A scene-driven virtual power plant planning-operation two-stage distribution robust optimization method

Abstract

The application discloses a kind of virtual power plant planning-operation two-stage distribution robust optimization methods based on scene driving, comprising: S1, based on the master-slave game framework of power supply, load and energy storage three parties, capacity configuration model and master-slave game model are constructed;S2, considering the correlation and uncertainty of wind turbine, photovoltaic output, construct the uncertainty set of wind light output probability distribution, and generate multiple typical scenarios by sampling and clustering algorithm;S3, under the driving of typical scenario, the confidence interval of uncertainty set is comprehensively norm-constrained, and the distribution robust optimization model of virtual power plant considering uncertainty is established in planning and operation two stages;Wherein, distribution robust optimization model contains capacity configuration model and master-slave game model;S4, based on the column and constraint generation algorithm of modified Kriging model, distribution robust optimization model containing master-slave game model is solved.Utilizing the application, the influence of uncertainty on virtual power plant planning rationality and operation economy can be reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of distributed robust optimization of virtual power plants, and particularly relates to a scenario-driven two-stage distributed robust optimization method for planning and operation of virtual power plants. BACKGROUND

[0002] Under the operation mechanism of the new power system mainly based on electric energy, the connection between heterogeneous energies such as electricity, gas, heat and cold is becoming closer and closer. However, the large-scale access of distributed resources such as wind power, photovoltaic and cogeneration units brings severe challenges to the dispatching and operation of the existing power system. Renewable energy generation has uncertainty, its capacity is small, and the volatility of power output is large, and large-scale grid connection will seriously affect the stability of the power system.

[0003] With the continuous increase of the renewable energy penetration rate in the distribution network and the continuous increase of the proportion of user terminal energy electrification, the power system gradually changes the operation mode of adjusting the output of generator units according to the load change, and transforms to the coordinated optimization operation of the power supply side, the load side and the energy storage side. Therefore, the distributed resources make the system operation mode and energy coupling relationship more complex, which greatly increases the difficulty of internal energy management of the system, and how to tap the interactive ability and adjustment ability of flexible resources such as source, load and storage is a key problem in the development of new power systems.

[0004] Virtual power plants break the geographical location and resource type restrictions, and through communication, metering and other control means, the distributed resources dispersed in adjacent areas are aggregated into an organic whole that can accept unified dispatching. As an entity integrated in the form of "virtual", the virtual power plant can independently carry out energy dispatching and market transactions, providing an effective solution for the reasonable consumption and centralized grid connection of distributed resources. Among the source, load and storage, the output of renewable energy on the source side and the user energy consumption on the load side both have uncertainty, which affects the rationality and economy of the internal resource capacity configuration and optimal dispatching of the virtual power plant.

[0005] At present, the main methods for dealing with the operation uncertainty of virtual power plants are random optimization and robust optimization. In the random optimization, the probability distribution of variables is fixed or estimated according to historical data. For uncertainty problems, the deterministic probability distribution is too optimistic, which may cause large errors and affect the economy of system operation. Robust optimization considers the worst-case scenario, but the actual operation scenario is not as extreme, so its optimization result is too pessimistic, leading to overly conservative decision-making. In terms of solution of optimization methods, the current mainly uses intelligent optimization algorithms, and the solution speed and accuracy still need to be further improved. In addition, the existing optimization methods focus on considering the influence of uncertainty on energy dispatching in the operation stage of virtual power plants, but lack effective quantitative evaluation of the rationality of capacity configuration in the planning stage.

[0006] In summary, in order to reduce the influence of uncertainty on the rationality of virtual power plant planning and the economy of operation, a new optimization method and a matching efficient solving algorithm are needed to be designed. SUMMARY

[0007] In view of the problems that the existing method cannot balance the conservativeness and economy, lacks quantitative evaluation of planning rationality, and solving performance needs to be improved, the purpose of the application is to provide a scenario-driven two-stage distribution robust optimization method for virtual power plant planning and operation, so as to effectively deal with the uncertainty problem and reduce the influence of uncertainty on the rationality of virtual power plant planning and the economy of operation.

[0008] A scenario-driven two-stage distribution robust optimization method for virtual power plant planning and operation, characterized in that it comprises the following steps:

[0009] S1, considering the capacity configuration and energy scheduling of the virtual power plant based on the master-slave game framework of the power supply, load and energy storage, a capacity configuration model and a master-slave game model are constructed;

[0010] S2, considering the correlation and uncertainty of wind turbine and photovoltaic output, an uncertainty set of wind and light output probability distribution is constructed, and a plurality of typical scenarios are generated through sampling and clustering algorithm;

[0011] S3, under the driving of typical scenarios, the confidence interval of the uncertainty set is comprehensively norm-constrained, and a two-stage distribution robust optimization model of virtual power plant considering uncertainty is established; wherein, the distribution robust optimization model comprises the capacity configuration model and the master-slave game model established in step S1;

[0012] S4, based on the modified Kriging model and the column and constraint generation algorithm, the distribution robust optimization model containing the master-slave game model is solved.

[0013] The application fully considers the influence of uncertainty on the medium and long-term capacity configuration and day-ahead optimal scheduling of internal resources of virtual power plant, solves the problem that the conservativeness and economy are difficult to balance in the current uncertainty optimization, and further more, the application solves the distribution robust optimization model by combining the agent model algorithm and the intelligent optimization algorithm, which can effectively improve the solving efficiency compared with the conventional solving method.

[0014] The specific steps of S1 are:

[0015] S101, taking the minimum investment cost of virtual power plant in the planning stage as the target, a capacity configuration model is constructed for the aggregated distributed resources in the virtual power plant, wherein the distributed resources include wind turbine, photovoltaic, combined heat and power unit, biomass boiler, electric ammonia system and energy storage system;

[0016] S102, an energy-benefit interaction model is established, in which the virtual power plant is the leader and the user is the follower, i.e., a master-slave game model; the virtual power plant formulates the energy sale price with the user in advance according to the load day-ahead prediction information, and the user adjusts the energy consumption through the demand response mechanism under the guidance of the price signal of the virtual power plant, and the two sequentially interact the pricing strategy and the energy consumption strategy, and the virtual power plant is in a dominant position in the interaction process and forms a master-slave game with the user;

[0017] S103, in the master-slave game relationship, the upper virtual power plant trades electricity, heat and ammonia with the lower user to minimize the day-ahead operation cost;

[0018] S104, for the price strategy issued by the upper virtual power plant in the day-ahead stage, the lower user adjusts the energy consumption through demand response to maximize the comprehensive benefit.

[0019] In step S101, the investment cost of the virtual power plant in the planning stage is as follows:

[0020]

[0021] In the formula, C inv represents the daily average investment cost of the VPP; i represents the type of equipment inside the VPP; Ω DG is the set of equipment inside the VPP, i∈Ω DG ; C DG,i represents the investment cost of the i-th type of equipment inside the VPP; N DG,i is the installation quantity of the i-th type of equipment; r inv represents the discount rate; TL i represents the economic service life of the i-th type of equipment; λ inv,i is the unit investment cost of the i-th type of equipment.

[0022] In step S102, the objective function and the constraint condition of the master-slave game model are as follows:

[0023] minC op (λ et ,λ ht ,λ at ,P et ,Q ht ,B at )

[0024]

[0025] In the formula, Ω vpp and Ω user represent the master-slave game strategy sets of the virtual power plant and the user, respectively; C op and C eu represent the objective functions of the virtual power plant and the user in the day-ahead stage, respectively; λet,t ht,t at,t respectively represent the electricity, heat and ammonia transaction prices of the virtual power plant and the user; P et,t ht,t at,t respectively represent the electricity, heat and ammonia transaction quantities of the virtual power plant and the user.

[0026] The specific steps of S2 are as follows:

[0027] S201, based on the kernel density estimation method, analyze the historical sample data of wind and light output of the virtual power plant, fit the probability distribution model of wind and light output, and thus deeply mine the wind and light distribution characteristics of the region where the virtual power plant is located;

[0028] S202, since the wind power and photovoltaic power are complementary in part of the time period and present a negative correlation relationship, the Frank-Copula function which takes into account the non-negative and negative correlation relationship between random variables is selected to describe the correlation of wind and light;

[0029] S203, use the Copularnd function to sample and generate a certain number of edge distribution function values (u, v) satisfying the Frank-Copula function relationship, and through inverse transformation of the wind and light edge distribution function values of each time period, obtain the wind and light output scenarios (x, y) of each time period satisfying the joint distribution function;

[0030] S204, to ensure the fitting accuracy of the Frank-Copula function, use the K-means clustering algorithm to cluster the sampling results, generate multiple typical scenarios, and obtain the probability of occurrence of each typical scenario.

[0031] The specific steps of S3 are as follows:

[0032] S301, based on the multiple typical scenarios reflecting the uncertainty and correlation of wind and light output generated in step S2, optimize the probability distribution uncertainty set in the worst scenario, and obtain the planning and operation two-stage decision scheme under the worst scenario; wherein, the distribution robust optimization model of the virtual power plant in the planning and operation two stages includes a capacity configuration model and a principal-agent game model, and the economic optimization of the two stages is taken as the target, and the target function is as follows:

[0033]

[0034] In the formula, x and y s respectively represent the decision variables of the planning stage and the operation stage; A is the coefficient matrix corresponding to the variables in the planning stage target function; B and C are respectively the coefficient matrices of the quadratic term and the linear term in the operation stage target function; d is the constant term coefficient of the operation stage target function; X is the set of planning stage variables; Ω π ​​​​​is the uncertainty set of wind and solar power output probability distribution; s is the typical scenario number; N s is the number of typical scenarios; π s represents the probability of scenario s occurring; Y(x, π s is a set of operating stage variables, representing the coupling relationship between x and y s in scenario s;

[0035] S302, the uncertainty set of wind and solar power output probability distribution and the confidence constraint satisfied by each scenario probability distribution are as follows:

[0036]

[0037]

[0038] wherein R + represents a positive real number; π s,0 represents the initial probability of a typical scenario; θ1 and θ ∞ correspond to the allowable probability fluctuation deviation under 1-norm and ∞-norm constraints, respectively; Pr(·) represents a probability distribution function; and S is the initial number of scenarios.

[0039] The specific steps of S4 are as follows:

[0040] S401, the interaction between the virtual power plant and the user in the principal-agent model is visualized by using the Kriging model, so as to simplify the double-layer optimization problem into a single-layer optimization problem;

[0041] S402, in order to reduce the influence of the number of sample data on the fitting accuracy of the model, a Kriging model correction mechanism considering dynamic updating of sample data is designed;

[0042] S403, the distribution robust optimization problem of the planning and operation two stages is decomposed into a master problem and a sub-problem, the principal-agent model is embedded into the master problem, and the column and constraint generation algorithm is used for stage-by-stage iterative solution.

[0043] Step S401 includes:

[0044] First, a certain number of initial sample points containing the unit time electricity, heat and ammonia transaction price of the virtual power plant and the user are generated for each discrete scenario by using Latin hypercube sampling, the number of each group of initial sample points is selected to be 10 times the number of transaction price variables, the initial sample points are substituted into the lower user model to obtain the electricity, heat and ammonia transaction volume corresponding to each group of transaction price, so as to construct an initial sample data set containing transaction price and transaction volume;

[0045] Then, the mapping relationship between the transaction price and the transaction volume in the sample data set is predicted using the Kriging model, so as to fit and replace the lower layer model according to the mapping relationship between the upper layer and the lower layer, and the single-layer optimization model of each scenario obtained after fitting is as follows:

[0046]

[0047] In the formula, H(·) represents the mapping relationship between the upper layer and the lower layer interaction obtained by using the Kriging model.

[0048] Step S402 includes:

[0049] First, the initial sample data set of each scenario is constructed by uniform sampling and calling the lower layer model, and the initial Kriging model is established based on the mapping relationship inside the data set;

[0050] Then, the main problem corresponding to the upper layer virtual power plant model is solved to obtain the optimal transaction price and transaction volume, but the transaction volume at this time is the predicted value based on the Kriging model rather than the true value, so the optimal transaction price needs to be substituted into the lower layer user model to obtain the real transaction volume, and the optimal transaction price and the real transaction volume corresponding thereto are added to the initial sample data set as a high-quality data set, so as to realize the correction of the Kriging model;

[0051] Finally, the sample data is dynamically updated through multiple iterations until the objective function value of the main problem converges.

[0052] In step S403, the main problem and the sub-problem obtained by decomposition are as follows:

[0053] Main problem:

[0054]

[0055] In the formula, l represents the iteration number, l∈{1,2,3,…,N l}, wherein N l is the total iteration number;

[0056] (2) Sub-problem;

[0057]

[0058] The max-min double-layer structure of the sub-problem includes an outer maximum problem and an inner minimum problem, and the probability distribution of the worst scenario in the confidence interval is found according to the solving result of the main problem; since the outer discrete scenario probability π s,l of the sub-problem and the inner running stage variable y s,lThe two sub-problems are independent of each other, so the two sub-problems are solved in two steps, that is, the inner minimum value problem is solved first, and then the outer maximum value problem is solved, wherein the inner minimum value problem can be obtained from the solving result of the main problem.

[0059] Compared with the prior art, the present application has the following beneficial effects:

[0060] 1、The present application fits the probability distribution model of wind and light output based on the kernel density estimation method, thereby deeply mining the wind and light distribution characteristics of the region where the virtual power plant is located, and selecting the Frank-Copula function which can consider the non-negative and negative correlation between random variables to effectively depict the correlation of wind and light output. In order to ensure the fitting accuracy of the Frank-Copula function, the present application uses a clustering algorithm to cluster the sampling results and generates multiple typical scenes reflecting uncertainty and correlation. Compared with existing uncertainty sampling methods, the present application comprehensively considers the time and space correlation between uncertainty variables, and therefore is more scientific.

[0061] 2、The present application drives under the multi-dimensional uncertainty scenario, adopts a distribution robust optimization method to construct an uncertainty set of wind and light output probability distribution, and based on a master-slave game framework of source-load-storage, designs a planning-operation two-stage distribution robust optimization model considering capacity configuration and energy scheduling of the virtual power plant. The model fully quantifies the influence of the uncertainty of the virtual power plant operation on medium and long-term planning and day-ahead scheduling, and is helpful for formulating more economical capacity configuration and optimization scheduling scheme.

[0062] 3、The present application designs a sampling point dynamic updating mechanism, modifies the traditional Kriging model, and uses the modified Kriging model to realize effective fitting and transformation of the master-slave game double-layer optimization model. The present application decomposes the distributed robust optimization problem into a main problem and a sub-problem, and nests the modified Kriging model into the column and constraint generation algorithm, and carries out hierarchical decoupling optimization and stage-by-stage iterative solving on the two-stage optimization problem containing the double-layer optimization model. Compared with the traditional analytic method and intelligent optimization algorithm, the present application can effectively reduce the solving time and reduce the complexity of the problem by nesting the surrogate model for solving. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 It is the overall framework diagram of the virtual power plant planning-operation two-stage distribution robust optimization method in the embodiment of the present application;

[0064] Figure 2 It is the flowchart of solving the distributed robust optimization problem based on the modified Kriging model and the column and constraint generation algorithm in the embodiment of the present application;

[0065] Figure 3The convergence process diagram of respectively embedding the modified Kriging model algorithm, the traditional Kriging model algorithm, the particle swarm algorithm and the KKT condition embedding column and constraint generation algorithm into the solving of the main problem and the sub-problem in the embodiment of the application. DETAILED DESCRIPTION

[0066] The application will be described in further detail below with reference to the drawings and embodiments, and it should be noted that the following embodiments are intended to facilitate the understanding of the application and do not limit the application in any way.

[0067] As shown in the embodiment, a scenario-driven virtual power plant planning-operation two-stage distribution robust optimization method is provided, including the following steps: Figure 1

[0068] S1, based on the principal-agent game framework of the source-load-storage three parties, a planning-operation conventional optimization model considering the capacity configuration and energy scheduling of the virtual power plant is established.

[0069] S2, the correlation and uncertainty of the wind turbine and photovoltaic output are considered to construct the uncertainty set of the wind-solar output probability distribution, and a plurality of typical scenarios are generated through a sampling and clustering algorithm.

[0070] S3, under the scenario driving, the confidence interval of the uncertainty set is comprehensively norm-constrained, and a virtual power plant planning-operation two-stage distribution robust optimization model considering the uncertainty is established.

[0071] S4, the distribution robust optimization problem containing the principal-agent game model is solved based on the modified Kriging model and the column and constraint generation algorithm.

[0072] Further, the detailed steps of step S1 include:

[0073] S101, taking the minimum long-term investment cost of the virtual power plant as the target, the capacity configuration is performed for the distributed resources such as wind turbines, photovoltaic, combined heat and power units, biomass boilers, electric ammonia systems and energy storage systems aggregated in the virtual power plant.

[0074] The investment cost of the virtual power plant in the planning stage is as follows:

[0075]

[0076] In the formula, C inv represents the daily average investment cost of the VPP; i represents the type of the internal equipment of the VPP; Ω DG is the set of the internal equipment of the VPP, i∈Ω DG ; C DG,i represents the investment cost of the i-th type of equipment in the VPP; N DG,i is the installation quantity of the i-th type of equipment; r inv represents the discount rate; TL​i denotes the economic life of the ith type of equipment; λ inv,i is the unit investment cost of the ith type of equipment.

[0077] S102, an energy-benefit interaction model is established with a virtual power plant as a leader and users as followers. The virtual power plant formulates energy sales prices with the users in priority according to load day-ahead prediction information, and the users adjust energy consumption through a demand response mechanism under the guidance of the virtual power plant price signal. The two parties sequentially interact pricing strategies and energy consumption strategies, and the virtual power plant is in a dominant position in the interaction process, forming a master-slave game with the users.

[0078] The objective function and constraint conditions of the master-slave game model between the virtual power plant and the users are as follows:

[0079] min C op (λ et ,λ ht ,λ at , P et , Q ht , B at ) (2)

[0080]

[0081] In the formula, Ω vpp and Ω user respectively represent the master-slave game strategy sets of the virtual power plant and the users; C op and C eu respectively represent the objective functions of the virtual power plant and the users in the day-ahead stage; λ et,t , λ ht,t , and λ at,t respectively represent the electricity, heat, and ammonia transaction prices of the virtual power plant and the users; P et,t , Q ht,t , and B at,t respectively represent the electricity, heat, and ammonia transaction quantities of the virtual power plant and the users.

[0082] S103, in the master-slave game relationship, the upper virtual power plant carries out electricity, heat, and ammonia transactions with the lower users with the objective of minimizing day-ahead operation cost.

[0083] The objective function C op of the virtual power plant in the day-ahead operation stage is as follows, specifically including unit fuel cost C raw , unit operation and maintenance cost C om , energy transaction cost with the upper network C grid , energy sales revenue with the users C user , and demand response cost C DR .

[0084] min C op= C raw + C om + C grid + C DR - C user (4)

[0085] Wherein, the unit fuel cost is shown as follows, including the natural gas purchase cost of the cogeneration unit and the biomass raw material purchase cost of the biomass boiler:

[0086]

[0087] In the formula, L g is the low heat value of natural gas; λ bio is the purchase price of biomass raw material; P gt,t represents the electric power output of the cogeneration unit; Q bfb,t represents the heat power output of the biomass boiler; ρ g is the density of biogas; η gt and η bfb are the operating efficiencies of the cogeneration unit and the biomass boiler; η g represents the volume of natural gas produced per unit of biomass raw material fermentation;

[0088] The unit operation and maintenance cost is shown as follows, including the daily inspection and fault recovery of the wind turbine, photovoltaic, cogeneration unit, biomass boiler, electric ammonia system and energy storage system.

[0089]

[0090] In the formula, P pv,t and P wt,t represent the electric power output of photovoltaic and wind power respectively; P PA,t represents the electric power consumption of P2A; P ch,t , Q ch,t , B ch,t and P dis,t , Q dis,t , B dis,t represent the energy absorbed and released by the battery, heat storage device and ammonia storage device in the energy storage system at each unit time; λ pv , λ wt , λ CHP , λ PA , λ bfb , λ BSS , λ HSS , λ GSS represent the unit power operation and maintenance cost of photovoltaic, wind turbine, cogeneration unit, electric ammonia system, biomass boiler, battery, heat storage device and ammonia storage device respectively.

[0091] The energy transaction cost with the upper network is as follows, including the heterogeneous energy transaction of the virtual power plant with the upper power network, the heat network and the ammonia network.

[0092]

[0093] In the formula, λ eb,t and λ es,t respectively represent the purchase and sale electricity prices of the virtual power plant with the upper power grid; P eb,t and P es,t respectively represent the purchase and sale electricity quantities of the virtual power plant with the upper power grid; λ hb and λ hs respectively represent the purchase and sale heat prices of the virtual power plant with the upper heat grid; Q hb,t and Q hs,t respectively represent the purchase and sale heat quantities of the virtual power plant with the upper heat grid; λ ab and λ as respectively represent the purchase and sale ammonia prices of the virtual power plant with the upper ammonia grid; B ab,t and B as,t respectively represent the purchase and sale ammonia quantities of the virtual power plant with the upper ammonia grid.

[0094] The revenue of the virtual power plant for selling electricity, heat and ammonia to users is as follows:

[0095]

[0096] The demand response cost of the virtual power plant with users is as follows, specifically including the compensation cost of reducing the electricity load, the heat load and the ammonia load on the user side.

[0097]

[0098] In the formula, β re , β rh and β ra respectively represent the compensation prices of the demand response for reducing the electricity load, the heat load and the ammonia load; P re,t , Q rh,t and B ra,t respectively represent the reduction quantities of the demand response for the electricity load, the heat load and the ammonia load.

[0099] S104, for the price strategy of the upper virtual power plant in the day-ahead stage, the lower user adjusts the own energy consumption through demand response with the maximum comprehensive benefit as the target.

[0100] The user comprehensive benefit C eu includes the user demand response revenue and the energy purchase cost, and the target function is as follows:

[0101] max C eu = C DR -C user (10)

[0102] Further, the detailed steps of step S2 include:

[0103] S201, based on the kernel density estimation method, analyze the historical sample data of wind and light output of the virtual power plant, fit the probability distribution model of wind and light output, and deeply mine the wind and light distribution characteristics of the region where the virtual power plant is located.

[0104] Let x1, x1,…, x n is a sample independent of wind turbine output x, then the probability density function f h (x) of the distribution to which x is subject to is estimated by the kernel density as follows. The kernel density estimation of the probability density function f h (y) of photovoltaic output y can be obtained in the same way.

[0105]

[0106] In the formula, n is the sample size; h is the window width; K(·) represents the kernel function, and the kernel function has various forms. The formula of the Gaussian kernel function selected in this chapter is as follows:

[0107]

[0108] S202, since wind power and photovoltaic power are complementary in some time periods and show a negative correlation, the Frank-Copula function which can take into account the non-negative and negative correlation between random variables is selected to describe the correlation between wind and light.

[0109] The introduced Frank-Copula function C(u,v) reflecting the relationship between the marginal distribution and the joint distribution of wind and light output in each period is as follows:

[0110]

[0111] In the formula, F X (x) and F Y (y) represent the marginal distribution functions of wind power and photovoltaic power respectively; F h (x,y) is the joint distribution function of wind and light output.

[0112] S203, use the Copularnd function to sample and generate a certain number of marginal distribution function values (u,v) satisfying the Frank-Copula function relationship, and through the inverse transformation of the marginal distribution function values of wind and light in each period, the wind and light output scenarios (x,y) satisfying the joint distribution function can be obtained.

[0113] ​S204. Since the number of initial scenarios obtained after sampling is relatively large, in order to ensure the fitting accuracy of the Frank-Copula function, the sampling results are clustered using the K-means clustering algorithm to generate multiple typical scenarios and obtain the probability of each typical scenario occurring.

[0114] Furthermore, the detailed steps of step S3 include:

[0115] S301. Based on multiple typical scenarios reflecting the uncertainty and correlation of wind and solar power output generated in step S2, the probability distribution uncertainty set in the worst scenario is optimized, and a planning-operation two-stage decision-making scheme under the worst scenario is obtained.

[0116] The distributed robust optimization model for the virtual power plant in the planning and operation stages includes the capacity configuration model established in step S101 and the master-slave game model established in step S102, with the economic optimization of the two stages as the goal. The objective function is as follows:

[0117]

[0118] Where x and y s Denote the decision variables in the planning stage and the operation stage respectively; A is the coefficient matrix corresponding to the variables in the objective function of the planning stage; B and C are the coefficient matrices of the quadratic term and the linear term in the objective function of the operation stage respectively; d is the constant term coefficient of the objective function of the operation stage; X is the set of variables in the planning stage; Ω π is the uncertainty set of wind and solar power output probability distribution; s is the typical scenario number; N s is the number of typical scenarios; π s represents the probability of scene s occurring; Y(x,π s ) is a set of variables in the running phase, which can be represented as x and y in scene s s coupling relationship.

[0119] S302: To ensure that the actual probability distribution of each scenario still converges to the theoretical probability distribution when the number of generated scenarios is large enough, the confidence intervals of the 1-norm and the ∞-norm are combined to constrain the fluctuation range of the probability distribution.

[0120] The uncertainty set of wind and solar power output probability distribution and the confidence constraints satisfied by the probability distribution of each scenario are shown in the following formula:

[0121]

[0122]

[0123] Where R + represents a positive real number; π s,0 represents the initial probability of a typical scenario; θ1 and θ∞ respectively correspond to the allowed probability fluctuation bias under 1-norm and ∞-norm constraints; Pr(·) represents the probability distribution function; S is the initial scenario number.

[0124] Let α1 and α ∞ respectively represent the confidence degrees corresponding to the 1-norm and ∞-norm confidence intervals of the probability distribution of each scenario, and let the quantile points of formula (16) be α1 and α ∞ Then the probability fluctuation bias can be calculated as follows:

[0125]

[0126] Further, Figure 2 The flowchart for solving the distribution robust optimization problem in step S4 based on the modified Kriging model and the column and constraint generation algorithm, a dynamic sampling point updating mechanism is designed on the basis of the traditional Kriging model for modification, and the modified Kriging model is embedded in the column and constraint generation algorithm, and the two-stage distribution robust optimization model containing the master-slave game double-layer optimization problem is iteratively solved, and the detailed steps include:

[0127] S401, the interaction between the virtual power plant and the user in the master-slave game model is displayed by the Kriging model, so as to simplify the double-layer optimization problem into a single-layer optimization problem.

[0128] Firstly, Latin hypercube sampling is used to generate a certain number of initial sample points containing the unit time electricity-heat-ammonia transaction price of the virtual power plant and the user for each discrete scenario, and the number of each group of initial sample points is selected to be 10 times the number of transaction price variables. The transaction volume corresponding to each group of transaction prices can be obtained by substituting the initial sample points into the lower user model, so as to construct an initial sample data set containing transaction price and transaction volume. Then the mapping relationship between the transaction price and the transaction volume in the sample data set is predicted by using the Kriging model, so as to fit and replace the lower model according to the mapping relationship between the upper and lower layers. The single-layer optimization model of each scenario obtained after fitting is shown in formula (18):

[0129]

[0130] In the formula, H(·) represents the mapping relationship between the upper and lower interaction quantities predicted by using the Kriging model.

[0131] S402, in order to reduce the influence of the number of sample data on the fitting accuracy of the model, a Kriging model modification mechanism considering dynamic updating of sample data is designed.

[0132] Firstly, the initial sample dataset of each scenario is constructed by uniform sampling and calling the lower model, and the initial Kriging model is established based on the mapping relationship inside the dataset. Then, the main problem corresponding to the upper virtual power plant model is solved to obtain the optimal transaction price and transaction volume. However, the transaction volume at this time is the predicted value based on the Kriging model rather than the true value, so the optimal transaction price needs to be substituted into the lower user model to obtain the true transaction volume, and the optimal transaction price and the true transaction volume corresponding thereto are added to the initial sample dataset as high-quality dataset to realize the modification of the Kriging model. Finally, the sample data is dynamically updated through multiple iterations until the objective function value of the main problem converges.

[0133] S403, decompose the distribution robust optimization problem of the planning and running two stages into a main problem and a sub-problem, embed the principal-agent model into the main problem, and solve the problem by using column and constraint generation algorithm for staged iteration.

[0134] The main problem and the sub-problem obtained by decomposition are as follows:

[0135] (1) Main problem:

[0136]

[0137] In the formula, l represents the iteration number, l∈{1, 2, 3, …, N l}, wherein N l is the total iteration number.

[0138] (2) Sub-problem;

[0139]

[0140] The max-min double-layer structure of the sub-problem includes an outer maximum problem and an inner minimum problem. According to the solving result of the main problem, the probability distribution of the worst scenario in the confidence interval is found. Since the outer discrete scenario probability π s,l and the inner running stage variable y s,l are independent of each other, the sub-problem can be solved in two steps, that is, the inner minimum problem is solved first, and then the outer maximum problem is solved, wherein the inner minimum problem can be obtained from the solving result of the main problem.

[0141] The detailed solving steps of step S403 include:

[0142] S4031: Set the lower bound L B of the model as -∞, the upper bound U B of the model as +∞, the iteration number l as 1, the convergence threshold ε r as 10 -2 , and the initial probability of each scenario as π s,0 .

[0143] S4032: solving the master problem to obtain the optimal solution and updating the lower bound

[0144] S4033: based on the optimal variable of the master problem solving the sub-problem to obtain the probability of the worst-case scenario and the variable and the objective function value of the sub-problem and updating the upper bound

[0145] S4034: if U B -L B ≤ε r , stop iteration and output the optimal solution Otherwise, let l = l + 1 and return to step 2 to update the probability of each scenario and add the variable y s,l+1 and the constraint related to the variable y s,l+1 to the master problem.

[0146] Further, in step S403, the modified Kriging model algorithm, the traditional Kriging model algorithm, the particle swarm algorithm, the KKT condition embedded column and constraint generation algorithm are used to solve the master problem and the sub-problem in the column and constraint generation algorithm, Figure 3 for the comparison of the convergence performance of the method of the present application and other methods. When the modified Kriging model algorithm designed by the present application is used in combination with the column and constraint generation algorithm, compared with the traditional Kriging model algorithm and the KKT condition, the iteration number is reduced by 37.50% and 61.54%, respectively, and when the maximum iteration number is reached, the particle swarm algorithm still does not converge. Therefore, the method designed by the present application can effectively improve the convergence speed of solving the distributed robust optimization problem.

[0147] The above-described embodiments have described the technical solutions and beneficial effects of the present application in detail. It should be understood that the above-described embodiments are only specific embodiments of the present application and are not used to limit the present application. Any modification, supplement and equivalent replacement made within the principle range of the present application should be included in the protection range of the present application.

Claims

1. A scene-driven virtual power plant planning-operation two-stage distribution robust optimization method, characterized in that, The method comprises the following steps: S1, based on the master-slave game framework of power supply, load and energy storage, considering the capacity configuration and energy scheduling of virtual power plant, a capacity configuration model and a master-slave game model are constructed; S2, considering the correlation and uncertainty of wind turbine and photovoltaic output, an uncertainty set of wind and light output probability distribution is constructed, and a plurality of typical scenes are generated through sampling and clustering algorithm; S3, under the driving of the typical scene, the confidence interval of the uncertainty set is comprehensively norm-constrained, and a distribution robust optimization model considering uncertainty is established for the planning and operation of the virtual power plant; wherein the distribution robust optimization model comprises the capacity configuration model and the master-slave game model established in step S1; S4, based on the modified Kriging model and the column and constraint generation algorithm, the distribution robust optimization model containing the master-slave game model is solved; the specific steps are as follows: S401, the interaction between the virtual power plant and the user in the master-slave game model is displayed through the Kriging model, so that the double-layer optimization problem is simplified into a single-layer optimization problem; specifically including: Firstly, Latin hypercube sampling is used to generate initial sample points containing unit time electricity, heat and ammonia transaction price of virtual power plant and user for each discrete scene, the number of each group of initial sample points is selected as 10 times the number of transaction price variables, the transaction volume corresponding to each group of transaction price is obtained by substituting the initial sample points into the lower user model, so as to construct an initial sample data set containing transaction price and transaction volume; Then, the mapping relationship between transaction price and transaction volume in the sample data set is predicted by using Kriging model, so as to fit and replace the lower model according to the mapping relationship between the upper and lower layers, and the single-layer optimization model of each scene after fitting is as follows: In the formula, H(·) represents the mapping relationship between the upper and lower interaction obtained by using Kriging model prediction; S402, in order to reduce the influence of the number of sample data on the fitting accuracy of the model, a Kriging model modification mechanism considering dynamic updating of sample data is designed; specifically including: Firstly, the initial sample data set of each scene is constructed by uniform sampling and calling the lower model, and the initial Kriging model is established based on the mapping relationship in the data set; Then, the optimal transaction price and transaction volume are obtained by solving the master problem corresponding to the upper virtual power plant model, but the transaction volume at this time is the predicted value based on the Kriging model rather than the true value, so it is necessary to substitute the optimal transaction price into the lower user model to solve the real transaction volume, and add the optimal transaction price and the real transaction volume corresponding thereto to the initial sample data set as high-quality data set, so as to modify the Kriging model; Finally, the sample data is dynamically updated through multiple iterations until the objective function value of the master problem converges; S403, the distribution robust optimization problem of planning and operation is decomposed into master problem and sub-problem, and the master-slave game model is embedded into the master problem, and the column and constraint generation algorithm is used for stage-by-stage iteration solution; the master problem and the sub-problem obtained by decomposition are as follows: Master problem: In the formula, l represents the number of iterations, l ∈ {1, 2, 3, …, N l} l is the total number of iterations; (2) Sub-problem; The max-min bi-level structure of the sub-problem contains an outer maximum problem and an inner minimum problem, which finds the probability distribution of the worst scenario within the confidence interval according to the solving result of the main problem; since the outer discrete scenario probability π s,l and the inner operation stage variable y s,l are independent of each other, the sub-problem is divided into two steps, i.e., the inner minimum problem is solved first, and then the outer maximum problem is solved, wherein the inner minimum problem is obtained from the solving result of the main problem.

2. The scenario-driven virtual power plant planning-operation two-stage distribution robust optimization method according to claim 1, characterized in that, The specific steps of S1 are as follows: S101, a capacity configuration model is built for the aggregated distributed resources in the virtual power plant, including wind turbines, photovoltaic, combined heat and power units, biomass boilers, ammonia production systems, and energy storage systems, with the objective of minimizing the investment cost in the planning stage of the virtual power plant; S102, an energy-benefit interaction model, i.e., a master-slave game model, is established with the virtual power plant as the leader and the users as the followers; the virtual power plant formulates the energy sales price with the users in advance according to the load forecast information, and the users adjust their energy consumption through demand response mechanisms under the guidance of the price signal of the virtual power plant; the two parties interact with each other in turn to determine the pricing strategy and the energy consumption strategy, and the virtual power plant is in a dominant position and forms a master-slave game with the users; S103, in the master-slave game relationship, the upper virtual power plant trades electricity, heat, and ammonia with the lower users with the objective of minimizing the operation cost in the day-ahead stage; S104, the lower users adjust their energy consumption through demand response with the objective of maximizing the comprehensive benefits according to the price strategy issued by the upper virtual power plant in the day-ahead stage.

3. The scenario-driven virtual power plant planning-operation two-stage distribution robust optimization method according to claim 2, characterized in that, In step S101, the investment cost of the virtual power plant in the planning stage is as follows: where C inv denotes the daily average investment cost of the VPP; i denotes the type of equipment inside the VPP; Ω DG is the set of equipment inside the VPP, i e Ω DG ; C DG,i denotes the investment cost of the i-th type of equipment inside the VPP; N DG,i is the number of installations of the i-th type of equipment; r inv denotes the discount rate; TL i denotes the economic useful life of the i-th type of equipment; l inv,i is the unit investment cost of the i-th type of equipment.

4. The scenario-driven virtual power plant planning-operation two-stage distribution robust optimization method according to claim 3, characterized in that, In step S102, the objective function and the constraint conditions of the master-slave game model are as follows: minC op (λ et ,λ ht ,λ at ,P et ,Q ht ,B at ) In the formula, Ω vpp and Ω user respectively represent the master-slave game strategy set of the virtual power plant and the user; C op and C eu respectively represent the objective function of the virtual power plant and the user in the day-ahead stage; λ et , λ ht , λ at respectively represent the electricity, heat, and ammonia transaction prices of the virtual power plant and the user; P et , Q ht , and B at respectively represent the electricity, heat, and ammonia transaction quantities of the virtual power plant and the user.

5. The scenario-driven virtual power plant planning-operation two-stage distribution robust optimization method according to claim 4, characterized in that, The specific steps of S2 are as follows: S201, based on the kernel density estimation method, the historical sample data of wind and light output of the virtual power plant are analyzed, and the probability distribution model of wind and light output is fitted, so as to deeply mine the wind and light distribution characteristics of the area where the virtual power plant is located; S202, since the wind power and photovoltaic output are complementary in some time periods and show a negative correlation, the Frank-Copula function which takes into account the non-negative and negative correlation between random variables is selected to describe the correlation between wind and light; S203, using Copularnd function to sample and generate a certain number of edge distribution function values (u, v) satisfying the Frank-Copula function relationship, and obtaining the wind-solar output scene (x, y) of each period satisfying the joint distribution function by performing inverse transformation on the edge distribution function values of each period of the wind-solar power. S204, in order to ensure the fitting accuracy of the Frank-Copula function, the K-means clustering algorithm is used to cluster the sampling results, generate multiple typical scenarios, and obtain the probability of occurrence of each typical scenario.

6. The scenario-driven virtual power plant planning-operation two-stage distribution robust optimization method according to claim 5, characterized in that, The specific steps of S3 are as follows: S301, based on the multiple typical scenarios reflecting the uncertainty and correlation of wind and light output generated in step S2, the probability distribution uncertainty set in the worst scenario is optimized, and the planning and operation two-stage decision scheme under the worst scenario is obtained; wherein, the distribution robust optimization model of the virtual power plant in the planning and operation two stages includes the capacity configuration model and the master-slave game model, and the objective function is as follows: where x and y s represent the decision variables in planning and operational stages respectively; A is the coefficient matrix of the variables in the planning stage objective function; B and C are the coefficient matrices of the quadratic and linear terms in the operational stage objective function respectively; d is the constant term coefficient of the operational stage objective function; X is the set of planning stage variables; Ω π is the uncertainty set of wind and solar power output probability distribution; s is the typical scenario number; N s is the number of typical scenarios; π s represents the probability of scenario s occurring; Y(x, π s ) is the set of operational stage variables, representing the coupling relationship between x and y s in scenario s. S302, the uncertainty set of the probability distribution of wind and light output and the confidence constraint satisfied by the probability distribution of each scenario are as follows: where R + denotes a positive real number; π s,0 denotes the initial probability of a typical scenario; θ1and θ ∞ correspond to the allowed probability fluctuation bias under 1-norm and ∞-norm constraints, respectively; Pr(·) denotes the probability distribution function; S is the initial number of scenarios.

Images