Self-adaptive multi-objective evolution method for optimizing capacity of off-grid wind-solar hydrogen production and energy storage system

Abstract

A self-adaptive multi-objective evolution method for optimizing the capacity of an off-grid wind and light hydrogen production and energy storage system comprises the following steps: firstly, according to an energy storage system model, taking unit hydrogen production cost, wind and light abandoning penalty and load shedding penalty as minimization objectives; secondly, the capacity of each component of the wind-solar power generation unit, the power storage unit and the hydrogen and oxygen storage unit is used as a decision variable, and the system is set to conform to the actual constraint; thirdly, initializing a population; individuals in the population are initialized according to the constructed search space, so that the individuals meet constraint conditions; then adaptively selecting a search strategy; the strategy adaptive network selects a proper search strategy based on individual decision variables and objective function values, and evolves a new sub-population; and fusing the parent population and the child population to calculate the non-dominated level and the congestion degree of all individuals, determining parameters of the reward updating strategy adaptive network according to the non-dominated level, and screening out a new population. According to the method, the randomness and blindness of algorithm search can be reduced, and the search efficiency of the algorithm is enhanced.

Description

Technical Field

[0001] The present invention relates to a multi-objective optimization method, in particular to an adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen production energy storage system, belonging to the technical field of energy storage system capacity optimization. Background Art

[0002] As global energy is further advanced, the proportion of renewable energy such as wind power and photovoltaics in the power grid continues to increase. Building a new power system with renewable energy as the core has become an important direction of the modern energy system. By the end of 2023, China's wind power and photovoltaic installed capacity reached 440 million kilowatts and 610 million kilowatts, respectively, ranking first in the world. However, wind and solar power generation has significant volatility and intermittency, which makes the power quality unstable and difficult to continuously and reliably supply power to the load. In order to ensure the stable operation of the power system, wind and solar abandonment occur from time to time, resulting in a large amount of waste of renewable energy and also causing great economic losses.

[0003] In response to these problems, hydrogen energy stands out as a clean and efficient energy source. The production of hydrogen through electrolysis of water using renewable energy (green electricity hydrogen production) can not only significantly reduce the environmental pollution caused by traditional hydrogen production (grey hydrogen), but also efficiently convert the excess electricity from wind and solar power generation into hydrogen energy, thereby improving the utilization rate of renewable energy. The green electricity hydrogen production method absorbs surplus electricity, avoids the occurrence of wind and solar power abandonment, and realizes the deep consumption of renewable energy and the full utilization of resources.

[0004] In the process of green electricity hydrogen production, off-grid energy storage systems play a key role. Off-grid energy storage systems can not only effectively cope with the intermittent and volatile nature of wind and solar power generation, ensure the stability of power supply through energy storage scheduling, but also use surplus electricity for hydrogen production, forming an efficient energy system of electricity-hydrogen complementarity. The independence of off-grid energy storage systems makes them particularly suitable for remote areas and scenarios that are not covered by the power grid, greatly improving energy utilization efficiency and reducing dependence on the main power grid.

[0005] Although off-grid energy storage systems have many advantages, the capacity optimization configuration becomes more complicated due to the close connection between various power generation, hydrogen production, power storage and hydrogen storage units in the off-grid mode. Excessive or small energy storage capacity may have an adverse effect on the economy and performance of the system. If the capacity of the energy storage system is too small, it will not be able to effectively absorb renewable energy, nor can it meet the load demand of the electrolyzer, which will lead to the occurrence of wind and solar power abandonment, increased load shedding penalties and a decrease in hydrogen production. If the capacity is too large, it will lead to an increase in system investment and maintenance costs, thus affecting economic efficiency. Since multiple conflicting objectives need to be taken into account, the optimization process must adopt a multi-objective optimization method to simultaneously consider investment, operating costs, wind and solar power abandonment penalties and maximum output. The wind and solar power generation energy storage system in the off-grid mode involves more units and the model is more complex. Traditional population-based algorithms usually use a single search strategy, which is easy to fall into the local optimum and difficult to quickly find the global optimal solution. Summary of the invention

[0006] The purpose of the present invention is to provide an adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen energy storage system. By adaptively selecting appropriate search strategies for individuals at different stages of evolution, a set of Pareto optimal solutions can be searched for users more quickly, so that users can formulate corresponding energy storage system capacity configuration plans according to their preferences. This can reduce the randomness and blindness of the algorithm search and enhance the algorithm's search efficiency.

[0007] In order to achieve the above object, the present invention provides an adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen energy storage system, comprising the following steps:

[0008] S1: Construct the objective function; according to the energy storage system model, take the unit hydrogen production cost, wind and solar power abandonment penalty and load shedding penalty as the minimization objectives;

[0009] S2: Construct a search space; take the capacity of each component of the wind and solar power generation unit, the power storage unit, and the hydrogen and oxygen storage unit as decision variables, and set constraints that meet the actual needs of the system;

[0010] S3: Initialize the population; according to the individuals in the initial population of the search space constructed in step S2, make them meet the constraints;

[0011] S4: Adaptive selection of search strategy; the strategy adaptive network selects the appropriate search strategy based on the individual decision variables and objective function values, and evolves a new sub-population; the population iteration process includes building a strategy pool, selecting methods, and a strategy adaptive network;

[0012] S5: Update the population and strategy adaptation network; fuse the parent and child populations to calculate the non-dominated level and crowding degree of all individuals, determine the parameters of the reward update strategy adaptation network based on the non-dominated level and screen out the new population.

[0013] The energy storage system model of the present invention includes structures such as a power generation unit, a power storage unit, a hydrogen and oxygen storage unit, and an electrolyzer; wherein the power generation unit is composed of a wind turbine and a photovoltaic generator, and the power generation unit converts wind energy and light energy into electrical energy to supply the electrolyzer with hydrogen and oxygen production, and then stores it in the hydrogen and oxygen storage unit; when the generated electricity is in surplus, the system will detect whether the capacity of the power storage unit is full, and if the power storage unit still has capacity, the power storage unit composed of a battery and a supercapacitor will start to store electricity, and if the capacity of the power storage unit is full, energy waste will occur, resulting in wind and light abandonment; the electrolyzer usually operates at rated power, and generates hydrogen and oxygen by electrolyzing water, and these hydrogen and oxygen are stored in hydrogen storage tanks and oxygen storage tanks respectively. If the sum of the discharges of the power generation unit and the power storage unit cannot meet the rated power of the electrolyzer, in order to maintain the stability of the system, the electrolyzer will cut load and operate, which will cause a decrease in the output of hydrogen and oxygen, affecting the economy of the system.

[0014] The objective function constructed in step S1 of the present invention is as follows:

[0015]

[0016] Where: C H is the annual unit hydrogen production cost;

[0017] C I is the annual investment cost;

[0018] C M is the annual maintenance cost;

[0019] m H is the mass of hydrogen per year;

[0020] C WP Punishment for abandoning the wind and light for the year;

[0021] C L Penalty for annual load shedding;

[0022] In formula (1)

[0023]

[0024] Where: α is the depreciation rate;

[0025] y is the useful life;

[0026] N W 、N P 、N B 、N S、N EB 、N H 、N O They are the capacities of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks;

[0027] are the investment costs per unit capacity of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks;

[0028] In formula (1),

[0029] Where: They are the annual maintenance costs per unit capacity of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks;

[0030] In formula (1),

[0031] Where: δ is the hydrogen production rate of the electrolyzer per unit capacity per unit time per unit power;

[0032] In formula (1),

[0033] in

[0034] Where: β is the penalty coefficient for wind and solar power abandonment;

[0035] It is the abandoned wind and solar power;

[0036] is the hourly wind power generation;

[0037] is the hourly photovoltaic power generation;

[0038] P EBmax is the rated power of the electrolyzer;

[0039] The hourly discharge power of the battery;

[0040] is the hourly discharge power of the supercapacitor;

[0041] In formula (1),

[0042]

[0043] Where: γ is the load shedding penalty coefficient;

[0044] The hourly operating power of the electrolyzer.

[0045] The decision variables in step S2 of the present invention are the capacities of the wind turbine, photovoltaic generator, battery, supercapacitor, electrolyzer, hydrogen storage tank and oxygen storage tank, and the power balance constraint is:

[0046]

[0047] Where: is the charging power of the battery;

[0048] is the charging power of the supercapacitor;

[0049] Range constraints: All variables involved above are non-negative and have upper and lower limit constraints.

[0050] The initial population Pop in step S3 of the present invention is composed of N individuals X, each individual X has n variables x, and the initialization method of each variable x of individual X is as follows:

[0051] x=x min +rand nor (x max -x min ) (8)

[0052] Where: x max and x min are the upper and lower bounds of the variable;

[0053] rand nor ∈[0,1] is a random number between 0 and 1 that conforms to a uniform distribution.

[0054] The strategy pool constructed in step S4 of the present invention is a new search strategy designed based on the existing efficient search strategy. The strategy pool is composed of formulas (11), (12), (13), (14), and (15). The specific formula is as follows:

[0055]

[0056] Where: x r1 , x r2 and x r3 is a randomly selected parent solution;

[0057] x best is the global optimal solution in the current population. In multi-objective optimization, it is the solution in the first non-dominated layer.

[0058] and CR are predefined parameters;

[0059] r is a random number sampled from [0,1] with uniform distribution;

[0060]

[0061] Where: μ is a uniformly distributed random value sampled in [0,1];

[0062] η is a predefined parameter;

[0063]

[0064] The strategy adaptation network is a multi-layer fully connected network, whose input is the position state of each individual in the decision space and the target space, and whose output is the probability of selecting different search strategies.

[0065] The specific steps of step S5 of the present invention are as follows:

[0066] S51: After the fusion of the parent and child populations, it is necessary to calculate the non-dominated level and crowding degree of all individuals, and use this as a basis to update the new population. In the optimization of energy storage system capacity configuration, since multiple objectives are involved, the new population cannot be screened by relying solely on the fitness value as in single-objective optimization. Therefore, the dominance relationship between all solutions must be calculated, and the solutions must be sorted by non-domination according to the dominance relationship, and the solutions with the highest non-dominated level must be selected to enter the new population. In order to ensure the diversity of the population, it is also necessary to evaluate the crowding degree of the solution in the target space, retain those solutions with relatively sparse distribution, and select some solutions from dense areas to join the new population.

[0067] The definition of dominance relation is as follows: given two solutions A and B, if A is at least as good as B in all objectives and A is better than B in at least one objective, then solution A dominates solution B; non-dominated level: solutions that are not dominated by other solutions are called the first non-dominated layer (level 1). After removing the first layer, find solutions that are not dominated by the remaining solutions to form the second non-dominated layer (level 2). The calculation formula of crowding is as follows:

[0068]

[0069] Where: f i (A next ) and f i (A prev ) are the target values ​​of the adjacent solutions of solution A on target i;

[0070] f i (max) and f i (min) is the maximum and minimum value of all solutions on target i;

[0071] S52: Determine the reward returns of different search strategies based on the dominance relationship and update the parameters of the strategy adaptive network. The main idea of ​​updating the parameters is as follows: If a strategy is good, then for all states S, the expected value of the state should be large. Therefore, the objective optimization function of the strategy adaptive network is:

[0072]

[0073] Where: s is the position of the decision space and the target space;

[0074] p is a search strategy;

[0075] θ is the parameter of the network;

[0076] π θ (p|s) represents the probability of taking action p when the individual is in state s when the parameter of the strategy adaptation network is θ;

[0077] r(s,p) represents the reward obtained by adopting strategy p in state s;

[0078] By updating the parameters θ of the policy adaptation network, the value of J(θ) becomes larger and larger, which means that the policy adaptation network becomes stronger and stronger. Gradient ascent can be used to solve the maximization problem.

[0079] Let the parameters of the current strategy adaptation network be θ now , then the new parameter θ new Updated via:

[0080]

[0081] Where: σ is the learning rate;

[0082] is the policy gradient;

[0083] Reward is the focus of gradient update. Taking Formula 8 as an example, the reward design after the search strategy is executed is as follows:

[0084]

[0085] Where: Indicates a dominant relationship. Indicates that x is better than x in at least one objective function r1 ;

[0086] The relationship between the rewards is as follows R4<0 <R1<R2<R3。

[0087] Compared with the prior art, the present invention selects appropriate search strategies for individual adaptation at different stages of evolution, and searches for a set of Pareto optimal solutions for users more quickly, so that users can formulate corresponding energy storage system capacity configuration plans according to their preferences; the strategy pool designed by the present invention introduces gradient information on the basis of several classic differential evolution search strategies with different characteristics and simulated binary crossover operators, and uses gradients to guide the search of the algorithm, which reduces the randomness and blindness of the algorithm search to a certain extent and enhances the search efficiency of the algorithm; the strategy adaptive network designed by the present invention can judge whether the individual tends to converge or explore according to the different positions of the individual in the decision space and the target space, and reflects it in the form of probability to the search strategies with different characteristics. This method solves the problem of the singularity of the search strategy, has good portability, and can be transplanted to other multi-peak optimization problems; the strategy adaptive network designed by the present invention does not directly use the strategy with the highest probability for search, but uses the ε greedy strategy and roulette method to increase the early exploration of the adaptive network, and can also balance exploration and utilization in the middle and late stages to help find more strategies, reducing the risk of the network falling into a local optimal solution. The gradient update of the policy network does not depend on other indicators. The reward during policy update is determined only based on the Pareto dominance relationship, which makes the adaptive network highly portable and applicable to other multi-objective evolutionary optimization methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is an operation block diagram of the wind-solar power generation hydrogen storage system of the present invention;

[0089] Figure 2 It is the overall flow chart of the adaptive multi-objective evolutionary algorithm in the present invention;

[0090] Figure 3 It is a flow chart of the population cycle iteration of the adaptive multi-objective evolutionary algorithm in the present invention;

[0091] Figure 4 Schematic diagram of the gradient-guided search strategy in the present invention;

[0092] Figure 5 It is a comparison chart of HV index between the evolutionary method of the present invention and the classic NSGAII algorithm;

[0093] Figure 6 It is the Pareto optimal solution set searched out by the embodiment of the present invention. DETAILED DESCRIPTION

[0094] The present invention will be further described below in conjunction with the accompanying drawings.

[0095] like Figure 1-Figure 4As shown, an adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen production and storage system, the energy storage system model of the present invention includes structures such as a power generation unit, a power storage unit, a hydrogen and oxygen storage unit, and an electrolyzer; wherein the power generation unit is composed of a wind turbine and a photovoltaic generator, the power generation unit converts wind energy and light energy into electrical energy to supply the electrolyzer with hydrogen and oxygen, and then stores it in the hydrogen and oxygen storage unit; when the generated electricity is surplus, the system will detect whether the capacity of the power storage unit is full, if the power storage unit still has capacity, the power storage unit composed of a battery and a supercapacitor starts to store electricity, if the capacity of the power storage unit is full, energy waste will occur, resulting in wind and solar abandonment; the electrolyzer usually operates at rated power, and generates hydrogen and oxygen by electrolyzing water, and these hydrogen and oxygen are stored in hydrogen storage tanks and oxygen storage tanks respectively, if the sum of the discharges of the power generation unit and the power storage unit cannot meet the rated power of the electrolyzer, in order to maintain the stability of the system, the electrolyzer will cut load and operate, which will cause a decrease in the production of hydrogen and oxygen, affecting the economy of the system.

[0096] The adaptive multi-objective evolutionary algorithm of the present invention comprises the following steps:

[0097] S1: Construct an objective function; according to the energy storage system model, the unit hydrogen production cost, wind and solar power abandonment penalty and load shedding penalty are taken as the minimization targets. Since the hydrogen and oxygen produced by water electrolysis are proportional, the present invention only considers the hydrogen cost. The formula of the objective function is as follows:

[0098] The objective function constructed in step S1 of the present invention is as follows:

[0099]

[0100] Where: C H is the annual unit hydrogen production cost;

[0101] C I is the annual investment cost;

[0102] C M is the annual maintenance cost;

[0103] m H is the mass of hydrogen per year;

[0104] C WP Punishment for abandoning the wind and light for the year;

[0105] C L Penalty for annual load shedding;

[0106] In formula (1)

[0107]

[0108] Where: α is the depreciation rate;

[0109] y is the useful life;

[0110] N W 、N P 、N B 、N S 、N EB 、N H 、N O They are the capacities of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks;

[0111] are the investment costs per unit capacity of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks;

[0112] In formula (1),

[0113] Where: They are the annual maintenance costs per unit capacity of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks;

[0114] In formula (1),

[0115] Where: δ is the hydrogen production rate of the electrolyzer per unit capacity per unit time per unit power;

[0116] In formula (1),

[0117] in

[0118] Where: β is the penalty coefficient for wind and solar power abandonment;

[0119] It is the abandoned wind and solar power;

[0120] is the hourly wind power generation;

[0121] is the hourly photovoltaic power generation;

[0122] P EBmax is the rated power of the electrolyzer;

[0123] The hourly discharge power of the battery;

[0124] is the hourly discharge power of the supercapacitor;

[0125] In formula (1),

[0126]

[0127] Where: γ is the load shedding penalty coefficient;

[0128] The hourly operating power of the electrolyzer.

[0129] S2: Construct a search space; take the capacity of each component of the wind and solar power generation unit, the power storage unit, and the hydrogen and oxygen storage unit as decision variables, and set constraints that meet the actual needs of the system;

[0130] The decision variables in step S2 of the present invention are the capacities of the wind turbine, photovoltaic generator, battery, supercapacitor, electrolyzer, hydrogen storage tank and oxygen storage tank, and the power balance constraint is:

[0131]

[0132] Where: is the charging power of the battery;

[0133] is the charging power of the supercapacitor;

[0134] Range constraints: All variables involved above are non-negative and have upper and lower limit constraints.

[0135] S3: Initialize the population; according to the individuals in the initial population of the search space constructed in step S2, make them meet the constraints;

[0136] The initial population Pop consists of N individuals X, each individual X has n variables, and the initialization method of each variable x of individual X is as follows:

[0137] x=x min +rand nor (x max -x min ) (8)

[0138] Where: x max and x min are the upper and lower bounds of the variable;

[0139] rand nor ∈[0,1] is a random number between 0 and 1 that conforms to a uniform distribution.

[0140] S4: Adaptive selection of search strategy; the strategy adaptive network selects the appropriate search strategy based on the individual decision variables and objective function values, and evolves a new sub-population; the population iteration process includes building a strategy pool, selecting methods, and a strategy adaptive network;

[0141] The strategy pool constructed in step S4 of the present invention is a new search strategy designed based on the existing efficient search strategy. The strategy pool is composed of formulas (11), (12), (13), (14), and (15). The specific formula is as follows:

[0142]

[0143] Where: x r1 , x r2 and x r3 is a randomly selected parent solution;

[0144] x best is the global optimal solution in the current population. In multi-objective optimization, it is the solution in the first non-dominated layer.

[0145] and CR are predefined parameters;

[0146] r is a random number sampled from [0,1] with uniform distribution;

[0147]

[0148] Where: μ is a uniformly distributed random value sampled in [0,1];

[0149] η is a predefined parameter;

[0150]

[0151] The strategy adaptive network is a multi-layer fully connected network, whose input is the position state of each individual in the decision space and the target space, and whose output is the probability of selecting different search strategies. Since the positions of the decision space and the target space are considered at the same time, the designed strategy adaptive network is more conducive to the prediction of strategy probability. Figure 3 As shown, after the strategy adaptive network calculates the probability of each strategy, it will select the specific search strategy to be executed through the ε greedy and roulette method, and the specific steps are as follows: the present invention will randomly select the search strategy with a probability of ε, and use the roulette method to select the search strategy with a probability of 1-ε according to the strategy probability calculated by the strategy adaptive network. In the early stage of network update, the present invention will randomly select the search strategy with a slightly larger ε, and in the later stage, it will select with a larger probability according to the calculated strategy probability.

[0152] S5: Update the population and strategy adaptation network; fuse the parent and child populations to calculate the non-dominated level and crowding degree of all individuals, determine the parameters of the reward update strategy adaptation network based on the non-dominated level and screen out the new population.

[0153] The specific steps of step S5 are as follows:

[0154] S51: After the fusion of the parent and child populations, it is necessary to calculate the non-dominated level and crowding degree of all individuals, and use this as a basis to update the new population. In the optimization of energy storage system capacity configuration, since multiple objectives are involved, the new population cannot be screened by relying solely on the fitness value as in single-objective optimization. Therefore, the dominance relationship between all solutions must be calculated, and the solutions must be sorted by non-domination according to the dominance relationship, and the solutions with the highest non-dominated level must be selected to enter the new population. In order to ensure the diversity of the population, it is also necessary to evaluate the crowding degree of the solution in the target space, retain those solutions with relatively sparse distribution, and select some solutions from dense areas to join the new population.

[0155] The definition of dominance relation is as follows: given two solutions A and B, if A is at least as good as B in all objectives and A is better than B in at least one objective, then solution A dominates solution B; non-dominated level: solutions that are not dominated by other solutions are called the first non-dominated layer (level 1). After removing the first layer, find solutions that are not dominated by the remaining solutions to form the second non-dominated layer (level 2). The calculation formula of crowding is as follows:

[0156]

[0157] Where: f i (A next ) and f i (A prev ) are the target values ​​of the adjacent solutions of solution A on target i;

[0158] f i (max) and f i (min) is the maximum and minimum value of all solutions on target i;

[0159] S52: Determine the reward returns of different search strategies based on the dominance relationship and update the parameters of the strategy adaptive network. The main idea of ​​updating the parameters is as follows: If a strategy is good, then for all states S, the expected value of the state should be large. Therefore, the objective optimization function of the strategy adaptive network is:

[0160]

[0161] Where: s is the position of the decision space and the target space;

[0162] p is a search strategy;

[0163] θ is the parameter of the network;

[0164] π θ (p|s) represents the probability of taking action p when the individual is in state s when the parameter of the strategy adaptation network is θ;

[0165] r(s,p) represents the reward obtained by adopting strategy p in state s;

[0166] By updating the parameters θ of the policy adaptation network, the value of J(θ) becomes larger and larger, which means that the policy adaptation network becomes stronger and stronger. Gradient ascent can be used to solve the maximization problem.

[0167] Let the parameters of the current strategy adaptation network be θ now , then the new parameter θ new Updated via:

[0168]

[0169] Where: σ is the learning rate;

[0170] is the policy gradient;

[0171] Reward is the focus of gradient update. Taking Formula 8 as an example, the reward design after the search strategy is executed is as follows:

[0172]

[0173] Where: Indicates a dominant relationship. Indicates that x is better than x in at least one objective function r1 ;

[0174] The relationship between the rewards is as follows R4<0 <R1<R2<R3。

[0175] The strategy pool of the present invention is a new gradient-guided search strategy designed based on the existing efficient search strategy. The specific construction process is as follows: Exploration and convergence operations usually accompany the entire process of algorithm search, but with the continuous iteration of the population, the algorithm should change from tending to explore the entire search space to tending to converge to the optimal solution or local optimal solution. Therefore, an excellent strategy pool should contain various search strategies from exploration to convergence. In the past nearly two decades, the differential evolution algorithm and its variants have repeatedly achieved excellent rankings in the single-objective optimization competition of IEEE CEC. The following are some differential evolution search strategies with different characteristics:

[0176]

[0177] Where: x = (x 1 ,…,x i ,…,x D ) is the sub-solution;

[0178] D is the dimension of the decision variables;

[0179] x r1 , x r2 and x r3is a randomly selected parent solution;

[0180] x best is the global optimal solution in the current population. In multi-objective optimization, it is the solution in the first non-dominated layer.

[0181] and CR are predefined parameters;

[0182] r is a random number sampled from [0,1] with uniform distribution.

[0183] Formula (9) has strong randomness, which is conducive to global search. It is generally suitable for the initial stage of search and increases the diversity of the population. Formula (10) converges quickly, but it will lead to insufficient diversity and is suitable for local convergence in the later stage. Formula (11) takes into account the advantages of global search and local convergence and is suitable for complex multi-peak optimization problems. In addition, simulated binary crossover is the most popular crossover operator for continuous optimization in genetic algorithms and decomposition-based multi-objective evolutionary algorithms.

[0184]

[0185] Where: μ is a uniformly distributed random value sampled in [0,1], and η is a predefined parameter.

[0186] Although the above formula has been widely used, there may still be some invalid search problems during the iteration process. Figure 4 As shown in the figure on the left, As a basis for solving two different and The possible distributions of solutions produced by the difference are all in However, from the entire function, we can see that the potential area is This phenomenon is caused by the wrong direction of the difference, which will seriously affect the search efficiency, especially the efficiency of local search. To solve this problem, we introduce the gradient to guide the search direction. Figure 4 In the same situation on the left, the solution generated by introducing the gradient will be distributed in the more promising area on the right. The modified formula is as follows

[0187]

[0188] Formula (13) is derived from formula (9). Formula (9) is too random and blind and is only suitable for the early stage of search to expand the diversity of the population. Formula (13) reduces randomness to a certain extent after introducing gradient information. Formula (14) is derived from formula (10). Formula (10) uses differential search near the optimal solution and converges quickly, but it is easy to fall into the local optimum and is not suitable for multi-objective optimization. For this reason, the present invention makes a slight modification so that the basic solution refers to the position of the optimal solution and its own gradient information, so that it is not easy to fall into the local optimum. Formula (15) is derived from formula (12). While retaining its global search capability, it increases the ability to converge to the optimal solution. Formula (11) contains two differential vectors, which is more complicated and difficult to add gradient information. Moreover, it itself takes into account both global search and local convergence, so it is not changed.

[0189] In summary, the strategy pool of the present invention will be composed of formulas (11), (12), (13), (14) and (15).

[0190] Based on the existing differential evolution operator and simulated binary crossover operator, the present invention designs a gradient-guided search strategy to reduce the randomness of the algorithm search. In addition, the present invention designs a strategy adaptive network to solve the problem of the singleness of the search strategy. The network can adaptively select the appropriate search strategy according to different optimization problems and different individual states, which can effectively enhance the search efficiency of the algorithm and has good portability.

[0191] An embodiment of the present invention is given

[0192] The actual data of wind speed and light intensity in a certain area are selected to minimize the annual unit hydrogen production cost, annual wind and solar abandonment penalty, and annual load shedding penalty as the optimization goal. The population size is set to 50, the number of iterations is set to 50, the capacity of the electrolyzer is fixed, and the capacity of the wind turbine, photovoltaic generator, battery, supercapacitor, hydrogen storage tank, and oxygen storage tank are optimized. Figure 5 As shown in the figure, the comparison results between the adaptive evolution method with gradient-guided search strategy pool proposed in the present invention and the classic NSGAII, the HV index is a commonly used index for analyzing the advantages and disadvantages of multi-objective algorithms, and the larger the value of the HV index, the better the convergence and distribution of the adopted method. Figure 5 It can be seen that the method of the present invention has a good search capability, can quickly converge to the optimal solution, and basically tends to be stable in the later stage of iteration. The HV index of the NSGAII algorithm has been oscillating continuously. This is because the population size of 50 is relatively small for NSGAII. Although it produces many non-dominant solutions, it has not been able to get close to the optimal solution. This also shows that the method proposed in the present invention has good search efficiency and convergence ability. Figure 6As shown, the Pareto optimal solution set searched by the method of the present invention, if the optimal solution selected for the three objectives with a preference of 1:1:1, its corresponding minimum optimization result is: the annual unit hydrogen production cost is 29, the annual wind and solar power abandonment penalty is 205, and the annual load shedding penalty is 252. The corresponding hydrogen cost is 29 yuan / kg, and the abandonment rate (wind and solar power abandonment divided by the total power generation) and the power shortage rate (load shedding power divided by the total power generation) are both less than 1%. If the user has different preferences, there is no need to repeatedly solve the problem. The results in the optimal solution set can be directly weighted and summed according to the corresponding preferences and sorted to find the optimal solution under the corresponding preferences.

[0193] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. An adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen energy storage system, characterized in that: The following steps are involved: S1: Construct the objective function; according to the energy storage system model, take the unit hydrogen production cost, wind and solar power abandonment penalty and load shedding penalty as the minimization objectives; S2: Construct a search space; take the capacity of each component of the wind and solar power generation unit, the power storage unit, and the hydrogen and oxygen storage unit as decision variables, and set constraints that meet the actual needs of the system; S3: Initialize the population; according to the individuals in the initial population of the search space constructed in step S2, make them meet the constraints; S4: Adaptive selection of search strategy; the strategy adaptation network selects the appropriate search strategy based on the individual decision variables and objective function values, and evolves a new subpopulation; The population iteration process includes building a strategy pool, selecting methods, and strategy adaptation networks; S5: Update the population and strategy adaptation network; fuse the parent and child populations to calculate the non-dominated level and crowding degree of all individuals, determine the parameters of the reward update strategy adaptation network based on the non-dominated level and screen out the new population.

2. According to claim 1, an adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen energy storage system is characterized in that: The energy storage system model includes a power generation unit, a power storage unit, a hydrogen and oxygen storage unit, and an electrolyzer; the power generation unit is composed of a wind turbine and a photovoltaic generator. The power generation unit converts wind energy and light energy into electrical energy to supply the electrolyzer with hydrogen and oxygen production, and then stores it in the hydrogen and oxygen storage unit; when there is a surplus of electricity generated, the system will detect whether the capacity of the power storage unit is full. If the power storage unit still has capacity, the power storage unit composed of batteries and supercapacitors will start to store electricity. If the capacity of the power storage unit is full, energy waste will occur, resulting in wind and solar power abandonment; the electrolyzer usually operates at rated power and produces hydrogen and oxygen by electrolyzing water. These hydrogen and oxygen are stored in hydrogen storage tanks and oxygen storage tanks respectively. If the sum of the discharges of the power generation unit and the power storage unit cannot meet the rated power of the electrolyzer, the electrolyzer will cut load and operate.

3. The adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen energy storage system according to claim 2 is characterized in that: The objective function constructed in step S1 is as follows: Where: C H is the annual unit hydrogen production cost; C I is the annual investment cost; C M is the annual maintenance cost; m H is the mass of hydrogen per year; C WP Punishment for abandoning the wind and light for the year; C L Penalty for annual load shedding; In formula (1) Where: α is the depreciation rate; y is the useful life; N W 、N P 、N B 、N S 、N EB 、N H 、N O They are the capacities of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks; are the investment costs per unit capacity of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks; In formula (1), Where: They are the annual maintenance costs per unit capacity of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks; In formula (1), Where: δ is the hydrogen production rate of the electrolyzer per unit capacity per unit time per unit power; In formula (1), in Where: β is the penalty coefficient for wind and solar power abandonment; It is the abandoned wind and solar power; is the hourly wind power generation; is the hourly photovoltaic power generation; P EBmax is the rated power of the electrolyzer; The hourly discharge power of the battery; is the hourly discharge power of the supercapacitor; In formula (1), Where: γ is the load shedding penalty coefficient; The hourly operating power of the electrolyzer.

4. The adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen energy storage system according to claim 2 is characterized in that: The decision variables in step S2 are the capacities of wind turbines, photovoltaic generators, batteries, supercapacitors, electrolyzers, hydrogen storage tanks, and oxygen storage tanks, and the power balance constraint is: Where: is the charging power of the battery; is the charging power of the supercapacitor; Range constraints: All variables involved above are non-negative and have upper and lower limit constraints.

5. The adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen energy storage system according to claim 2, characterized in that: The initial population Pop in step S3 consists of N individuals X, each of which has n variables. The initialization method of each variable x of individual X is as follows: x=x min +rand nor (x max -x min ) (8) Where: x max and x min are the upper and lower bounds of the variable; rand nor ∈[0,1] is a random number between 0 and 1 that conforms to a uniform distribution.

6. The adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen energy storage system according to claim 2, characterized in that: The strategy pool constructed in step S4 is to design a new gradient-guided search strategy based on the existing efficient search strategy. The strategy pool is composed of formulas (11), (12), (13), (14), and (15). The specific formulas are as follows: Where: x r1 , x r2 and x r3 is a randomly selected parent solution; x best is the global optimal solution in the current population; and CR are predefined parameters; r is a random number sampled from [0,1] with uniform distribution; Where: μ is a uniformly distributed random value sampled in [0,1]; v is a predefined parameter; The strategy adaptation network is a multi-layer fully connected network, whose input is the position state of each individual in the decision space and the target space, and whose output is the probability of selecting different search strategies.

7. The adaptive multi-objective evolutionary method for optimizing the capacity of an off-grid wind-solar hydrogen energy storage system according to claim 2, characterized in that: The specific steps of step S5 are as follows: S51: After the fusion of the parent and child populations, it is necessary to calculate the non-dominated level and crowding degree of all individuals, and use this as a basis to update the new population. In the optimization of energy storage system capacity configuration, since multiple objectives are involved, the new population cannot be screened by relying solely on the fitness value as in single-objective optimization. Therefore, the dominance relationship between all solutions must be calculated, and the solutions must be sorted by non-domination according to the dominance relationship, and the solutions with the highest non-dominated level must be selected to enter the new population. In order to ensure the diversity of the population, it is also necessary to evaluate the crowding degree of the solution in the target space, retain those solutions with relatively sparse distribution, and select some solutions from dense areas to join the new population. The definition of dominance relation is as follows: given two solutions A and B, if A is at least as good as B in all objectives and A is better than B in at least one objective, then solution A dominates solution B; non-dominated level: solutions that are not dominated by other solutions are called the first non-dominated layer. After removing the first layer, find solutions that are not dominated by the remaining solutions to form the second non-dominated layer. The calculation formula of crowding degree is as follows: Where: f i (A next ) and f i (A prev ) are the target values ​​of the adjacent solutions of solution A on target i; f i (max) and f i (min) is the maximum and minimum value of all solutions on target i; S52: Determine the reward returns of different search strategies based on the dominance relationship and update the parameters of the strategy adaptive network. The main idea of ​​updating the parameters is as follows: If a strategy is good, then for all states S, the expected value of the state should be large. Therefore, the objective optimization function of the strategy adaptive network is: Where: s is the position of the decision space and the target space; p is a search strategy; θ is the parameter of the network; π θ (p|s) represents the probability of taking action p when the individual is in state s when the parameter of the strategy adaptation network is θ; r(s,p) represents the reward obtained by adopting strategy p in state s; By updating the parameters θ of the policy adaptation network, the value of J(θ) becomes larger and larger, which means that the policy adaptation network becomes stronger and stronger. Gradient ascent is used to solve the maximization problem. Let the parameters of the current strategy adaptation network be θ now , then the new parameter θ new Updated via: Where: σ is the learning rate; is the policy gradient; Reward is the focus of gradient update. Taking Formula 8 as an example, the reward design after the search strategy is executed is as follows: Where: Indicates a dominant relationship. Indicates that x is better than x in at least one objective function r1 ; The relationship between the rewards is as follows R4<0 <R1<R2<R3。

Images