition. The rolling time-domain multi-condition switching optimization algorithm based on Lyapunov stability constraints transforms the system stability conditions from engineering experience into quantifiable optimization constraints, enabling the switching decision to find the best solution while ensuring system stability. This overcomes the problem that traditional fixed-parameter controllers cannot simultaneously handle multi-port power-voltage coupling oscillations in multi-port strongly coupled scenarios, improving the reliability of multi-condition switching and the overall system operating efficiency.

[0042] The principle of the virtual impedance decoupling control is as follows: a virtual impedance element is introduced at the output end of each converter control loop. The impedance characteristics are simulated through a control algorithm, and the output impedance frequency response of each port is reshaped in real time. This ensures that the impedance characteristics of multiple parallel converters meet the stability criterion, thereby suppressing the DC 750V bus coupling oscillation caused by the superposition of negative impedance characteristics and achieving decoupling of power control at each port. The virtual impedance decoupling control changes the port impedance characteristics only by adjusting the control algorithm parameters. In multi-port parallel operation scenarios, it features high flexibility and fast response speed, effectively solving the problem of difficult power control decoupling caused by multi-port dynamic coupling.

[0043] The state of charge (SBC) of the energy storage battery is the ratio of the battery's current remaining charge to its rated capacity, reflecting the degree of charging and discharging of the battery.

[0044] The SCADA database is a data acquisition and monitoring system database used to store real-time operating data of each port of the multi-port electro-hydrogen coupling energy router.

[0045] The subgradient method is an iterative method for solving non-smooth optimization problems. In each iteration, the dual variable is updated along the subgradient direction to gradually approach the optimal solution of the Lagrange dual problem. It is applicable to mixed integer optimization problems with non-differentiable constraints.

[0046] The multi-port electro-hydrogen coupling energy router consists of a 750V DC bus, 8 port interfaces, and a corresponding current transformer. The 8 ports include: 6 DC interfaces for alkaline electrolyzers, proton exchange membrane electrolyzers, fuel cells, photovoltaics, energy storage, and DC loads; and 2 AC interfaces for 10kV grid and AC loads. The 10kV grid interface uses a parallel input and series output power electronic transformer to achieve high-voltage access and electrical isolation. All ports achieve energy collection and power routing through the 750V DC bus. The control system adopts a hierarchical and partitioned architecture, divided into system-level and converter-level layers. System-level control includes panoramic monitoring, production planning decision-making, optimization control, photovoltaic absorption and utilization rate analysis, photovoltaic power generation prediction, energy consumption analysis, energy storage charging and discharging statistics, and operating condition control functions. Converter-level control includes frequency and voltage stabilization, current inner-loop control, PQ calculation, and PWM phase-shift control functions.

[0047] The specific implementation of step S01 is as follows: The system collects seven types of operational data in real time from the SCADA database of the multi-port electro-hydrogen coupled energy router, with a time resolution of 1 minute: voltage, current, power, state of charge of the energy storage battery, alkaline concentration, membrane water content, and DC 750V bus voltage. These collected data are then fed into a cross-modal spatiotemporal graph attention model. This model abstracts the system topology into a graph structure, with the eight ports serving as graph nodes, power flow between ports as directed edges, and the weights of these directed edges dynamically assigned by the real-time transmitted power. The bottom-layer multi-head graph attention layer assigns attention weights to the neighboring graph nodes of each graph node and calculates the spatial correlation weighted embedding vector. The middle-layer bidirectional gated recurrent unit temporal encoder extracts the historical temporal features of each graph node in both forward and reverse directions, capturing the long-range dependence and short-term fluctuation features of the power sequence. The temporal encoding output is concatenated with the bottom-layer embedding vector and then fed into the top layer. The top-layer cross-modal fusion layer encodes the electrical and chemical branches separately through fully connected layers and then fuses them through attention-weighted concatenation. The graph convolutional layer performs three iterations of neighborhood aggregation. After each aggregation, residual skip connections are used to retain the original graph node features to avoid over-smoothing. During the training phase, Kirchhoff's current law residual is added to the loss function as a physical constraint loss term to force the model output to meet the power conservation constraint. The model output layer is divided into two task heads: the first task head outputs the predicted power demand values ​​for each port in the next 15 minutes, and the second task head outputs the predicted values ​​of the DC 750V bus voltage fluctuation amplitude. The training dataset uses historical data from the past 60 minutes as the input window, covering the complete historical records of three operating conditions. The training set, validation set, and test set are divided in a 7:2:1 ratio. The Adam optimizer is used for training, and an early stopping mechanism is introduced. When the average absolute error of the power demand prediction values ​​of each port in the test set exceeds the preset error threshold, the model structure or hyperparameters are readjusted.

[0048] The specific implementation of step S02 is as follows: The current at each port collected in step S01 and the real-time cell voltage feedback value of the alkaline electrolyzer are input into the enhanced electrochemical model. The core of the enhanced electrochemical model is the bubble coverage dynamic equation, which describes the nonlinear dynamic process of the bubble coverage on the electrode surface of the alkaline electrolyzer changing with current density. The equation parameters are determined by the current density coverage of 0.1... Up to 2.0 Bubble generation and desorption experiments were conducted within the specified range, and the nonlinear response data of the cell voltage feedback value and the current at each port were fitted and analyzed. The model takes the real-time current at each port and the cell voltage feedback value as input, solves the bubble coverage correction value under the current operating conditions, and corrects the estimated hydrogen production efficiency of the alkaline electrolyzer online based on the correction value, outputting the corrected hydrogen production efficiency estimate for subsequent optimization algorithms.

[0049] The specific implementation of step S03 is as follows: Based on the predicted value of the DC 750V bus voltage fluctuation amplitude output in step S01, the bus voltage stability assessment value is calculated. The bus voltage stability assessment value is based on the zoning threshold. and ( The system is divided into three intervals. The partition thresholds are determined by statistically analyzing the time-domain response data of the DC 750V bus voltage deviation and the power deviation at each port under at least 50 sets of experiments with different load disturbance amplitudes and different port switching combinations. The bus voltage stability assessment value is set at no more than [a certain threshold value]. When the system is in the stable region, a small virtual impedance parameter is issued to reduce power transmission loss; when the bus voltage stability assessment value is within the range of... and During this period, the system is in the transition zone, and a moderate virtual impedance parameter is issued to balance losses and oscillation suppression; when the bus voltage stability assessment value exceeds... When the system is in the oscillation risk zone, a large virtual impedance parameter is issued to prioritize the suppression of multi-port coupled oscillations. Virtual impedance decoupling control introduces a virtual impedance element at the output of each converter control loop, uses a control algorithm to simulate impedance characteristics, and reshapes the frequency response of the output impedance of each port in real time, so that the impedance characteristics of multiple parallel converters meet the stability criteria.

[0050] The specific implementation of step S04 is as follows: The predicted power demand values ​​for each port in the next 15 minutes output from step S01, the corrected hydrogen production efficiency estimates output from step S02, and the current operating condition identifier are input into a rolling time-domain multi-condition switching optimization algorithm based on Lyapunov stability constraints. This algorithm models the multi-condition switching problem as a mixed integer optimal control problem with switching constraints. The prediction time domain is 15 minutes, the control time domain is 1 minute, and the binary variables for operating condition switching are integer decisions, while the power commands for each port are continuous decisions, jointly optimizing the system operating efficiency. The algorithm constructs a Lyapunov function with the DC 750V bus voltage deviation and the power deviations for each port as state components, requiring that the Lyapunov function value strictly decreases monotonically after each control step. This condition is embedded as a hard constraint in the optimization problem. Integer variables are solved using the branch-and-bound method, Lagrange relaxation is applied to the Lyapunov stability constraints, and dual variables are updated using the subgradient method. The upper limit of the inner iteration is 20 times, and the entire process calculation time is controlled within 30 seconds to meet online requirements. Within each control step, the algorithm outputs power commands and operating condition switching commands for each port.

[0051] The specific implementation of step S05 is as follows: Based on the operating condition switching command output in step S04, determine whether the current operating condition is a 10kV grid-connected fuel cell power generation condition, a new energy off-grid hydrogen production condition, or a new energy weak grid-connected hydrogen production condition. Then, issue operating mode commands and power commands to each port converter according to the control strategy corresponding to the operating condition. Under the 10kV grid-connected fuel cell power generation condition, the AC 10kV grid port operates in V / F mode to support the DC 750V bus voltage, the energy storage port operates in P / Q mode to maintain power balance, and the fuel cell port connects to the DC 750V bus for power generation in P / Q mode. Under the new energy off-grid hydrogen production condition, the energy storage port operates in V / F mode to support the DC 750V bus voltage, the photovoltaic port operates in maximum power point tracking mode, and the alkaline electrolyzer port and proton exchange membrane electrolyzer port operate in P / Q mode. The control system adjusts the electrolyzer operating power according to the photovoltaic power output and the state of charge of the energy storage battery. Under the condition of weak grid connection of new energy hydrogen production, the interaction power with the AC 10kV grid does not exceed 10% of the hydrogen production load power. The AC 10kV grid port operates in V / F mode, the energy storage port maintains power balance in P / Q mode, and the photovoltaic port operates in maximum power point tracking mode.

[0052] The specific implementation of step S06 is as follows: After receiving the operating mode command and power command issued in step S05, the converter-level controller executes current inner-loop control, PQ calculation, and PWM phase-shift control. Current inner-loop control ensures current tracking accuracy through closed-loop adjustment of the output current at each port; PQ calculation calculates the active and reactive power of each port in real time based on the operating mode command, providing a reference value for phase-shift control; PWM phase-shift control achieves precise power distribution among the converters by adjusting the phase relationship of the switching signals, ultimately maintaining stable DC 750V bus voltage and power balance at each port.

[0053] It should be noted that the key technologies of this invention include: a cross-modal spatiotemporal graph attention model that jointly models electrical and chemical quantities in a unified graph structure; a bottom-level graph attention layer that captures the spatial topological relationships between ports; a middle-level bidirectional gated cyclic unit time-series encoder that captures the temporal dynamics of the power sequence; and a top-level cross-modal fusion layer that merges electrical and chemical quantity branches into a unified feature, enabling the model to simultaneously perceive the impact of the electrolytic cell's electrochemical state on power demand. A physical constraint loss term ensures the consistency of the prediction results at the global power conservation level, thereby improving power prediction accuracy in scenarios with strong multi-port coupling and heterogeneous physical quantities. Compared to a single time-series model or simple regression methods, this technology can simultaneously model high-quality features in both spatial topology and time series dimensions. The virtual impedance partitioning adjustment technology, through a closed-loop mechanism of prediction-evaluation-parameter adjustment, completes the output impedance shaping of each port before oscillation occurs, ensuring that the impedance characteristics of the parallel converter meet the stability criterion, thus solving the fundamental defect that fixed-parameter controllers cannot perceive the dynamic changes in oscillation risk. The rolling time-domain optimization technique based on Lyapunov stability constraints transforms stability conditions from engineering experience into quantifiable hard constraints, ensuring that the switching decisions do not mathematically compromise system stability. The synergistic effect of these three key technologies enables power prediction to provide feedforward for impedance shaping, impedance shaping to provide oscillation suppression for switching, and switching optimization to seek the best outcome while ensuring stability, forming a complete stable control closed loop from perception to decision-making to execution.

[0054] It should be noted that this invention also solves the following technical problem: In current multi-port electro-hydrogen coupling energy router systems, the online estimation of hydrogen production efficiency in alkaline electrolyzers typically uses a fixed electrochemical parameter model, neglecting the real-time impact of the nonlinear change in bubble coverage with current density on the cell voltage. This leads to a deviation between the estimated and actual hydrogen production efficiency values, resulting in inaccurate decision-making basis for the optimization algorithm and unreasonable power allocation in the electrolyzer. This invention establishes an enhanced electrochemical model, using the current at each port and the real-time cell voltage feedback value of the alkaline electrolyzer as input, to solve the dynamic equation of bubble coverage online. It incorporates the nonlinear dynamics of bubble generation and desorption into the efficiency estimation, providing real-time correction to the estimated hydrogen production efficiency value. This allows the optimization algorithm to obtain more accurate efficiency input, making the calculated power commands for each port closer to the actual optimal operating point, thus improving the overall operating efficiency and the rationality of hydrogen production power allocation in the multi-port electro-hydrogen coupling energy router.

[0055] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, perform the aforementioned multi-port electro-hydrogen coupling energy router control method.

[0056] A third aspect of the present invention provides a multi-port electro-hydrogen coupling energy router control system, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0057] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems is that the essence of multi-port coupled oscillation is the superposition of negative impedances of the output impedances of each port converter within a specific frequency band. When the superposition amplitude exceeds the damping capacity of the bus, the system becomes unstable. This invention uses a cross-modal spatiotemporal graph attention model to predict the bus voltage fluctuation amplitude in advance, enabling the control system to obtain a quantified stability assessment value before the oscillation risk evolves into actual oscillation. This triggers the partitioning adjustment of the virtual impedance parameters in advance, shaping the frequency response of the output impedances of each port to within the stable domain, thus suppressing the negative impedance superposition effect below the threshold. The rolling time-domain optimization algorithm based on Lyapunov stability constraints ensures from the decision-making level that the system state evolves along the strictly decreasing direction of the Lyapunov function after each control step, meaning that the operating condition switching command itself will not destroy the system stability. The two mechanisms act on the impedance shaping layer and the decision optimization layer respectively, forming a dual stability guarantee, enabling the system to achieve reliable operating condition switching and oscillation suppression in multi-port strongly coupled scenarios.

[0058] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0059] The specific implementation method of step S01 is as follows.

[0060] Historical operating data, including port voltage, port current, port power, energy storage battery state of charge, alkaline concentration, membrane water content, and DC 750V bus voltage, were collected from the SCADA database of the multi-port electro-hydrogen coupled energy router at a 1-minute acquisition resolution. This data covered complete operating records under three operating conditions: fuel cell 10kV grid-connected power generation, off-grid hydrogen production, and weak grid-connected hydrogen production. Missing values ​​were imputed, outliers were removed, and the data was normalized. Using the historical data of the past 60 minutes as the input window, and the predicted power demand of each port and the predicted fluctuation range of the DC 750V bus voltage for the next 15 minutes as labels, time-series sample pairs were constructed and divided into training, validation, and test sets in a 7:2:1 ratio.

[0061] The topology of a multi-port electro-hydrogen coupling energy router system is abstracted as a graph structure. ,in This is a set of graph nodes, containing a total of 8 nodes, corresponding to the photovoltaic port, energy storage port, alkaline electrolyzer port, proton exchange membrane electrolyzer port, fuel cell port, AC 10kV grid port, DC load port, and AC load port, respectively. Let be a set of directed edges, and let the directed edge weight matrix be... The Middle Line number Column elements are ,in For nodes To the node Real-time transmission power (unit) ), Rated power reference value (unit) ), The weight is a dimensionless weight.

[0062] The underlying multi-head graph attention layer focuses on nodes. The formula for calculating attention weights is as follows: ; In the formula, and They are nodes With nodes The normalized dimensionless input feature vector For feature dimension, and The first The query matrix and key matrix of each attention head are both dimensionless learnable weight matrices. For the number of attention heads, For nodes For nodes In the Dimensionless attention score under each attention head. The function is a leaky linear rectified activation function, when the input... Time output When input Time output ,in The leakage coefficient has an empirical value of 0.01.

[0063] ; In the formula, For nodes The set of neighboring nodes, For nodes For neighboring nodes In the A person's attention is focused on Normalized dimensionless attention weights Normalization involves taking the exponent of each element and dividing by the sum of the exponents of all elements, so that the sum of the outputs is 1. For the neighboring node index.

[0064] node Spatial correlation weighted embedding vector It is obtained by splicing multiple heads, and the formula is expressed as follows: ; In the formula, This represents a vector concatenation operation. For the first The value matrix of each attention head is a dimensionless learnable weight matrix. It is a dimensionless embedding vector with dimension . , The value vector dimension is defined for each attention head. The convolutional layer performs three iterations of neighborhood aggregation, and after each aggregation, the original node features are preserved through residual skip connections. The formula for updating node features after the second aggregation is expressed as follows: ; In the formula, For the first Secondary aggregation node dimensionless eigenvectors Number of iterations , and They are nodes With nodes The degree in a graph structure (i.e., the number of directed edges connected to that node). For the first The dimensionless learnable weight matrix of the sub-aggregate for Activation function, i.e. , It is also a dimensionless vector.

[0065] Mid-level bidirectional gated cyclic unit timing encoder to node Historical time series feature sequences Perform bidirectional extraction in both forward and reverse directions, with the forward hidden state being... With reverse hidden state Both are dimensionless vectors with a default dimension of 128. Concatenating them yields the time-series encoded output. , The dimension is 256, and the weighted embedding vector is associated with the underlying space. After concatenation, the vector is fed into the top layer. Let be a dimensionless vector, where Index for the current time.

[0066] The top-level cross-modal fusion layer branches electrical quantities. Branches of chemical quantities After being fully connected encoded separately, they are then concatenated and fused into a unified graph node feature vector through attention-weighted concatenation. The fusion formula is expressed as follows: ; In the formula, For nodes Port voltage (unit) ), For nodes Port current (unit) ), Alkali concentration (unit) ), Membrane water content (unit) ), State of charge (SOC) of energy storage battery (dimensionless, representing the ratio of the battery's current remaining charge to its rated capacity). and These are fully connected coding layers for electrical and chemical quantities, respectively. Their inputs are normalized to be dimensionless quantities, and their outputs... and All are dimensionless vectors. and The dimensionless learnable attention weighting coefficients satisfy the following conditions: , This is a dimensionless fused feature vector.

[0067] The output layer has two task headers. The first task header outputs the predicted power demand values ​​for each port over the next 15 minutes. ( ,unit The second task head outputs the predicted value of the DC 750V bus voltage fluctuation. (unit ).

[0068] The total loss function during the training phase consists of a weighted sum of three terms, expressed in the following formula: ; In the formula, For nodes No. Actual power at any time (unit) ), The number of ports is set to 8. The actual value of the DC 750V bus voltage fluctuation amplitude (unit: ), The rated bus voltage is taken as 750 (unit). ), The prediction step count is set to 15. , , These are the dimensionless weighting coefficients for the power prediction mean square error term, the bus voltage prediction mean square error term, and the Kirchhoff current law residual physical constraint term, respectively. These three coefficients were determined through grid search experiments on the validation set, and their empirical values ​​are as follows: , , All three have passed by dividing by or Complete the dimensionless transformation. The total loss is dimensionless. The Adam optimizer is used, with an initial learning rate of 0.001, batch size of 64, and a maximum training epoch of 200. An early stopping mechanism is introduced, terminating training when the validation set loss does not decrease for 20 consecutive epochs. After training, the prediction performance is evaluated on the test set. If the mean absolute error of the predicted power demand values ​​for each port on the test set exceeds a preset error threshold, the model structure or hyperparameters are readjusted and retrained. The preset error threshold is determined through cross-validation experiments on multiple sets of historical data.

[0069] The specific implementation method of step S02 is as follows.

[0070] The currents collected at each port by S01 and the real-time cell voltage feedback values ​​of the alkaline electrolyzer are input into the enhanced electrochemical model to solve the dynamic equation for bubble coverage. The formula for the dynamic equation for bubble coverage is as follows: ; In the formula, Bubble coverage (dimensionless, value range 0 to 1). For time (unit) ), This is a dimensionless bubble generation rate coefficient. This is a dimensionless bubble desorption rate coefficient. Real-time current density of alkaline electrolyzer (unit) ), by the current at each port (unit Divide by the effective area of ​​the alkaline electrolytic cell electrode. (unit ) to obtain, that is ,in Obtained from the parameters on the equipment nameplate. The reference current density is set to 1.0 (unit). ), It is the dimensionless power exponent of bubble desorption. Bubble dynamic response time constant (units) The left side of the equation Units are In the right-end molecule and All are dimensionless quantities, and the whole is divided by (unit The unit after ) is The dimensions at both ends are consistent. The above parameters... , , , By covering a current density of 0.1–2.0 in the alkaline electrolytic cell... Bubble generation and desorption experiments were conducted within a certain range, and the nonlinear response data of the real-time cell voltage feedback value and the current at each port of the alkaline electrolyzer were obtained through fitting analysis. The experimental steps were as follows: Step 1, at current densities of 0.1, 0.3, 0.5, 0.8, 1.0, 1.5, and 2.0... Under typical operating conditions, the real-time cell voltage of the alkaline electrolytic cell was recorded. (unit ) and corresponding current density The time-domain response curve; Step 2, use the nonlinear least squares method to fit the parameters of the dynamic equation of bubble coverage, so as to minimize the sum of squared residuals between the simulated tank voltage and the measured tank voltage, and obtain , , , The optimal estimate.

[0071] Based on bubble coverage correction value The estimated hydrogen production efficiency of the alkaline electrolyzer is corrected online, and the corrected formula for the hydrogen production efficiency is expressed as follows: ; In the formula, This is a revised estimate of hydrogen production efficiency (dimensionless). The uncorrected baseline hydrogen production efficiency estimate (dimensionless) is calculated using the rated current density. The steady-state hydrogen production rate was determined by an experiment on an alkaline electrolyzer, and the ratio of the measured hydrogen production rate to the theoretical Faraday hydrogen production rate was used for calculation. This is the dimensionless correction value for bubble coverage obtained by solving the dynamic equation for bubble coverage. The bubble coverage rate (dimensionless) under reference operating conditions was determined through a steady-state experiment at the rated current density. , , and Both are dimensionless quantities, with the same dimensions at both ends.

[0072] The specific implementation method of step S03 is as follows.

[0073] Predicted DC 750V bus voltage fluctuation based on S01 output Calculate the bus voltage stability assessment value The formula is expressed as follows: ; In the formula, Units are , Units are , This is a dimensionless evaluation value. (Partition threshold) and ( All are dimensionless quantities. They were obtained by conducting at least 50 sets of experiments under different load disturbance amplitudes and different port switching combinations, recording the time-domain response data of the DC 750V bus voltage deviation and the power deviation at each port, and then determining the stability domain boundary through statistical analysis. When the system is in the stable region, it sends small virtual impedance parameters to the converter-level controller; when At that time, the system is in the transition region, and medium virtual impedance parameters are issued; when At this time, the system is in the oscillation risk zone, and large virtual impedance parameters are issued. The virtual impedance parameters corresponding to each interval are determined through impedance characteristic experiments and oscillation suppression effect analysis under the corresponding operating conditions. Virtual impedance decoupling control introduces a virtual impedance link at the output of each converter control loop. The impedance characteristics are simulated through the control algorithm, and the frequency response of the output impedance of each port is reshaped in real time, so that the impedance characteristics of multiple parallel converters meet the stability criterion, thereby suppressing the DC 750V bus coupling oscillation caused by the superposition of negative impedance characteristics.

[0074] The specific implementation method of step S04 is as follows.

[0075] The predicted power demand for each port output by S01 over the next 15 minutes, the corrected hydrogen production efficiency estimate output by S02, and the current operating condition identifier are input using a rolling time-domain multi-condition switching optimization algorithm based on Lyapunov stability constraints. The multi-condition switching problem of the multi-port electro-hydrogen coupled energy router is modeled as a mixed-integer optimal control problem with switching constraints, with a prediction time domain of 15 minutes and a control time domain of 1 minute. System state vector. Due to DC 750V bus voltage deviation (Dimensionless) and power deviation at each port (dimensionless, It consists of ) for Bus voltage deviation at any time (unit) ), for Time of the first Port power deviation (unit) ), It is a dimensionless state vector. Decision variables include binary variables for operating condition switching. ( For operating condition number, ) and dimensionless power commands for each port .

[0076] The formula for constructing the Lyapunov function is as follows: ; In the formula, It is a dimensionless positive definite weight matrix. The dimension of the state vector is set to 9. As a dimensionless scalar, its specific form and the weight coefficients of each element are determined through iterative analysis of simulation experiments and actual machine disturbance experiments under three operating conditions. It is necessary to ensure that the deceleration rate of the Lyapunov function value under typical disturbances in each operating condition meets the preset convergence speed requirements.

[0077] The Lyapunov stability constraint formula for each control step is expressed as follows: ; In the formula, A convergence rate coefficient is preset for the dimensionless form, with an empirical value of 0.05. and All are dimensionless scalars. This constraint is embedded as a hard constraint in the optimization problem to ensure that the system does not become unstable during the switching of operating conditions.

[0078] The objective function is expressed as follows: ; In the formula, For the first port number Power command at any time (unit) ), The output of S01 port number Power demand forecast at any time (unit) The difference between the two is divided by The latter is a dimensionless quantity. This is a dimensionless penalty coefficient for switching operating conditions, with an empirical value of 10, used to suppress frequent switching. and All variables are dimensionless binary variables, and all terms in the objective function are dimensionless. Integer variables are solved using the branch and bound method, with Lagrange relaxation applied to the Lyapunov stability constraints. Dual variables... The dual variable is updated using the subgradient method. In each iteration, the dual variable is updated along the subgradient direction, gradually approximating the optimal solution to the Lagrange dual problem. The formula for updating the dual variable is as follows: ; In the formula, For iteration counting, For the first The dimensionless dual variable in the next iteration For the first The dimensionless subgradient of the Lyapunov stability constraint at the nth iteration For the first The dimensionless step size of the next iteration adopts a decreasing step size strategy, as expressed by the following formula: ; In the formula, The initial step size is dimensionless, with an empirical value of 0.1. The upper limit of the inner iteration is 20 times, and the total calculation time is controlled within 30 seconds.

[0079] The specific implementation methods for steps S05 and S06 are as follows.

[0080] Based on the operating condition switching command output by S04, the current operating condition is determined to be either 10kV grid-connected fuel cell power generation, off-grid hydrogen production from renewable energy, or weak grid-connected hydrogen production from renewable energy. The corresponding control strategy is then used to issue operating mode and power commands to each port converter. Upon receiving the operating mode and power commands, the converter-level controller executes current inner-loop control, power calculation, and pulse-width modulation phase-shift control to maintain stable DC 750V bus voltage and power balance at each port.

[0081] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: The purpose of this invention is to solve the problems of hydrogen load differential capacity configuration and complex multi-port independent control brought about by the multi-port electro-hydrogen coupling energy router after hydrogen energy is connected. It proposes a hierarchical and zoned control strategy applicable to various operating conditions. Through photovoltaic power generation prediction, energy storage charging and discharging statistics, production planning decision-making, photovoltaic consumption and utilization rate analysis, energy consumption analysis and other means, it realizes the efficient integration of source, grid, load and hydrogen storage through the multi-port electro-hydrogen coupling energy router, meets the needs of coordinated regulation at the power dispatch and chemical operation levels under various modes, and supports load following source and safe and economical grid operation.

[0082] This system achieves multi-source data fusion across hydrogen generation, grid, load, and storage, extracts relevant data features, and uses techniques such as LSTM / GRU time series models, attention mechanism models, BP neural network fitting, and second-order inertial model fitting to predict new energy output, load, equipment characteristics, and frequency / voltage response. It also achieves multi-objective optimization of the system through power balance constraints, hydrogen storage capacity constraints, and voltage / frequency safety constraints. Finally, it utilizes techniques such as PID control, fuzzy control, Bayesian networks, random forests, game theory, and reinforcement learning to implement decision-making and control of a multi-port electro-hydrogen coupled energy router.

[0083] The topology diagram of the multi-port electro-hydrogen coupling energy router system with hydrogen energy access function described in this invention is as follows: Figure 2 As shown, by constructing a 750V DC bus and configuring 8 interfaces, including 6 DC interfaces for ALK electrolyzers, PEM electrolyzers, fuel cells, photovoltaics, energy storage, and DC loads, and 2 AC interfaces for 10kV AC grids and AC loads, the system realizes the conversion, transmission, and routing of electrical energy with different electrical parameters, and can achieve reasonable power allocation according to the operating conditions of each input and output port.

[0084] The multi-port electro-hydrogen coupling energy router management system of this invention adopts a hierarchical partitioning architecture, and the system structure is as follows: Figure 3 As shown, the control system is divided into a two-layer control architecture: system-level and converter-level. The system-level control, based on J2EE technology, MVC development model, MyEclipse development tool, and SCADA database technology, includes functions such as panoramic monitoring, production planning decision-making, optimization control, photovoltaic power consumption analysis, photovoltaic power generation prediction, energy consumption analysis, energy storage charging and discharging statistics, and operating condition control. The converter-level control includes functions such as frequency and voltage stabilization, current inner-loop control, PQ calculation, and PWM phase-shift control, ensuring real-time power balance, rapid frequency and voltage stabilization, maximum local consumption of new energy sources, safe and reliable power supply, and proactive emergency fault defense for the multi-port electro-hydrogen coupled energy router.

[0085] To verify the effectiveness of this invention, technicians set up a test environment, using a multi-port electro-hydrogen coupling energy router system as the object. The system was configured with a photovoltaic rated power of 500kW, an energy storage rated capacity of 1000kWh, an alkaline electrolyzer rated power of 200kW, a proton exchange membrane electrolyzer rated power of 100kW, and a fuel cell rated power of 150kW. A 10kV AC grid side was connected to the grid using a power electronic transformer with a parallel input and series output structure. A 750V DC bus collected energy from each port. The SCADA database continuously collected operating data with a time resolution of 1 minute. The historical data covered complete operating records under three operating conditions, with a total data volume of more than 180 days of continuous records.

[0086] Technicians first established a training dataset for the cross-modal spatiotemporal graph attention model using historical data. The dataset used the past 60 minutes of historical data as the input window, and the predicted power demand at each port and the predicted fluctuation amplitude of the DC 750V bus voltage for the next 15 minutes as labels. Time-series sample pairs were constructed and divided into training, validation, and test sets in a 7:2:1 ratio. The Adam optimizer was used during training, with an initial learning rate of 0.001 and a batch size of 64. An early stopping mechanism was introduced, terminating training when the validation set loss did not decrease for 20 consecutive rounds, with a maximum of 200 training rounds. The loss function consisted of a weighted sum of the mean square error of power prediction, the mean square error of DC 750V bus voltage prediction, and the physical constraint term of Kirchhoff's current law residual. The weights of these three terms were determined through grid search experiments on the validation set. The mean absolute error of the predicted power demand at each port in the test set met the preset error threshold. After model training, the model was deployed to the system-level control platform.

[0087] The parameters of the dynamic equation for bubble coverage in the enhanced electrochemical model are obtained by adjusting the current density in the alkaline electrolyzer at 0.1. Up to 2.0 Bubble generation and desorption experiments were conducted within the specified range, and the nonlinear response data of the tank voltage feedback value and the current at each port were fitted and analyzed. Under typical operating conditions, the current density is 1.0. At that time, the bubble coverage correction value was about 0.18, and the online correction to the hydrogen production efficiency estimate was about 3.2 percentage points. After correction, the hydrogen production efficiency estimate was closer to the measured value.

[0088] The partition threshold of the bus voltage stability assessment value and Experiments were conducted under 50 different load disturbance amplitudes and port switching combinations to record the time-domain response data of the DC 750V bus voltage deviation and the power deviation at each port. The stability domain boundary was determined through statistical analysis. The virtual impedance parameters corresponding to each interval were determined through impedance characteristic experiments and oscillation suppression effect analysis under the corresponding operating conditions, as shown in Table 1. Table 1. Correspondence between bus voltage stability assessment values ​​by zone and virtual impedance parameters

[0089] In a typical test scenario, technicians simulated the process of switching from off-grid hydrogen production to 10kV grid-connected power generation using fuel cells. Before the switch, the energy storage port supported the 750V DC bus voltage in V / F mode, the photovoltaic port operated in maximum power point tracking mode, and the alkaline electrolyzer and proton exchange membrane electrolyzer ports operated in P / Q mode. At the moment of switching trigger, the cross-modal spatiotemporal graph attention model predicted an increase in the predicted amplitude of the 750V DC bus voltage fluctuation, indicating that the bus voltage stability assessment value had entered the oscillation risk zone. The system immediately issued a large virtual impedance parameter to the converter-level controller to shape the output impedance of each port and suppress coupled oscillations. Simultaneously, the rolling time-domain multi-condition switching optimization algorithm based on Lyapunov stability constraints, with the corrected hydrogen production efficiency estimate and power prediction as input, completed the solution within 30 seconds and output the condition switching command and the power command for each port. After the switchover, the AC 10kV grid port operates in V / F mode to support the DC 750V bus voltage, the energy storage port operates in P / Q mode to maintain power balance, and the fuel cell port connects to the DC 750V bus in P / Q mode to generate electricity. The system smoothly transitions to the fuel cell 10kV grid-connected power generation condition. Figure 4 As shown.

[0090] Technicians further tested the operational performance under both off-grid hydrogen production and weak grid-connected hydrogen production conditions, such as... Figure 5 and Figure 6 As shown in Table 2, under the condition of weak grid connection of new energy hydrogen production, the interaction power between the system and the 10kV AC grid is strictly controlled within 10% of the hydrogen production load power. The control system dynamically adjusts the operating power of the alkaline electrolyzer and the proton exchange membrane electrolyzer according to the photovoltaic power output and the state of charge of the energy storage battery, so as to maximize the consumption of new energy. Typical daily operating data is shown in Table 2. Table 2. Statistical Table of Operating Parameters for Three Typical Daily Operating Conditions

[0091] From a fundamental perspective, this invention represents an improvement over traditional methods: Traditional fixed-parameter controllers cannot detect dynamic changes in oscillation risk in multi-port parallel scenarios; operating condition switching relies on engineering experience thresholds; and power prediction is based solely on a single time-series feature. These three aspects are independent and lack coordination. This invention, however, uses a cross-modal spatiotemporal graph attention model to jointly model electrical and chemical quantities within a unified graph structure. This allows power prediction to simultaneously perceive the influence of the electrolytic cell's electrochemical state. The prediction results directly drive the partitioned adjustment of virtual impedance parameters, forming a feedforward oscillation suppression mechanism. Furthermore, the rolling time-domain optimization algorithm based on Lyapunov stability constraints embeds stability conditions as hard constraints, ensuring mathematical stability in operating condition switching decisions. This fundamentally eliminates the problem of coexisting oscillation risk and decision instability in traditional methods.

[0092] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0093] Table 3. Variable Explanation Table (Part 1)

[0094] Table 4. Variable Explanation Table (Part Two)

[0095] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A control method for a multi-port electro-hydrogen coupling energy router, characterized in that, Includes the following steps: The SCADA database of the multi-port electro-hydrogen coupled energy router collects the voltage, current, power, state of charge of the energy storage battery, alkaline concentration, membrane water content, and DC 750V bus voltage of each port. The collected data are input into the cross-modal spatiotemporal graph attention model, and the predicted values ​​of future power demand for each port and the predicted values ​​of DC 750V bus voltage fluctuation are output. The collected port currents and real-time cell voltage feedback values ​​of the alkaline electrolyzer are input into the enhanced electrochemical model to solve the dynamic equation of bubble coverage, output the bubble coverage correction value, and perform online correction on the estimated value of hydrogen production efficiency of the alkaline electrolyzer based on the bubble coverage correction value, and output the corrected estimated value of hydrogen production efficiency. The bus voltage stability assessment value is calculated based on the predicted value of the DC 750V bus voltage fluctuation amplitude. According to the range to which the bus voltage stability assessment value belongs, the corresponding virtual impedance parameters are sent to the converter-level controller to shape the output impedance of each port and suppress multi-port coupled oscillation. The predicted future power demand of each port, the corrected hydrogen production efficiency estimate, and the current operating condition identifier are input into the rolling time-domain multi-operating condition switching optimization algorithm based on Lyapunov stability constraints. Within each control step, the power command and operating condition switching command of each port are output. Based on the operating condition switching command, determine the current operating condition, and issue operating mode commands and power commands to each port converter according to the control strategy corresponding to the operating condition. The converter-level controller receives operating mode commands and power commands, and performs current inner loop control, PQ calculation and PWM phase shift control to maintain the stability of the DC 750V bus voltage and the power balance of each port.

2. The multi-port electro-hydrogen coupling energy router control method according to claim 1, characterized in that, The cross-modal spatiotemporal graph attention model specifically abstracts the system topology of a multi-port electro-hydrogen coupled energy router into a graph structure, with 8 ports serving as graph nodes, power flow between ports serving as directed edges, and the weight of the directed edges dynamically assigned by the real-time transmission power. The network adopts a three-layer stacked structure: a bottom-layer multi-head graph attention layer, a middle-layer bidirectional gated cyclic unit timing encoder, and a top-layer cross-modal fusion layer.

3. The multi-port electro-hydrogen coupling energy router control method according to claim 2, characterized in that, The top-level cross-modal fusion layer specifically encodes the electrical quantity branch and the chemical quantity branch separately through fully connected encoding, and then merges them into a unified graph node feature vector through attention weighting. The electrical quantity branch includes the voltage and current of each port, and the chemical quantity branch includes the alkali concentration, membrane water content and energy storage battery state of charge.

4. The multi-port electro-hydrogen coupling energy router control method according to claim 3, characterized in that, The training of the cross-modal spatiotemporal graph attention model involves adding Kirchhoff's current law residual as a physical constraint loss term to the loss function. The total loss function is composed of a weighted sum of the power prediction mean square error term, the DC 750V bus voltage prediction mean square error term, and the physical constraint loss term. The three weights are determined through grid search experiments on the validation set.

5. The multi-port electro-hydrogen coupling energy router control method according to claim 4, characterized in that, The training dataset for the cross-modal spatiotemporal graph attention model is specifically collected from historical operating data in the SCADA database. The historical data is used as the input window, and the predicted power demand of each port and the predicted voltage fluctuation of the DC 750V bus are used as labels to construct time series sample pairs. The dataset is divided according to the proportion of the training set, the proportion of the validation set, and the proportion of the test set.

6. The multi-port electro-hydrogen coupling energy router control method according to claim 5, characterized in that, The parameters of the dynamic equation for bubble coverage in the enhanced electrochemical model are obtained by performing bubble generation and desorption experiments within a preset range of current density coverage in the alkaline electrolyzer, and by fitting and analyzing the nonlinear response data of the real-time cell voltage feedback value and the current at each port of the alkaline electrolyzer.

7. The multi-port electro-hydrogen coupling energy router control method according to claim 6, characterized in that, The bus voltage stability assessment value is based on the zoning threshold. and The system is divided into a stable zone, a transition zone, and an oscillation risk zone. When the bus voltage stability assessment value does not exceed... Currently, small virtual impedance parameters are issued. When the bus voltage stability assessment value is located at... and When the intermediate virtual impedance parameter is issued, the bus voltage stability assessment value exceeds Currently, large virtual impedance parameters are being issued.

8. The multi-port electro-hydrogen coupling energy router control method according to claim 7, characterized in that, The partition threshold and Specifically, this is achieved by conducting at least a preset number of experiments under different load disturbance amplitudes and different port switching combinations, recording the time-domain response data of the DC 750V bus voltage deviation and the power deviation of each port, and then determining the stability domain boundary through statistical analysis.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the multi-port electro-hydrogen coupling energy router control method according to any one of claims 1-8.

10. A multi-port electro-hydrogen coupling energy router control system, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images