A green electricity hydrogen production coordinated control method satisfying multiple control targets and multiple constraint conditions

Abstract

The application discloses a green electricity hydrogen production coordinated control method meeting multiple control targets and multiple constraint conditions, and relates to the field of green electricity hydrogen production coordinated control.The method comprises the following steps: based on a multi-target constraint model, a population structure is set by using an improved constraint grouping multi-population evolution algorithm, a multi-population evolution framework is obtained, and the population in the multi-population evolution framework is independently evolved and recombined by combining a dynamic relaxation and a hibernation activation mechanism; the recombined population is co-evolved, a physical barrier perception memory graph is constructed according to the evolved population, and population transition is performed, so that a Pareto frontier solution set is obtained; based on a verification and evaluation algorithm of a good and bad solution distance method and a meteorological extreme error, the Pareto frontier solution set is screened, and a weight parameter optimal solution is updated to a model predictive controller, so that green electricity hydrogen production coordinated control is realized.The application gives consideration to double-time-scale coordinated control of global optimization and real-time response, and realizes the synergistic improvement of hydrogen production benefits and new energy consumption benefits maximization.

Description

Technical Field

[0001] This invention relates to the field of coordinated control of green electricity hydrogen production, and more specifically, to a coordinated control method for green electricity hydrogen production that satisfies multiple control objectives and multiple constraints. Background Technology

[0002] By the end of 2025, China's grid-connected installed capacity of wind power and photovoltaic power reached 640 million kilowatts and 1.2 billion kilowatts respectively, with a combined installed capacity of 1.84 billion kilowatts, accounting for 47% of the country's total installed power capacity, surpassing thermal power for the first time to become the main source of power generation. As the penetration rate of new energy sources continues to rise, their intermittency and volatility pose increasingly prominent challenges to the safe and stable operation of the power grid.

[0003] Hydrogen production through water electrolysis offers significant advantages as an adjustable and flexible load, including rapid response and zero carbon emissions. Compared to traditional energy storage media, hydrogen energy boasts high energy density and the technical and economic feasibility of cross-seasonal storage, enabling spatiotemporal decoupling between renewable energy generation and end-use energy. Grid-connected hydrogen production systems not only help smooth fluctuations in wind and solar power output and reduce curtailment caused by insufficient grid absorption capacity, but also participate in grid peak-shaving ancillary services, enhancing the system's dynamic regulation capabilities and providing crucial support for the safe, stable, and economical operation of the power grid under a high-proportion renewable energy integration environment.

[0004] Therefore, how to fully leverage the regulation potential of electrolysis hydrogen production technology to meet the needs of multiple objectives such as smoothing power fluctuations, expanding the space for renewable energy consumption, and participating in grid peak shaving ancillary services has become a key issue that urgently needs to be addressed.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes a coordinated control method for green electricity hydrogen production that satisfies multiple control objectives and multiple constraints, thereby overcoming the aforementioned technical problems existing in existing related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] A coordinated control method for green electricity-to-hydrogen production that satisfies multiple control objectives and constraints includes the following steps:

[0009] Based on the tuning problem of the weight parameter matrix in the model predictor of a wind-solar-storage grid-connected hydrogen production system, a multi-objective constrained model is constructed.

[0010] Based on the multi-objective constraint model, an improved constraint grouping multi-population evolution algorithm is used to set the population structure, resulting in a multi-population evolution framework. The dynamic relaxation and dormancy activation mechanism is combined to independently evolve and recombine the populations in the multi-population evolution framework, resulting in a recombined population.

[0011] A hybrid strategy of fitness landscape co-analysis was used to co-evolve the recombined population to obtain the evolved population. A physical barrier perception memory map was constructed based on the evolved population, and population transition based on spectral clustering was performed based on the physical barrier perception memory map to obtain the Pareto front solution set of the weight parameters.

[0012] Based on the superior-inferior solution distance method and the verification and evaluation algorithm of meteorological extreme errors, the optimal solution of the Pareto front solution set of the weight parameters is screened to obtain the optimal solution of the weight parameters. The optimal solution of the weight parameters is then smoothly weighted and updated to the model predictive controller to achieve coordinated control of green electricity to hydrogen production.

[0013] The beneficial effects of this invention are as follows:

[0014] 1. This invention achieves multi-objective coordinated control of an electric-hydrogen coupled system based on the improved constrained grouped multi-population evolutionary algorithm (CGMEA) and model predictive control (MPC) collaborative optimization. To address the dynamic balance between renewable energy consumption and high-value hydrogen production under grid-connected conditions, a multi-objective constrained model predictive control framework is introduced, and a coordinated control strategy for green electricity-to-hydrogen production that satisfies multiple control objectives and constraints is presented. The proposed control method addresses the strong coupling and multi-objective conflict between renewable energy consumption and high-value hydrogen production under grid-connected conditions. The outer loop uses a slow-timescale improved CGMEA for global Q and R parameter optimization, while the inner loop uses a fast-timescale MPC for local rolling execution. Specifically, before the control parameters are formally updated, this invention introduces a first-order spatiotemporal correlation recursive equation based on a verification and evaluation algorithm that includes meteorological extreme errors. This reconstructs meteorological errors into a continuously evolving spatiotemporally extreme disturbance field, accurately capturing the temporal propagation inertia of meteorological disturbances and the spatial coupling effect of wind and solar energy. By quantifying extreme value perturbation boundaries and extrapolating dynamic robustness boundaries, potential parameters that could cause the system to exceed limits under abnormal operating conditions are eliminated in advance, thus achieving parameter verification from static optimality to dynamic reliability. Under the premise of strictly meeting multiple constraints such as grid interaction, equipment operation, and power change rate, this architecture balances global optimization and real-time response in dual-timescale coordinated control, achieving a synergistic improvement that maximizes hydrogen production revenue and renewable energy consumption benefits.

[0015] 2. This invention introduces a dynamic relaxation function that transitions from loose to tight and a dormancy and activation mechanism based on the displacement of the population center point to dynamically identify effective constraints and schedule limited computing power, thus meeting the computing power requirements for short-cycle online real-time control of the electro-hydrogen coupling system. Simultaneously, addressing the order-of-magnitude difference in response time between electromagnetic transient devices (such as batteries) and thermochemical devices (such as electrolyzers), it utilizes non-uniform simulated binary crossover operations based on physical inertial mapping to overcome the blindness of the traditional operator's globally unified probability. This allows for flexible reconfiguration of battery parameters while isolating the fatigue impact of high-frequency disturbances on slow-response devices such as electrolyzers.

[0016] 3. Compared with intelligent optimization algorithms that use traditional blind crossover and mutation operators, this invention addresses the problem of out-of-limit solutions easily generated at the operational limits of green electricity-to-hydrogen systems by introducing a hybrid strategy based on fitness-based landscape collaborative analysis. When individuals in the population approach the output limit of the electrolyzer or the state of charge (SOC) boundary of the battery, the algorithm adaptively freezes random mutations and initiates a topology-guided directional walk operator. To prevent out-of-limit solutions from being generated due to excessively large fixed step sizes when the algorithm slides close to physical constraint boundaries, an exponential spatial damping function considering the physical safety margin of the system is introduced, giving the algorithm environmental awareness of the system's physical boundaries. Simultaneously, a multi-scale topological memory graph is constructed, and spectral clustering is used to intelligently guide trapped populations to efficiently escape infeasibility traps, avoiding random restarts. This ensures the continuity of the Pareto front and the completeness of the exploration of constraint boundaries, improving the convergence speed and solution determinism of the MPC weight matrix optimization. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0018] Figure 1 This is a flowchart of a coordinated control method for green electricity-to-hydrogen production that satisfies multiple control objectives and multiple constraints according to an embodiment of the present invention;

[0019] Figure 2 This is a schematic diagram of a coordinated control system for green electricity-to-hydrogen production that satisfies multiple control objectives and multiple constraints according to an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0021] Figure 4 This is a flowchart of the improved CGMEA algorithm according to an embodiment of the present invention;

[0022] Figure 5 This is a flowchart of the collaborative control of the CGMEA outer loop optimizer and the MPC inner loop rolling optimization according to an embodiment of the present invention.

[0023] In the picture:

[0024] 1. Multi-objective constraint model construction module; 2. Multi-population evolution module; 3. Physical barrier perception memory graph construction module; 4. Optimal solution selection module. Detailed Implementation

[0025] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0026] According to embodiments of the present invention, a coordinated control method for green electricity-to-hydrogen production that satisfies multiple control objectives and multiple constraints is provided.

[0027] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a coordinated control method for green electricity-to-hydrogen production that satisfies multiple control objectives and multiple constraints is provided. The method includes the following steps:

[0028] Based on the tuning problem of the weight parameter matrix in the model predictive controller of a wind-solar-storage grid-connected hydrogen production system, a multi-objective constrained model is constructed.

[0029] Specifically, based on the tuning problem of the weight parameter matrix in the model predictive controller of a wind-solar-storage grid-connected hydrogen production system, a multi-objective constrained model is constructed, including:

[0030] Establish the power balance equation of the wind-solar-storage grid-connected hydrogen production system, as well as the power balance equation of net power and controllable units, and construct the continuous-time state-space equation based on the dual-time-scale coordinated control architecture.

[0031] Based on the forward Euler method, the continuous-time state-space equation is discretized to obtain the state-space prediction model of the wind-solar-storage-grid-hydrogen production system.

[0032] Based on multiple constraints, boundary restrictions are imposed on the state variables and constraint output variables in the state-space prediction model, and a multi-objective constraint model containing multiple control objectives is constructed.

[0033] Specifically, such as Figure 4 and Figure 5As shown, the model predictive controller in this invention utilizes model predictive control to achieve coordinated control of multiple variables and objectives in an electro-hydrogen coupling system. However, considering that traditional model predictive control relies on empirically selected weight matrices, it is difficult to achieve quantitative balance among multiple conflicting objectives such as maximizing hydrogen production, maximizing renewable energy consumption, and minimizing grid interaction fluctuations. This can easily lead to excessive compromise or neglect of certain operational objectives, making it difficult to solve the optimization problem of the electro-hydrogen coupling system. Therefore, an improved constrained grouping multi-population evolutionary algorithm is adopted to dynamically optimize the calculation process of the model predictive control weight matrix, forming a dual-loop collaborative control architecture of the improved constrained grouping multi-population evolutionary algorithm (CGMEA) and model predictive control (MPC) to achieve multiple control objectives such as maximizing hydrogen production and maximizing renewable energy consumption, thus enabling multi-objective coordinated optimization control of the electro-hydrogen coupling system.

[0034] Step 1: Employing a dual-timescale architecture of inner-layer MPC (fast-timescale control) and outer-layer CGMEA (slow-timescale optimization), the electro-hydrogen coupling system model is discretized using the forward Euler method, establishing a linear discrete state-space prediction model containing the system matrix and control matrix. The system's state variables, control variables, controlled output variables, and constraint output variables are defined. This discrete state-space model forms the prediction basis for the inner-loop MPC and is used to quantify the multi-objective fitness of individuals in the population and accurately verify the degree of constraint boundary violation.

[0035] Step 2: Considering multiple constraints such as grid connection power, grid connection electricity, rate of change of renewable energy power generation, and rate of change of load power, establish objective functions for multiple control objectives such as maximizing hydrogen production and maximizing renewable energy consumption; all inequality constraints and equality constraints in the system are boundary restrictions on the state variables and constraint output variables in Step 1.

[0036] Step 3: Introduce an improved Constrained Grouping Multipopulation Evolutionary Algorithm (CGMEA). This algorithm, based on the original CGMEA, adds a population activation and dormancy mechanism and a dynamic relaxation function. The former dynamically switches between dormant and activated states based on the convergence state of each subpopulation, rationally allocating computational power; the latter adjusts the strictness of constraints from loose to tight as the evolution progresses, balancing the diversity and accuracy of the solution set and enhancing the algorithm's ability to handle complex constraints. Addressing the blindness of the traditional algorithm's globally uniform crossover probability, the algorithm extracts the inherent physical time constants of the underlying physical devices corresponding to each decision variable, such as electrolyzers and batteries, and uses an exponential function to construct crossover probabilities. This ensures that the weights of batteries, which have extremely fast responses, maintain high-frequency recombination, while the weights of electrolyzers, which have extremely high inertia, tend to be inherited smoothly. In the later stages of evolution, a hybrid strategy based on fitness-landscape collaborative analysis is introduced, adding a topology analysis algorithm to guide the population to directional walk along low-curvature directions in narrow feasible channels. Simultaneously, an exponential spatial damping step size related to the physical constraint margin is introduced. The closer an individual is to the constraint boundary, the greater the spatial damping of the step size, achieving high-quality selection of Q and R parameters.

[0037] Step 4: Combine the improved CGMEA with Model Predictive Control (MPC) to form a dual-loop collaborative control framework. After Step 3, the improved CGMEA outputs a Pareto-optimal combination of Q and R parameters. It then includes a verification and evaluation algorithm for meteorological extreme errors, calculating the dynamic deviation trajectory of the system under extreme weather disturbances. This deviation trajectory is then substituted into all physical constraints for secondary safety verification. If all constraints are still satisfied, it proves that the parameters are not only statically optimal but also dynamically robust, and the verification passes. If a constraint is broken, the parameter is automatically removed, and the next best parameter is extracted sequentially along the TOPSIS (Top-Solution Distance Method) for re-verification. If the maximum number of attempts fails, a fault-tolerance mechanism is triggered, and the safety parameters from the previous cycle are used. After successful verification, a first-order low-pass filtering strategy is used to smoothly update the weight parameters to the inner-loop MPC, avoiding abrupt changes in control increments. The MPC solves a quadratic programming problem based on the updated parameters to form control strategies under different constraint conditions.

[0038] This invention improves the CGMEA algorithm by using a dynamic relaxation function to balance the feasibility of the population in the early stages of iteration and the convergence in the later stages, and by using a population dormancy and activation mechanism to improve the utilization efficiency of computational resources, thus enriching population diversity while ensuring evolutionary efficiency. Furthermore, a hybrid evolutionary strategy based on fitness-landscape co-analysis is introduced. By evaluating the local curvature of the target space online, conventional blind mutations are frozen when individuals in the population approach the physical hard limit, and topological directional walks along the tangent plane of the activation constraint are initiated, thereby guiding individuals to perform gradient sliding mining and reducing the probability of generating invalid solutions at high-dimensional physical boundaries. The resulting Pareto optimal Q and R parameter combinations are used to select the optimal solution based on TOPSIS for the next rolling optimization step of MPC. The constructed model simultaneously considers multiple control objectives such as maximizing hydrogen production and maximizing the new energy consumption rate, forming different control strategies under different constraints and objectives.

[0039] The green electricity hydrogen production system includes a new energy power generation unit, a battery unit, an electrolyzer unit, a hydrogen energy storage unit, and a load unit. Each unit is connected to a DC bus via a converter. The power generation unit is a wind power generation unit. This invention, based on an improved coordinated control strategy combining CGMEA and MPC for the green electricity hydrogen production system, includes the following steps:

[0040] Step 1: Establish a state-space prediction model for a wind-solar-storage grid-connected hydrogen production system, which includes:

[0041] 1) Establish the system power balance equation:

[0042] ;

[0043] Among them, P wt P represents the output power of the wind power generation unit. pv P represents the output power of the photovoltaic power generation unit. el P is the input power of the electrolytic cell. fc P represents the active power of the fuel cell. bat P represents the charging and discharging power of the battery cell. load P represents the total load power. grid This refers to the interaction power between the power grid and the coupled system.

[0044] 2) The power balance equations for the system's net power and the controllable unit are as follows:

[0045] ;

[0046] In the formula, net power P net P represents the difference between renewable energy and load demand. gen Power generated from controllable energy sources.

[0047] 3) Adopting a dual-time-scale coordinated control architecture:

[0048] Fast timescale: Inner MPC control, sampling period T s It lasts for 1 minute and is used for real-time tracking and dynamic response.

[0049] Slow timescale: outer CGMEA optimization, optimization period T opt It is allotted 30 minutes for strategy adjustments and multi-objective trade-offs.

[0050] 4) Discretize the linear discrete system using the forward Euler method and establish its state-space expression:

[0051] ;

[0052] In the formula, k is the current time; k+1 is the next time; x is the state variable; u is the control variable; y c y is the controlled output variable; b The output variables are constrained; A and B are the system matrix and control matrix, respectively; C c C is the controlled output matrix; b To constrain the output matrix.

[0053] This invention selects the electrolytic cell power P el Battery power P bat Fuel cell power P fc Battery SOC and hydrogen storage tank SOH are used as state variables; the electrolyzer output increment ΔP is selected. el Battery output increment ΔP bat Fuel cell output increment ΔP fc And the incremental difference P between renewable energy and load demand net As a control variable, the electrolytic cell power P is selected. el Battery power P bat Fuel cell power P fc As the controlled output y c Select the electrolytic cell power P el Battery power P bat Fuel cell power P fc Battery SOC, hydrogen storage tank SOH, grid interconnection power P grid As a constraint output y b Its discretization process is as follows:

[0054] Let the continuous-time state-space equation be:

[0055] ;

[0056] Where x(t) is the state vector, u(t) is the control vector, d(t) is the disturbance vector, and A c B c Ec For a continuous system matrix.

[0057] Discretization is performed using the first-order forward Euler method, with a sampling period of T. s Then the derivative is approximately:

[0058] ;

[0059] Where x(k) represents the state value at the k-th sampling time.

[0060] Substitute the approximation into the continuity equation and let t = kT s ,have to:

[0061] ;

[0062] Both sides are multiplied by T s By rearranging the terms, we obtain the discrete state-space equations:

[0063] ;

[0064] I represents the system matrix A c Identity matrices of the same dimension; defining the discrete system matrix:

[0065] ;

[0066] The discretized state-space expression is then:

[0067] ;

[0068] Therefore, the state-space expression of this system is:

[0069] ;

[0070] ;

[0071] ;

[0072] The elements in the control matrix B can be represented as:

[0073] ;

[0074] In the formula, Represents the ideal gas constant; Indicates the temperature of the hydrogen storage tank; Indicates Faraday efficiency; Indicates the number of electrolytic cells connected in series; Indicates the sampling period; Indicates the volume of the hydrogen storage tank; Indicates the number of electrons transferred in each reaction; Denotes Faraday's constant; Indicates the output voltage of the electrolytic cell; Indicates the upper limit of the hydrogen storage tank pressure; Indicates the number of fuel cells connected in series; Indicates the voltage of the fuel cell stack; Indicates the number of batteries connected in series in the battery pack; Indicates the number of batteries connected in series in the battery pack; Indicates the rated capacity of the battery; This indicates the battery voltage.

[0075] Based on a multi-objective constraint model, an improved constraint grouping multi-population evolution algorithm is used to set the population structure, resulting in a multi-population evolution framework. Then, dynamic relaxation and dormant activation mechanisms are combined to independently evolve and recombine the populations in the multi-population evolution framework, resulting in a recombined population.

[0076] Specifically, based on a multi-objective constraint model, an improved constraint grouping multi-population evolution algorithm is used to set the population structure, resulting in a multi-population evolution framework. Then, dynamic relaxation and dormant activation mechanisms are combined to independently evolve and recombine the populations within this framework, yielding the recombinated population, which includes:

[0077] The number of population structures is determined based on the number of constraints in the multi-objective constraint model, and a multi-population evolutionary framework including the main population, unconstrained population, and single-constrained subpopulation is established based on the number of population structures.

[0078] For each population in the multi-population evolution framework, the population individuals and states are initialized to obtain the activated population.

[0079] Specifically, for each population in the multi-population evolutionary framework, the population individuals and states are initialized, resulting in activated populations including:

[0080] In each population of a multi-population evolutionary framework, several candidate weight parameter individuals are randomly generated. These candidate weight parameter individuals are then substituted into an embedded model predictive control simulation environment based on a state-space prediction model to perform forward extrapolation and obtain the future state evolution trajectory.

[0081] By combining the future state evolution trajectory with the various constraints in the multi-objective constraint model, the corresponding constraint violation degree and objective function value are calculated.

[0082] Each population is initialized based on the constraint violation degree and the objective function value to obtain an initial population. The population states of all initial populations are marked as active to obtain the activated population.

[0083] After activation, the population is subjected to selection, crossover, and mutation operations to obtain offspring individuals. The fitness of all offspring individuals is evaluated using a dynamic relaxation factor to obtain the effective violation rate.

[0084] Specifically, the activated population undergoes selection, crossover, and mutation operations to obtain offspring individuals. The fitness of all offspring individuals is then evaluated using a dynamic relaxation factor to obtain the effective violation rate, which includes:

[0085] For each activated population, a preset number of candidate weight parameter individuals are randomly selected, and the winners are determined according to the fitness comparison criteria. The determined winners are used as the parents.

[0086] By performing simulated binary crossover and polynomial mutation operations on each pair of parent individuals, several offspring individuals are generated. All offspring individuals generated by the activated population are merged into a temporary offspring pool.

[0087] The original constraint violation degree is calculated for each individual in the temporary offspring pool, and a dynamic relaxation factor is introduced to correct the original constraint violation degree to obtain the effective violation degree.

[0088] Based on the effective violation degree and the objective function value, a population dormancy and activation mechanism based on the center point displacement and adaptive threshold is implemented for the population to which the offspring individuals belong. Combined with multi-dimensional population evaluation indicators, the population is reorganized to obtain the reorganized population.

[0089] Specifically, based on the effective violation degree and the objective function value, a population dormancy and activation mechanism based on centroid displacement and adaptive threshold is implemented for the population to which the offspring individuals belong. Combined with multi-dimensional population evaluation indicators, population recombination is performed, resulting in a recombined population including:

[0090] Based on the population type to which offspring individuals belong, the effective violation rates and objective function values ​​corresponding to each constraint in the multi-objective constraint model are non-dominatedly ranked to obtain the non-dominated ranking hierarchy:

[0091] Based on the non-dominated ranking hierarchy and crowding distance of the population, select a number of individuals from the parent individuals and the temporary offspring pool of each population to form the next generation population.

[0092] After each generation of the population ends, the displacement and adaptive threshold based on the current generation population center point are calculated for each population. Based on the comparison results of the displacement and the adaptive threshold, the population is subjected to iterative dormancy and activation operations to obtain the adjusted population.

[0093] The performance of each adjusted population is evaluated based on multi-dimensional population evaluation indicators. The performance evaluation results are then used to perform adaptive population merging and collaborative update operations based on dominance relationships and diversity enhancement on all adjusted populations to obtain the recombined population.

[0094] Specifically, step 2: For the multi-constraint, multi-objective problem of green electricity hydrogen production system, establish the objective function and constraints, including:

[0095] 1) For the CMOP problem of this system, the following objective function is established:

[0096] ;

[0097] In the formula, x corresponds to the decision variable vector, representing the sequence of power setpoints for all adjustable devices within a future forecast period; This represents the total number of inequality constraints. This indicates the total number of equality constraints. To represent inequality constraints, This indicates equality constraints.

[0098] ;

[0099] ;

[0100] In the formula, To maximize hydrogen production, It refers to the efficiency of the electrolytic cell. It is the power of the electrolytic cell at time t. To maximize the consumption of new energy sources, P cur Indicates the amount of wind power curtailed; Indicates the amount of wind power curtailed; This represents the total number of discrete time steps in the prediction time domain.

[0101] 2) Establish multiple inequality constraints for the green electricity hydrogen production system. It includes:

[0102] a. Power inequality constraint between grid connection and grid connection:

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] In the formula, Indicates the power exchange between power grids; This represents the electricity purchase status variable, which is 1 when electricity is purchased and 0 when no electricity is purchased. This indicates the amount of electricity purchased from the power grid; This represents the electricity sales status variable; it is 1 when electricity is purchased and 0 when no electricity is purchased. This indicates the amount of electricity sold to the power grid; This indicates the maximum value of the power exchange between power grids; This represents the maximum amount of electricity purchased from the power grid. This represents the maximum amount of electricity sold to the power grid.

[0109] b. Constraint on the power inequality between online and offline networks:

[0110] ;

[0111] In the formula, This indicates the maximum cumulative electricity sales limit to the grid (connected to the grid) within a scheduling cycle; This indicates the maximum cumulative electricity purchase (or removal) limit from the grid within a scheduling cycle.

[0112] c. Establish the controllable energy power constraint conditions as follows:

[0113] ;

[0114] In the formula, P elmin P fcmin P batmin These represent the minimum power of the electrolyzer, fuel cell, and battery, respectively, and P. elmax P fcmax P batmax These represent the maximum power of the electrolyzer, fuel cell, and battery, respectively.

[0115] d. The constraints for battery SOC and hydrogen storage tank SOH are as follows:

[0116] ;

[0117] In the formula, This represents the charging and discharging power of the battery at time t. This indicates the maximum charging and discharging power of the battery; This indicates the amplitude limit value when the battery is charged; This indicates the upper limit of the battery's state of charge.

[0118] e. The control increment constraint is:

[0119] ;

[0120] 3) Establish multiple equality constraints for the green electricity hydrogen production system. It includes:

[0121] a. Power balance constraints of each unit in the system:

[0122] ;

[0123] Net power balance constraint:

[0124] ;

[0125] In the formula, This represents the total power generation within the system at time t; This represents the total load demand power of the system at time t.

[0126] 4) Define constraint violation degree:

[0127] ;

[0128] ;

[0129] It is the sum of the severity of a control scheme x violating all constraints. =0 indicates that the solution is completely feasible. If a control solution exceeds the constraints, its corresponding... It is 1. Among them, It is a relaxation factor for handling equality constraints.

[0130] Step 3: Introduce an improved Constrained Grouping Multipopulation Evolutionary Algorithm (CGMEA). This algorithm, based on the original CGMEA, incorporates population activation and dormancy mechanisms as well as a dynamic relaxation function. Its algorithm flow includes five stages:

[0131] Phase 1: Population Initialization

[0132] 1.1) Set the population structure quantity:

[0133] Based on the number of constraints p+q, establish p+q+2 populations, each undertaking a different optimization task. The initial CGMEA populations include: a primary population, which performs global optimization considering all constraints; an unconstrained population, used purely for goal space exploration to prevent the algorithm from prematurely getting trapped in local optima; and p+q auxiliary populations, each specifically handling the satisfaction of a single set of constraints.

[0134] 1.2) Population Individual Initialization:

[0135] In each population, N individuals are randomly generated, and each individual represents a set of candidate Q and R parameters. For each set of candidate Q and R parameters carried by each individual, they are substituted into the embedded MPC simulation environment based on the model in step 1 for time-domain N... p Forward derivation. Obtain the future state evolution trajectory and control sequence Δu output by the state-space model under the candidate parameters. Subsequently, based on the obtained state trajectory and the constraints in step 2, calculate the constraint violation degree CV(x) and objective function value F(x) corresponding to this set of parameters.

[0136] 1.3) Setting initial parameter values:

[0137] Set parameters such as dynamic relaxation factor ε(0), target dimension, crossover mutation probability, and number of generations.

[0138] 1.4) Population state initialization:

[0139] All populations are marked as active. Active populations participate in offspring generation and environmental selection in each generation; dormant populations suspend offspring generation but still participate in environmental selection to receive individuals from other populations.

[0140] Phase 2: Independent Evolution of Populations

[0141] 2.1) Generate offspring:

[0142] For each active population, a predetermined number of individuals are randomly selected, and the winners are determined as parents based on fitness comparison criteria. Then, a simulated binary crossover (SBX) operation and a polynomial mutation operation are performed on each pair of parent individuals to generate N offspring individuals. Dormant populations do not generate offspring. The specific implementation steps are as follows:

[0143] A) Fitness comparison criterion: Each time, two candidate individuals x are randomly selected from the current population. A and x B The winner will be determined based on the following logic:

[0144] Logic 1: If x A All constraints are satisfied, and x B If not satisfied, then x A Win; otherwise x B Victory.

[0145] Logic 2: If neither of the constraints is satisfied, then it has a smaller CV. soft The individual with the highest value wins.

[0146] Logic 3: If both satisfy the constraints, then compare based on Pareto dominance: If x A Dominate x B Then x A The individual with the greater crowding distance wins; if neither can dominate the other, the individual with the greater crowding distance wins. Repeat this process to obtain a mating pool.

[0147] B) Non-uniform simulated binary crossover operation based on physical inertial mapping: Traditional algorithms use globally fixed crossover probabilities, which can easily lead to high-frequency abrupt changes in the control parameters of slow-dynamic devices. This invention innovatively extracts the physical intrinsic time constant τ of the controlled device (such as a battery or electrolytic cell) corresponding to the k-th dimension control weight of the individual parameter vector. k From the mating pool, according to the crossover probability p cWhen selecting a parent pair (x1, x2) to generate offspring (y1, y2), an independent adaptive crossover probability p is assigned to the k-th dimension decision variable. c,k :

[0148] ;

[0149] In the formula, As the baseline crossover probability, This is the sampling period for Model Predictive Control (MPC). If the crossover probability is satisfied... Then, perform the following SBX crossover on the k-th dimension:

[0150] ;

[0151] ;

[0152] Among them, the expansion factor The calculation depends on the random number u∈[0,1] and the distribution exponent η. c When u≤0.5, ;otherwise .

[0153] C) Polynomial mutation operation: The generated offspring are mutated according to the mutation probability p. m Perform a perturbation. The mutation logic for the k-th dimension variable is as follows:

[0154] ;

[0155] in, , These are the upper and lower bounds of the parameter. Disturbance factor. Given random numbers r∈[0,1] and the variation distribution exponent η m Decision: When r ≤ 0.5, ;otherwise .

[0156] 2.2) Fitness evaluation combined with dynamic relaxation:

[0157] All offspring individuals generated by the activated population are merged into a temporary offspring pool O. off Calculate the original constraint violation degree for each individual x in the pool: In the formula, C represents the total number of constraints; j represents the j-th constraint.

[0158] By introducing a dynamic relaxation factor ε(0), the original constraint violation degree CV(x) is modified into an effective violation degree:

[0159] ;

[0160] When CV(x) ≤ ε, the individual is regarded as a temporarily feasible solution and is treated equally with feasible solutions in non-dominated sorting; otherwise, the excess part is retained as the infeasibility measure.

[0161] 2.3) Multi-population differential constraint evaluation:

[0162] According to the population type to which the individual belongs, different constraint sets are used for non-dominated sorting: for the main population, the CV of all constraints soft (x) participates in the sorting to approximate the true feasible front; for the unconstrained population, all constraints are ignored and sorting is only based on the objective function values to provide unconstrained front information. For the auxiliary population, only the original violation degree of its corresponding constraint and the objective function are jointly sorted, focusing on the optimal solution under a single constraint.

[0163] 2.4) Environmental selection:

[0164] Each population P selects N individuals from its own parental individuals and the temporary offspring pool to form the next generation. The selection is based on the non-dominated sorting level and crowding distance corresponding to the population. Although the dormant population does not generate offspring, it still screens individuals from O off to update its own population.

[0165] 2.5) Population dormancy and activation:

[0166] After each generation, the center point and displacement of each population (including the dormant population) are calculated:

[0167] Center point of the current generation population: ; where N represents the total number of individuals in the population; represents the vector composed of the fitness values of the i-th individual on all objectives.

[0168] Displacement: .

[0169] Adaptive threshold: .

[0170] In the formula, represents the total number of objective functions of the multi-objective optimization problem; represents the specific evaluation value of the i-th individual on the j-th objective function.

[0171] For the activated population, if ΔC < CT and there are no temporarily feasible solutions in the population, that is, CV of all individuals soft (x) > 0, then the population is dormant; for the dormant population, if ΔC > CT, then the population is reactivated.

[0172] 2.6) Iterative loop: Repeat steps 2.1 - 2.4 until the preset maximum number of generations T1 in the first stage is reached.

[0173] Phase 3: Population Reorganization

[0174] 3.1) Subpopulation performance analysis and constrained grouping:

[0175] Performance evaluation is performed on each subpopulation: its non-dominated solution ratio, convergence distance from the population center to the global ideal point, and population activity (dormancy duration and activation count) are calculated as diversity indicators; and its activity is measured by combining dormancy duration and activation count. Based on the above indicators, the constraints corresponding to all subpopulations are divided into effective constraint groups (constraints where the subpopulation is still active or, although dormant, has high-quality non-dominated solutions and strong complementarity with the main population) and ineffective constraint groups (constraints that are dormant for a long time and have low-quality non-dominated solutions or are highly redundant with other populations). The multi-dimensional evaluation indicators and specific calculation steps are defined as follows:

[0176] A) The proportion of non-dominated solutions R ND : Calculate the percentage of individuals belonging to level 1 within the subpopulation relative to the population size N.

[0177] B) Convergence distance D c : Calculate the average Euclidean distance from all feasible solutions in the population to the global ideal point.

[0178] C) Target space coverage S cov : Calculate the volume of the hypercube formed by the extreme solutions in the population, and measure its diversity distribution in the target space.

[0179] D) Activity (Act) Measurement Mechanism: Combining dormant and active historical logs, a comprehensive activity calculation formula is defined:

[0180] ;

[0181] In the formula, T represents the cumulative number of generations that this subpopulation has been in a dormant state during its evolutionary process. active T represents the number of times a device transitions from a dormant to an active state. total This represents the current total algebra; ω1 to ω5 are the normalized weight coefficients. According to... Metric values, setting activity thresholds. If If the constraint is valid, then the constraint corresponding to that subpopulation is classified as a valid constraint group; otherwise, it is classified as an invalid constraint group.

[0182] 3.2) Population merging and competitive optimization:

[0183] If an invalid subpopulation's non-dominated solutions are completely dominated by the valid subpopulation, then its individuals are merged into the valid subpopulation; otherwise, they are merged into the main population. The main population and the unconstrained population each absorb some individuals from the best individuals of all subpopulations to enhance diversity. The unconstrained population is then mixed with the retained subpopulations, and N individuals are selected through non-dominated sorting to update both populations.

[0184] A hybrid strategy of fitness landscape co-analysis was used to co-evolve the recombined population to obtain the evolved population. A physical barrier perception memory map was constructed based on the evolved population, and population transition based on spectral clustering was performed based on the physical barrier perception memory map to obtain the Pareto front solution set of the weight parameters.

[0185] Specifically, a hybrid strategy of fitness-landscape co-evolution is used to co-evolve the recombined population, resulting in an evolved population. A physical barrier perception memory map is constructed based on this evolved population, and population transitions are performed using spectral clustering based on this map. The Pareto front solution set for the weighted parameters is obtained, including:

[0186] The dynamic relaxation factor of the recombined population is gradually tightened using an exponential decay strategy, and inter-population knowledge transfer and reference vector guidance are performed on several activated populations in the recombined population to obtain a population with information sharing.

[0187] A hybrid strategy based on fitness landscape synergy analysis is used to evaluate the local curvature of the population after information sharing, and to perform constrained activation discrimination on the population after information sharing to obtain the activation constraint boundary discrimination results.

[0188] Based on the local curvature and the results of the active constraint boundary discrimination, a topological orientation walk strategy along the effective constraint tangent plane is adopted, and combined with the exponential spatial damping function of physical safety margin, the population after information sharing is guided to slide along the constraint boundary to generate the evolved population.

[0189] A physical barrier perception memory map is constructed based on the new offspring individuals in all evolved populations. The physical barrier perception memory map is then subjected to spectral clustering and safe manifold partitioning based on the Laplace matrix to obtain safe subclusters. Trapped safe subclusters are subjected to cross-manifold centroid-oriented transitions to obtain the Pareto front solution set of the weight parameters.

[0190] Specifically, a physical barrier perception memory map is constructed based on the new offspring individuals in all evolved populations. This map is then subjected to spectral clustering based on the Laplace matrix and safe manifold partitioning to obtain safe subclusters. Trapped safe subclusters are then subjected to cross-manifold centroid-oriented transitions to obtain the Pareto front solution set for the weight parameters, including:

[0191] Non-dominated history solutions are extracted from the new offspring individuals in all evolved populations, and a vertex set is constructed based on the extraction results. A connectivity weight matrix is ​​set based on the Euclidean distance and physical barrier determination results.

[0192] A physical barrier perception memory graph is constructed based on the vertex set and connectivity weight matrix, and the degree matrix of the physical barrier perception memory graph is defined. The non-normalized Laplacian matrix of the physical barrier perception memory graph is calculated based on the degree matrix.

[0193] The non-normalized Laplacian matrix is ​​subjected to eigenvalue decomposition, and the eigenvectors corresponding to the first few smallest eigenvalues ​​are extracted based on the eigenvalue decomposition results to form a feature matrix. The row vectors of the feature matrix are clustered using a clustering algorithm to obtain several unconnected safe subclusters.

[0194] For the current new offspring individuals trapped in the trap, the original trapped sub-cluster is identified. Among the remaining safe sub-clusters, the sub-cluster with the best average target fitness is selected as the target escape sub-cluster. The current new offspring individuals are then subjected to a directional jump based on the centroid of the target sub-cluster to generate new escape offspring individuals.

[0195] Repeatedly perform co-evolution and diversity enhancement operations until the proportion of feasible solutions in the main population reaches a preset value and the change in the non-dominated front is less than a threshold or reaches the maximum number of generations for several consecutive generations, to obtain the final generation population. Extract all truly feasible and non-dominated individuals from all the final generation populations to form the Pareto front solution set with weight parameters.

[0196] Specifically, Phase 4: Co-evolution and Enhanced Diversity

[0197] 4.1) Dynamic relaxation factor gradually tightens:

[0198] An exponential decay strategy is used to gradually decrease ε(t+1) to 0:

[0199] ;

[0200] In the formula, This represents the exponential decay coefficient, used to control the tightening speed of the dynamic relaxation factor.

[0201] 4.2) Inter-population knowledge transfer and reference vector guidance:

[0202] In each generation, several active populations are randomly selected, and some individuals in their front solution sets are exchanged to achieve information sharing. At the same time, a set of uniformly distributed reference vectors are introduced into the main population and the unconstrained population to decompose the target space into multiple sub-regions. During environment selection, the nearest individual is retained in each sub-region to maintain the uniform distribution of the population across the entire front.

[0203] 4.3) Hybrid strategy evolution based on fitness-landscape synergy analysis:

[0204] To address the "infeasible solution trap" problem caused by the strong physical constraints of green electricity hydrogen production systems, a hybrid strategy evolution based on terrain perception is implemented:

[0205] A) Local curvature calculation: Every preset number of generations, extract the non-dominated solution set from the current generation of the active population. For individual x, use finite difference to approximate the objective function Hessian matrix H(x) at the adjacent solution x, and calculate the trace of this matrix, which is defined as the local curvature of the objective space location of the solution. .

[0206] B) Constraint Activation Determination: Extract the physical inequality constraints corresponding to individual x. .

[0207] Introducing the dynamic relaxation factor for the current iteration step If an individual x satisfies a certain inequality constraint (in If the boundary value is a very small positive number, then the individual is determined to have approached or touched the activation constraint boundary, and the specific constraint function evaluation value of the triggered activation state is recorded as . .

[0208] C) Topology-guided hybrid operator adaptive scheduling:

[0209] Smooth exploration mode: if local curvature This indicates that the population is currently in a relatively flat feasible region with a high physical safety margin. The standard SBX crossover and polynomial mutation steps in step 2.1 are then performed to facilitate a large-scale global exploration of the population.

[0210] Targeted wandering mode: If Furthermore, since the individual is at the activation constraint boundary, blind mutation at this point is highly likely to produce out-of-bounds and invalid solutions. Therefore, random mutation operations on this individual are forcibly frozen. A low-curvature directional walk operator based on topology analysis is then activated to generate a new offspring individual x. new :

[0211] ;

[0212] In the formula, To synthesize the gradient of the objective function, the walking direction is forced to always face the side that improves fitness by using the sign function sign and dot product operation; Let λ be a small Gaussian perturbation applied only within the null space of the activation constraints. Let λ be the adaptive damped walk step size based on the physical boundary margin. To prevent the algorithm from generating out-of-limit solutions due to excessively large fixed step sizes when sliding close to the physical constraint boundaries, an exponential spatial damping function considering the system's physical safety margin is introduced:

[0213] ;

[0214] In the formula, λ max The set maximum walking step length; D margin(x) represents the normalized safety margin (range [0,1]) of the nearest active physical constraint boundary to the current individual; D ref It is a preset baseline margin constant, and its mechanism of action is similar to that of the physical time constant in the cross operator. It is used to adjust the sensitive scale of the boundary damping response.

[0215] D) Trap escape mechanism based on multi-scale topological memory graph and spectral graph clustering:

[0216] When an individual is in a directional walking pattern, if the local curvature of the continuous m generations is... Continuously above the deadlock threshold Furthermore, the fitness increment is extremely small, indicating that individuals are trapped in an "infeasibility trap" composed of multiple physical hard constraints. To avoid the forgetting of evolutionary information caused by conventional random restarts, a multi-scale topological memory graph is constructed, and spectral graph clustering is used to guide the population's intelligent leap:

[0217] Construct a physical barrier perception memory graph G=(V,E) and a weight matrix W: Extract all currently surviving non-dominated historical solutions to form a vertex set. (N is the total number of historical solutions). The size is defined as... The connectivity weight matrix W. Matrix elements W ij It combines Euclidean distance and physical obstacle determination:

[0218] ;

[0219] In the formula, σ is a preset spatial scale constant that controls the width of the Gaussian kernel. Ω ij This is a penalty term for penalizing penetration of physical boundaries. The specific determination rule is: in solving x... i To x j S discrete sampling points are uniformly extracted from the high-dimensional connection line and substituted into the physical inequality constraints in step 2; if any constraint is broken, it is determined that there is a physical obstacle between the two points, and Ω is forced. ij =1; if all sampling points satisfy the physical constraints, then Ω ij =0.

[0220] Spectral clustering and secure manifold partitioning based on Laplacian matrix:

[0221] Define the degree matrix D of graph G as a diagonal matrix, with diagonal elements... .

[0222] Calculate the nonnormalized Laplacian matrix L of the memograph:

[0223] ;

[0224] Eigenvalue decomposition of the Laplacian matrix L ( The eigenvectors corresponding to the k smallest eigenvalues ​​are extracted to form a feature matrix. The row vectors of this feature matrix are then clustered using the K-means algorithm to divide the high-dimensional feasible region into k mutually disconnected safe sets. .

[0225] Warping: Directed transitions across manifold centroids

[0226] Identify the original trapped subcluster where the currently trapped individual resides. Among the remaining safe subclusters, select the subcluster with the best average target fitness as the target escape subcluster S. target Calculate the target subcluster S target spatial geometric centroid vector μ target Spatial distribution variance vector Var in each dimension target :

[0227] ;

[0228] ;

[0229] The ineffective blind mutation of trapped individuals is forcibly frozen, causing them to leap to the vicinity of the centroid of the target subcluster, and new escaped offspring individuals x are generated using their spatial variance. escape :

[0230] ;

[0231] In the formula, R is a random vector uniformly distributed in the interval [-1,1], and the symbol "." represents element-wise multiplication of the vector.

[0232] 4.4) Loop and convergence judgment:

[0233] Repeat steps 4.1-4.3 until the proportion of feasible solutions in the main population reaches 100% and the changes in the non-dominated frontier are less than the threshold or the maximum number of generations is reached.

[0234] Phase 5: Extracting the Optimal Solution and Updating MPC Weights

[0235] 5.1) Extracting the Pareto front:

[0236] From the final generation of all populations, extract all truly feasible and non-dominated individuals to form the Pareto front solution set. Each solution corresponds to a set of Q and R parameters.

[0237] Based on the superior-inferior solution distance method and the verification and evaluation algorithm of meteorological extreme errors, the optimal solution of the Pareto front solution set of the weight parameters is screened to obtain the optimal solution of the weight parameters. The optimal solution of the weight parameters is then smoothly weighted and updated to the model predictive controller to achieve coordinated control of green electricity to hydrogen production.

[0238] Specifically, based on the superior-inferior solution distance method and the verification and evaluation algorithm of meteorological extreme errors, the optimal solution of the Pareto front solution set of the weight parameters is screened to obtain the optimal solution of the weight parameters. The optimal solution of the weight parameters is then smoothly weighted and updated to the model predictive controller to achieve coordinated control of green electricity to hydrogen production, including:

[0239] The Pareto front solution set of the weight parameters is sorted and filtered using the superior-inferior solution distance method to obtain the candidate optimal parameter solution set;

[0240] Candidate optimal parameters are injected into the virtual model prediction and control environment, and the discrete meteorological prediction error statistics are reconstructed into a meteorological disturbance field using a verification and evaluation algorithm for meteorological extreme errors. The dynamic deviation trajectory under weather disturbances is then deduced through the meteorological disturbance field.

[0241] Specifically, the discrete meteorological forecast error statistics are reconstructed into a meteorological disturbance field using a verification and evaluation algorithm for meteorological extreme errors. The dynamic deviation trajectory under weather disturbances is then deduced from this meteorological disturbance field, including:

[0242] Based on the natural evolution law of wind and solar power output, a joint perturbation vector containing error components of wind power and photovoltaic power is constructed;

[0243] The joint perturbation vector is recursively derived using a first-order linear recursive equation to obtain a continuous spatiotemporal evolution perturbation sequence, and a meteorological perturbation field is generated based on the continuous spatiotemporal evolution perturbation sequence.

[0244] In a virtual model predictive control environment, a quadratic programming solution is performed on the continuous spatiotemporal evolution disturbance sequence in the meteorological disturbance field to obtain the dynamic deviation trajectory under weather disturbance.

[0245] The dynamic deviation trajectory is substituted into all physical constraints of the multi-objective constraint model for secondary safety verification. Based on the safety verification results, the candidate optimal parameter solution set is screened to obtain the optimal solution of the weight parameters.

[0246] By using a low-pass filtering strategy, the optimal solution of the weight parameters is smoothly updated to the inner-loop model predictive controller. The quadratic programming problem is then solved through the model predictive controller to obtain a coordinated control strategy for green electricity to hydrogen production under different constraints.

[0247] Specifically, 5.2) Selecting the optimal solution based on TOPSIS

[0248] Step 4: Construct a dual-loop cooperative control framework based on the combination of Constrained Grouping Multi-Population Evolutionary Algorithm (CGMEA) and Model Predictive Control (MPC), which includes:

[0249] 1) Before performing MPC rolling optimization, based on the discrete system state-space expression established in step 1, derive the prediction time-domain N. pThe multi-step prediction model within the model links the sequence of control variables Δu with the future state evolution x and output y. c Directly related.

[0250] Let the current time be k, and in the control time domain N c and prediction time domain N p Within, the state prediction equation for the j-th future step is iteratively expanded as follows:

[0251] ;

[0252] Combined with the controlled output matrix C defined in step 1 c The predicted value of the controlled output in the j-th step is:

[0253] ;

[0254] 2) Predict the output Substitute the values ​​and construct the objective function for MPC-based rolling optimization:

[0255] ;

[0256] ;

[0257] In the formula, y represents the prediction of time k+j from time k; ref Given a reference trajectory; Δu is the increment of the control variable; Np is the prediction time domain; Nc is the control time domain; matrix Q m q is the weighting matrix for the output error; i The weights of the state variables; matrix R m Let λ be the weighting matrix for the control increment. i (i=1, 2, 3) represent the weighting factors controlling the increment, for P net No control is applied; λ1, λ2, and λ3 represent the effects on the electrolytic cell power P, respectively. el Battery power P bat Fuel cell power P fc The degree of importance attached to it.

[0258] 3) Dynamic reliability verification of optimal Q and R parameters based on continuous spatiotemporal evolution perturbation field:

[0259] The optimal parameters (Q(k), R(k)) to be verified are injected into the virtual MPC environment, and the verification time domain N is set. v (where 1≤N) v ≤N p Perform a quadratic programming solution in a virtual environment to obtain the tentative state prediction trajectory x. v With control increment sequence Δu vBased on the verification and evaluation algorithm for extreme meteorological errors, the discrete meteorological forecast error statistics are reconstructed into a continuously evolving extreme meteorological disturbance field, thereby capturing the propagation inertia and spatial coupling effect of meteorological disturbances. This enables parameter verification from static optimality to dynamic reliability. The specific steps are as follows:

[0260] A) Constructing an extreme meteorological disturbance field with continuous spatiotemporal evolution:

[0261] Considering the natural evolution of wind and solar power output, a joint perturbation vector incorporating error components from both wind and solar power is constructed. Using a first-order linear recursive equation, a time-continuous sequence of extreme weather disturbances is generated: In the formula, The time lag coefficient matrix is ​​used to characterize the continuation effect of meteorological disturbance error at the previous time (t-1) on the current time (t), reflecting the smooth transition characteristics of gusts or cloud cover in time. This is a cross-coupling matrix of wind and solar variables, where the off-diagonal elements represent the maximum extreme value error of the input at the current moment. The degree of mutual influence between the two physical variables of wind power and photovoltaic power; This is the input vector for the current moment's limit prediction error based on historical statistics.

[0262] B) Deducing Dynamic Robust Boundaries: This involves generating a continuous spatiotemporal evolution perturbation sequence. Inject a virtual prediction model to calculate the dynamic deviation trajectory Δx that the system may reach under the worst operating conditions. v The formula for the deviation state in the j-th future step is: In the formula, Φ(m) represents the ideal predicted state without disturbance; Φ(m) is the net power disturbance state transition response coefficient derived based on the state-space system matrix in step 1.

[0263] C) Exceedance Diagnosis and Safety Assessment: Dynamic deviation from the trajectory Δx... v and control sequence Δu v Substitute the physical inequality constraints established in step 2 A second security check is performed.

[0264] If all constraints are satisfied This demonstrates that the parameter combination is not only optimal under static conditions, but its robust boundary can also completely enclose extreme weather disturbances, and the verification is successful. If any constraint is broken, the parameters are considered dynamically unstable.

[0265] D) Triggering directional fault tolerance rollback: If dynamic verification fails, the algorithm immediately removes the current parameter combination and selects the second-best parameter combination according to the TOPSIS score order to re-perform the above verification; if the maximum number of verifications is reached and all failures are made, the fault tolerance mechanism is triggered, and the safety weight parameters (Q(k-1), R(k-1)) of the previous cycle are directly used to ensure the absolute safety of the underlying control.

[0266] 4) Gradually update the weight matrix in MPC using a smooth transition strategy:

[0267] A first-order low-pass filtering strategy is used for smooth updating of the weight matrix. The specific mathematical expression for the update is:

[0268] ;

[0269] In the formula, Q(k) and R(k) are the actual operating weight matrices of the inner-loop MPC controller at the current moment, which are ultimately injected; Q(k-1) and R(k-1) are the weight matrices of the previous control cycle; α∈(0,1] is the smoothing transition factor. The value of α can be adaptively adjusted according to the current operating conditions of the system: when the power grid interaction fluctuation is small and the system is stable, α takes a larger value to speed up the adaptation of parameters to the new scenario; when the wind and solar power fluctuations are severe, α takes a smaller value to enhance the smoothness and anti-disturbance of the equipment control. and These represent the optimal state error weight matrix and the optimal control increment weight matrix, respectively, obtained and verified by the outer-loop CGMEA algorithm within the current optimization cycle.

[0270] Throughout the process, the improved CGMEA, acting as the outer-loop optimizer, runs at a slower cycle (e.g., every 30 minutes) to provide the inner-loop MPC controller with optimal Q and R parameters adapted to the latest scenarios. MPC performs rolling optimization at a faster cycle (e.g., every 1 minute) to achieve precise control of equipment such as electrolyzers and batteries.

[0271] like Figure 2 As shown, according to another embodiment of the present invention, a coordinated control system for green electricity-to-hydrogen production that satisfies multiple control objectives and multiple constraints is provided. The system includes:

[0272] Multi-objective constraint model construction module 1 is used to construct a multi-objective constraint model based on the tuning problem of the weight parameter matrix in the model predictive controller of the wind-solar-storage grid-connected hydrogen production system.

[0273] Multi-population evolution module 2 is used to set the population structure based on a multi-objective constraint model and an improved constraint grouping multi-population evolution algorithm to obtain a multi-population evolution framework. It also combines dynamic relaxation and dormancy activation mechanisms to independently evolve and recombine the populations in the multi-population evolution framework to obtain the recombined population.

[0274] The physical barrier perception memory map construction module 3 is used to perform co-evolution of the recombined population using a hybrid strategy of fitness landscape co-analysis to obtain the evolved population. Based on the evolved population, a physical barrier perception memory map is constructed, and population transition based on spectral clustering is performed based on the physical barrier perception memory map to obtain the Pareto front solution set of the weight parameters.

[0275] The optimal solution screening module 4 is used to screen the optimal solution of the Pareto front solution set of the weight parameters based on the superior-inferior solution distance method and the verification and evaluation algorithm of meteorological extreme error, obtain the optimal solution of the weight parameters, and smoothly weight and update the optimal solution of the weight parameters to the model predictive controller to realize the coordinated control of green electricity to hydrogen production.

[0276] In summary, this invention, based on a dual-timescale collaborative architecture, models the tuning problem of weight matrices Q and R in the MPC controller as a constrained multi-objective optimization problem. The optimization objectives include system performance indicators such as tracking error, control increment, and renewable energy absorption rate. Constraints cover physical limitations such as equipment output limits, ramp rate, and power balance. To efficiently solve this complex constrained optimization problem, a multi-population evolutionary framework is constructed, including a main population, an unconstrained population, and a single-constraint subpopulation. A dynamic relaxation function is introduced to loosen the constraint boundaries in the early stages of evolution, allowing individuals that slightly violate the constraints to compete and cross infeasible regions. Simultaneously, a population dormancy activation mechanism based on center point displacement and adaptive thresholds is designed to dynamically freeze stagnant populations to concentrate computational resources. Addressing the order-of-magnitude timescale differences in the electro-hydrogen system equipment, a non-uniform crossover operator based on physical inertial mapping is used to extract the inherent physical time constants of the underlying equipment, achieving decoupled control weight reorganization and effectively isolating the fatigue impact of high-frequency disturbances on slow-response equipment such as electrolyzers.

[0277] In the later stages of evolution, to overcome the "infeasibility trap" caused by high-dimensional physical constraints, this invention introduces a hybrid strategy based on fitness-based landscape collaborative analysis, building upon the identification of effective constraints through constraint grouping. This involves freezing conventional random crossover mutations within a narrow feasible region closely adhering to the physical limits by evaluating local curvature online, initiating a topologically oriented walk strategy along the effective constraint tangent plane, and supplementing this with an exponential spatial damping step size strongly bound to the system's physical constraint margin to guide the population to slide along the constraint boundary. Based on topological analysis, a multi-scale topological memory map is constructed, and spectral clustering is used to intelligently guide the population to efficiently escape the infeasibility trap, avoiding random restarts and ensuring the continuity of the Pareto front and the completeness of constraint boundary exploration. The TOPSIS method is used to select the optimal Q and R parameters. Before parameter updates, dynamic robustness verification and failure mode diagnosis, including meteorological extreme errors, are implemented to reconstruct discrete meteorological prediction error statistics into a continuously evolving extreme meteorological perturbation field, rather than a simple superposition of extreme values. Then, by quantifying the extreme value perturbation boundary and deriving the dynamic robust boundary, the static optimality of the parameters is ensured, thereby capturing the propagation inertia and spatial coupling effect of meteorological disturbances. After successful verification, the parameters are smoothly weighted and updated to the MPC controller after short-time domain simulation verification, realizing the parameter verification from static optimality to dynamic reliability. In the field of multi-objective optimization and automatic control, the specific steps include: First, establishing a multi-objective constraint model for MPC weight optimization; second, designing a multi-population evolutionary architecture and introducing dynamic relaxation and dormancy activation mechanisms; then, adopting a hybrid evolutionary strategy based on fitness landscape collaborative analysis, achieving hybrid strategy evolution through topological directional walk, spatial damping step size, and spectral clustering escape to obtain the Pareto front; finally, based on the dynamic robust derivation of the spatiotemporal meteorological evolution field and TOPSIS (i.e., the best-in-best solution distance method), parameter optimization and smooth update are achieved.

[0278] Furthermore, the present invention also provides an electronic device. For example... Figure 3 The diagram illustrates the hardware operating environment of an electronic device, which may include: a processor (e.g., CPU), memory, a user interface, a network interface, and a communication bus. The communication bus is used to enable communication between components. The user interface may include a display screen and an input unit such as a keyboard; optionally, the user interface may also include a standard wired interface or a wireless interface. The network interface may optionally include a standard wired interface or a wireless interface. The memory may be high-speed RAM or stable non-volatile memory, such as disk storage. Alternatively, the memory may be a storage device independent of the aforementioned processor.

[0279] Those skilled in the art will understand that Figure 3 The electronic devices shown do not constitute a limitation on electronic devices and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0280] like Figure 3 As shown, a memory, as a type of computer storage medium, may include an operating system, a network communication module, a user interface module, and device management programs. The operating system is a program that manages and controls the hardware and software resources of electronic devices, supporting the operation of electronic devices and other software or programs. Figure 3 In the electronic device shown, the user interface is mainly used to connect to the terminal and communicate with the terminal, such as receiving user signaling data sent by the terminal; the network interface is mainly used to communicate with the backend server; the processor can be used to call the program stored in the memory and execute the steps of the method or system described above.

[0281] Furthermore, the present invention also proposes a computer-readable storage medium storing a device management program, which, when executed by a processor, implements the steps of the method or system described above.

[0282] The specific embodiments of the computer-readable storage medium of the present invention are basically the same as those of the above-described methods or systems, and will not be repeated here. Furthermore, to achieve the above objectives, the present invention also provides a computer program product, comprising: a computer program, which, when executed by a processor, implements the steps of the methods or systems described above.

[0283] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0284] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A coordinated control method for green electricity-to-hydrogen production that satisfies multiple control objectives and constraints, characterized in that, The method includes: Based on the tuning problem of the weight parameter matrix in the model predictor of a wind-solar-storage grid-connected hydrogen production system, a multi-objective constrained model is constructed. Based on the multi-objective constraint model, an improved constraint grouping multi-population evolution algorithm is used to set the population structure, resulting in a multi-population evolution framework. The dynamic relaxation and dormancy activation mechanism is combined to independently evolve and recombine the populations in the multi-population evolution framework, resulting in a recombined population. A hybrid strategy of fitness landscape co-analysis was used to co-evolve the recombined population to obtain the evolved population. A physical barrier perception memory map was constructed based on the evolved population, and population transition based on spectral clustering was performed based on the physical barrier perception memory map to obtain the Pareto front solution set of the weight parameters. Based on the superior-inferior solution distance method and the verification and evaluation algorithm of meteorological extreme errors, the optimal solution of the Pareto front solution set of the weight parameters is screened to obtain the optimal solution of the weight parameters. The optimal solution of the weight parameters is then smoothly weighted and updated to the model predictive controller to achieve coordinated control of green electricity to hydrogen production.

2. The coordinated control method for green electricity-to-hydrogen production satisfying multiple control objectives and multiple constraints as described in claim 1, characterized in that, The process of tuning the weight parameter matrix in the model predictive controller of the wind-solar-storage grid-connected hydrogen production system, and constructing a multi-objective constrained model, includes: Establish the power balance equation of the wind-solar-storage grid-connected hydrogen production system, as well as the power balance equation of net power and controllable units, and construct the continuous-time state-space equation based on the dual-time-scale coordinated control architecture. Based on the forward Euler method, the continuous-time state-space equation is discretized to obtain the state-space prediction model of the wind-solar-storage-grid-hydrogen production system. Based on multiple constraints, boundary restrictions are imposed on the state variables and constraint output variables in the state-space prediction model, and a multi-objective constraint model containing multiple control objectives is constructed.

3. The coordinated control method for green electricity-to-hydrogen production satisfying multiple control objectives and multiple constraints as described in claim 1, characterized in that, The multi-objective constraint model utilizes an improved constraint grouping multi-population evolution algorithm to define the population structure, resulting in a multi-population evolution framework. Furthermore, dynamic relaxation and dormant activation mechanisms are combined to independently evolve and recombine the populations within this framework, yielding a recombined population including: The number of population structures is determined based on the number of constraints in the multi-objective constraint model, and a multi-population evolutionary framework including the main population, unconstrained population, and single-constrained subpopulation is established based on the number of population structures. For each population in the multi-population evolutionary framework, the population individuals and states are initialized to obtain the activated population; After activation, the population is subjected to selection, crossover and mutation operations to obtain offspring individuals. The fitness of all offspring individuals is evaluated by combining dynamic relaxation factors to obtain the effective violation degree. Based on the effective violation degree and the objective function value, a population dormancy and activation mechanism based on the center point displacement and adaptive threshold is implemented for the population to which the offspring individuals belong. Combined with multi-dimensional population evaluation indicators, the population is reorganized to obtain the reorganized population.

4. The coordinated control method for green electricity-to-hydrogen production satisfying multiple control objectives and multiple constraints as described in claim 3, characterized in that, The process of initializing the individuals and states of each population in the multi-population evolutionary framework to obtain the activated population includes: In each population of a multi-population evolutionary framework, several candidate weight parameter individuals are randomly generated. These candidate weight parameter individuals are then substituted into an embedded model predictive control simulation environment based on a state-space prediction model to perform forward extrapolation and obtain the future state evolution trajectory. By combining the future state evolution trajectory with the various constraints in the multi-objective constraint model, the corresponding constraint violation degree and objective function value are calculated. Each population is initialized based on the constraint violation degree and the objective function value to obtain an initial population. The population states of all initial populations are marked as active to obtain the activated population.

5. The coordinated control method for green electricity-to-hydrogen production satisfying multiple control objectives and multiple constraints as described in claim 3, characterized in that, The activated population undergoes selection, crossover, and mutation operations to obtain offspring individuals. The fitness of all offspring individuals is then evaluated using a dynamic relaxation factor to obtain the effective violation score, which includes: For each activated population, a preset number of candidate weight parameter individuals are randomly selected, and the winners are determined according to the fitness comparison criteria. The determined winners are used as the parents. By performing simulated binary crossover and polynomial mutation operations on each pair of parent individuals, several offspring individuals are generated. All offspring individuals generated by the activated population are merged into a temporary offspring pool. The original constraint violation degree is calculated for each individual in the temporary offspring pool, and a dynamic relaxation factor is introduced to correct the original constraint violation degree to obtain the effective violation degree.

6. The coordinated control method for green electricity-to-hydrogen production satisfying multiple control objectives and multiple constraints as described in claim 3, characterized in that, The process involves implementing a population dormancy and activation mechanism based on centroid displacement and adaptive thresholds for the population to which offspring individuals belong, according to the effective violation degree and objective function value. This is combined with multi-dimensional population evaluation indicators for population recombination, resulting in a recombined population including: Based on the population type to which offspring individuals belong, the effective violation rates and objective function values ​​corresponding to each constraint in the multi-objective constraint model are non-dominatedly ranked to obtain the non-dominated ranking hierarchy: Based on the non-dominated ranking hierarchy and crowding distance of the population, select a number of individuals from the parent individuals and the temporary offspring pool of each population to form the next generation population. After each generation of the population ends, the displacement and adaptive threshold based on the current generation population center point are calculated for each population. Based on the comparison results of the displacement and the adaptive threshold, the population is subjected to iterative dormancy and activation operations to obtain the adjusted population. The performance of each adjusted population is evaluated based on multi-dimensional population evaluation indicators. The performance evaluation results are then used to perform adaptive population merging and collaborative update operations based on dominance relationships and diversity enhancement on all adjusted populations to obtain the recombined population.

7. The coordinated control method for green electricity-to-hydrogen production satisfying multiple control objectives and multiple constraints as described in claim 1, characterized in that, The hybrid strategy utilizing fitness-landscape co-evolution analysis is used to co-evolve the recombined population, resulting in an evolved population. A physical barrier perception memory map is constructed based on this evolved population, and population transitions based on spectral clustering are performed using this map to obtain the Pareto front solution set for the weighted parameters, including: The dynamic relaxation factor of the recombined population is gradually tightened using an exponential decay strategy, and inter-population knowledge transfer and reference vector guidance are performed on several activated populations in the recombined population to obtain a population with information sharing. A hybrid strategy based on fitness landscape synergy analysis is used to evaluate the local curvature of the population after information sharing, and to perform constrained activation discrimination on the population after information sharing to obtain the activation constraint boundary discrimination results. Based on the local curvature and the results of the active constraint boundary discrimination, a topological orientation walk strategy along the effective constraint tangent plane is adopted, and combined with the exponential spatial damping function of physical safety margin, the population after information sharing is guided to slide along the constraint boundary to generate the evolved population. A physical barrier perception memory map is constructed based on the new offspring individuals in all evolved populations. The physical barrier perception memory map is then subjected to spectral clustering and safe manifold partitioning based on the Laplace matrix to obtain safe subclusters. Trapped safe subclusters are subjected to cross-manifold centroid-oriented transitions to obtain the Pareto front solution set of the weight parameters.

8. The coordinated control method for green electricity hydrogen production satisfying multiple control objectives and multiple constraints as described in claim 7, characterized in that, The physical barrier perception memory map is constructed based on the new offspring individuals from all evolved populations. Then, spectral clustering and safe manifold partitioning based on the Laplace matrix are performed on the physical barrier perception memory map to obtain safe subclusters. Trapped safe subclusters are subjected to cross-manifold centroid-oriented transitions to obtain the Pareto front solution set of the weight parameters, including: Non-dominated history solutions are extracted from the new offspring individuals in all evolved populations, and a vertex set is constructed based on the extraction results. A connectivity weight matrix is ​​set based on the Euclidean distance and physical barrier determination results. A physical barrier perception memory graph is constructed based on the vertex set and connectivity weight matrix, and the degree matrix of the physical barrier perception memory graph is defined. The non-normalized Laplacian matrix of the physical barrier perception memory graph is calculated based on the degree matrix. The non-normalized Laplacian matrix is ​​subjected to eigenvalue decomposition, and the eigenvectors corresponding to the first few smallest eigenvalues ​​are extracted based on the eigenvalue decomposition results to form a feature matrix. The row vectors of the feature matrix are clustered using a clustering algorithm to obtain several unconnected safe subclusters. For the current new offspring individuals trapped in the trap, the original trapped sub-cluster is identified. Among the remaining safe sub-clusters, the sub-cluster with the best average target fitness is selected as the target escape sub-cluster. The current new offspring individuals are then subjected to a directional jump based on the centroid of the target sub-cluster to generate new escape offspring individuals. Repeatedly perform co-evolution and diversity enhancement operations until the proportion of feasible solutions in the main population reaches a preset value and the change in the non-dominated front is less than a threshold or reaches the maximum number of generations for several consecutive generations, to obtain the final generation population. Extract all truly feasible and non-dominated individuals from all the final generation populations to form the Pareto front solution set with weight parameters.

9. The coordinated control method for green electricity-to-hydrogen production satisfying multiple control objectives and multiple constraints as described in claim 1, characterized in that, The verification and evaluation algorithm based on the superior-inferior solution distance method and meteorological extreme error filters the optimal solution of the Pareto front solution set of the weight parameters to obtain the optimal solution of the weight parameters, and smoothly and weightedly updates the optimal solution of the weight parameters to the model predictive controller to realize the coordinated control of green electricity to hydrogen production, including: The Pareto front solution set of the weight parameters is sorted and filtered using the superior-inferior solution distance method to obtain the candidate optimal parameter solution set; Candidate optimal parameters are injected into the virtual model prediction and control environment, and the discrete meteorological prediction error statistics are reconstructed into a meteorological disturbance field using the verification and evaluation algorithm of meteorological extreme errors. The dynamic deviation trajectory under weather disturbance is then deduced through the meteorological disturbance field. The dynamic deviation trajectory is substituted into all physical constraints of the multi-objective constraint model for secondary safety verification. Based on the safety verification results, the candidate optimal parameter solution set is screened to obtain the optimal solution of the weight parameters. By using a low-pass filtering strategy, the optimal solution of the weight parameters is smoothly updated to the inner-loop model predictive controller. The quadratic programming problem is then solved through the model predictive controller to obtain a coordinated control strategy for green electricity to hydrogen production under different constraints.

10. A coordinated control method for green electricity-to-hydrogen production satisfying multiple control objectives and multiple constraints as described in claim 9, characterized in that, The verification and evaluation algorithm utilizing meteorological extreme errors reconstructs discrete meteorological forecast error statistics into a meteorological disturbance field. The dynamic deviation trajectory under weather disturbances is then deduced from this meteorological disturbance field, including: Based on the natural evolution law of wind and solar power output, a joint perturbation vector containing error components of wind power and photovoltaic power is constructed; The joint perturbation vector is recursively derived using a first-order linear recursive equation to obtain a continuous spatiotemporal evolution perturbation sequence, and a meteorological perturbation field is generated based on the continuous spatiotemporal evolution perturbation sequence. In a virtual model predictive control environment, a quadratic programming solution is performed on the continuous spatiotemporal evolution disturbance sequence in the meteorological disturbance field to obtain the dynamic deviation trajectory under weather disturbance.

Images