A bi-level distribution robust scheduling method for virtual power plant considering shared energy storage

Abstract

A kind of virtual power plant double-layer distribution robust scheduling method considering shared energy storage belongs to the field of power system optimization scheduling, steps are as follows: step 1, based on R-fig Copula and the upper layer global robust scheduling of distribution robust optimization, determine global scheduling plan;Step 2, based on the lower layer distributed collaborative scheduling of alternative direction multiplier method ADMM, execute and decompose the global scheduling plan.The present application ensures that all participants including independent shared energy storage operators can achieve profit by the above method, builds a feasible business model for multi-agent collaborative participation in grid service, and provides effective economic incentives.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system optimal scheduling, in particular to a virtual power plant double-layer distribution robust scheduling method considering shared energy storage. BACKGROUND

[0002] With the increasing penetration of renewable energy such as wind and solar in power grids, their inherent volatility and uncertainty have brought great challenges to the safe and stable operation of power systems. As an effective technical path for integrating distributed energy, controllable load and energy storage systems, virtual power plant (VPP) has become a key to improving the flexibility of power grids. At the same time, shared energy storage (SES) as a new business model can provide energy shifting services for multiple subjects, further enhancing system regulation capacity.

[0003] However, the coordinated scheduling of VPP and SES faces two major problems. First, each participating subject (such as VPP operators, SES operators, users, etc.) is an independent economic entity, and out of business confidentiality considerations, they are reluctant to share their internal private information such as costs and bids, which makes it difficult to implement traditional centralized optimization methods. Second, resources such as wind, light, and load have complex nonlinear correlations and tail dependencies under extreme events. Traditional stochastic programming or robust optimization methods either assume that uncertainty variables are independent of each other or use simple probability distribution models, which are difficult to accurately depict such complex coupling characteristics, resulting in scheduling schemes that may be too aggressive or conservative in actual application, failing to balance economic efficiency and safety.

[0004] In existing research, although distributed algorithms such as alternating direction method of multipliers (ADMM) are used to solve multi-subject collaborative optimization problems to protect privacy, most are based on deterministic models and do not fully consider the impact of uncertainty. While traditional robust optimization methods can handle uncertainty, their "worst-case" assumption often leads to overly conservative results. Distributional robust optimization (DRO) can model the uncertainty of the probability distribution itself under data-driven conditions, but its effectiveness is highly dependent on the accurate depiction of the joint distribution of uncertainty. Therefore, there is an urgent need for a VPP and SES coordinated scheduling method that can accurately model multiple uncertainties, protect the privacy of each party, and balance economic efficiency and robustness. SUMMARY

[0005] The present application aims to overcome the shortcomings of the prior art and provide a virtual power plant double-layer distribution robust scheduling method considering shared energy storage, which can accurately depict the complex correlation of source-load uncertainty, protect the business privacy of each participating party, and achieve a balance between the economic efficiency and robustness of system operation.

[0006] This invention is achieved through the following technical solution: a two-layer distributed bar scheduling method for virtual power plants considering shared energy storage, comprising the following steps:

[0007] Step 1: Use upper-level global robust scheduling based on R-Teng Copula and distributed robust optimization to determine the global scheduling plan.

[0008] Step 1.1: Construct a source-load joint probability distribution model: Collect historical data of wind power, photovoltaic power and load, and use R-Teng Copula theory to construct a high-dimensional joint probability distribution model that can capture the nonlinearity, asymmetry and tail correlation characteristics between variables;

[0009] To describe the complex dependency among wind power, solar power, and load, we first introduce Copula theory. According to Sklar's theorem, any multidimensional joint distribution function is expressed as a combination of its marginal distribution functions and a Copula function. Let the three random variables—wind power, solar power, and load—be denoted as follows: Joint probability density function Decomposed into:

[0010]

[0011] in, For the first Marginal probability density function of each variable; For the first Marginal cumulative distribution function of variables The result of the conversion is A variable that is uniformly distributed over an interval; It is the Copula density function, used to describe the dependency structure between variables;

[0012] To capture asymmetry and tail correlation, R-vine Copula is employed. R-vine Copula decomposes the high-dimensional Copula density function into a product of a series of paired two-dimensional Copulas through a multi-layered tree structure. For a three-dimensional variable, its R-vine Copula density function decomposes as follows:

[0013]

[0014] in, and It is a pairing copula of the first-level tree structure, which describes the pairs of variables. and Direct dependencies between them; It is a pairing copula in a second-level tree structure, describing the variables Given conditions, variables and The conditional dependencies between them; and The tree structure is determined using a data-driven approach, with the conditional distribution function as the basis, and the optimal family of Copula functions and parameters are selected for each pair.

[0015] Step 1.2: Generate and reduce typical scenarios: Based on the constructed R-vine Copula model, Monte Carlo sampling is performed to generate random scenarios that can reflect the true characteristics of the source load. Then, the K-means clustering algorithm is used to reduce the massive number of scenarios into a group of representative typical scenarios and initial probabilities.

[0016] Based on the constructed R-vine Copula model A second Monte Carlo sampling is performed to generate a combined wind-solar-load scene. To reduce the computational complexity of subsequent optimization, the K-means clustering algorithm is used to cluster the generated scenes. The scenario was reduced to A representative typical scenario ,in For scene indexing, Indexed by time period, for each typical scenario Represents a cluster center with its initial probability It is determined by the proportion of the number of samples included in the cluster to the total number of samples.

[0017] Step 1.3: Construct and solve a two-stage robust DRO model: Establish a DRO model with the goal of minimizing the total upper-level scheduling cost. This model searches for the worst-case scenario probability distribution within a probability fuzzy set defined by the 1-norm and the ∞-norm, and minimizes the expected operating cost based on this distribution to obtain a global scheduling plan that balances economy and robustness. The global scheduling plan includes the globally optimal energy interaction power between the virtual power plant and the shared energy storage.

[0018] Step 1.3.1, Model Construction:

[0019] A two-stage DRO model is established, in the following form:

[0020]

[0021] in, The first-stage decision variable represents the scheduling decision determined previously and independent of the specific scenario; the feasible region is... ; For the cost of the first phase, For the scene The true probability, all The constructed probability vector Constrained in a probabilistic fuzzy set Inside;

[0022] The cost function for the second stage, denoted as is determined and the scenario occurs, the minimum operation cost by the optimal real-time dispatch is

[0023]

[0024] where is the decision variable of the second stage, also called recourse variable, is the real-time adjustment decision under the scenario ; is the cost coefficient vector of the second stage; is the feasible region of the second stage decision, defined by operation constraints, including

[0025] Power balance constraints:

[0026]

[0027] where is the output of the conventional generator in the time period; and are the actual grid-connected power of wind and photovoltaic, respectively; and are the discharging and charging power of the shared energy storage, respectively; and are the power purchase and power sale to the grid, respectively; is the actual load in the time period;

[0028] Energy storage operation constraints:

[0029]

[0030] where is the state of charge of the energy storage at the end of the time period; and are the charging and discharging efficiencies, respectively; is the dispatch time interval, the state of charge and the charging and discharging power of the energy storage need to satisfy the capacity and power upper and lower bound constraints;

[0031] Other operation constraints: including the output upper and lower bounds and ramping constraints of the conventional generators, the exchange power limits with the grid, the constraint that the output of the renewable energy is no more than its prediction, and the adjustment constraints of the demand response load;

[0032] The probability fuzzy set characterizes the real probability distribution deviation range between empirical probability distribution , the fuzzy set is jointly constrained by 1-norm and ∞-norm:

[0033]

[0034] wherein, and are the deviation radii determined according to the preset confidence level, used to limit the fluctuation range of the probability distribution.

[0035] Step 1.3.2, model solving:

[0036] The two-stage distribution robust optimization problem is solved iteratively by using the column and constraint generation C&CG algorithm, which is decomposed into a main problem M and a sub-problem to be solved alternately until the upper and lower bounds converge.

[0037] Step 2, lower-layer distributed collaborative scheduling based on alternating direction multiplier method ADMM, used to execute and decompose the global scheduling plan:

[0038] Step 2.1, constructing local optimization model of each participant: a local optimization model with the objective of optimal self-operation cost is established for each independent aggregation unit of the operator, including its own internal device constraints and energy interaction variables with other units;

[0039] The entire system is decomposed into multiple independent aggregation units, each unit The objective is to minimize its local cost function Each unit is connected to each other through energy exchange, and the interaction power between unit and unit is , and needs to satisfy the consistency constraint: .

[0040] Step 2.2, establishing augmented Lagrangian function and iterative solving: the global optimal energy interaction power determined in step 1.3 is taken as the consistency constraint of the lower-layer distributed optimization, the augmented Lagrangian function of the system is constructed, and the ADMM algorithm is used for iterative solving, the global scheduling plan is decomposed into the optimal operation strategy of each unit, until the algorithm converges.

[0041] To solve this distributed optimization problem, an augmented Lagrangian function is constructed, and the ADMM algorithm is used for iteration, in the q+1 th iteration:

[0042] Local optimization step, i.e. the update of : each aggregation unit , given the interaction powers of other units and the current Lagrange multiplier , solve the local optimization problem to update its internal decision variables and the planned interaction power The optimization objective function is:

[0043]

[0044] where is the coefficient of the quadratic penalty term, in this step, each unit only uses its own private information to make decisions.

[0045] The price update step, i.e. The update of : After all units complete the local optimization, the coordination center collects the planned interaction power of each unit, and updates the Lagrange multiplier according to the violation degree of the consistency constraint, i.e. the residual :

[0046]

[0047] The updated multiplier will be sent to each unit as the price signal for the next round of iteration;

[0048] The above steps are repeated until the primal residual and the dual residual are both less than the preset threshold, indicating that the algorithm converges and all parties reach a consensus on the energy exchange plan.

[0049] Compared with the prior art, the beneficial effects of the present application are: the present application more accurately characterizes the complex nonlinearity and tail correlation between wind power, photovoltaic and load, provides more reliable decision support for extreme uncertainty scenarios, and significantly improves the balance between the robustness and economy of the virtual power plant dispatching scheme. While ensuring the safe operation of the system, the present application effectively reduces the total operating cost of the system, compared with the traditional non-coordinated robust model, the total operating cost of the virtual power plant can be reduced by 12.6%, and the renewable energy rejection rate can be reduced to 2.3%, thereby realizing more than 95% of new energy consumption level.

[0050] In addition, in the process of realizing collaborative optimization among each independent market subject, the present application does not need each participant to share its internal cost function, operating constraints and other business sensitive data, thereby effectively protecting the business privacy of each party. By establishing a mutually beneficial and win-win collaborative mechanism, the present application ensures that all participants including independent shared energy storage operators can achieve profits, builds a feasible business model for multi-subject collaborative participation in power grid services, and provides effective economic incentives. BRIEF DESCRIPTION OF DRAWINGS

[0051] The application will be further described below in conjunction with the accompanying drawings and specific embodiments. The scope of protection of the application is not limited to the following content.

[0052] Figure 1 The schematic diagram of the double-layer distributed robust scheduling framework proposed by the application.

[0053] Figure 2 The schematic diagram of the tree structure of the wind-light-load three-dimensional joint distribution constructed for the R-Frank Copula model.

[0054] Figure 3 The schematic diagram of the convergence process of the ADMM algorithm in the lower-layer distributed scheduling, showing the iteration curves of the cost / profit of each aggregation unit.

[0055] Figure 4 The optimization scheduling results of the virtual power plant and shared energy storage obtained by the method of the application under a typical severe scenario. DETAILED DESCRIPTION

[0056] The application will be further described below in conjunction with the accompanying drawings and specific embodiments. The scope of protection of the application is not limited to the following content. The application proposes a virtual power plant double-layer distributed robust scheduling method considering shared energy storage, the overall technical process of which is shown in Figure 1 , which is divided into two levels of upper-layer global robust optimization and lower-layer distributed collaborative optimization.

[0057] First stage: upper-layer global robust optimization

[0058] The goal of this stage is to develop a global day-ahead scheduling plan that balances economic efficiency and robustness at the system level, considering the worst-case source-load uncertainty scenario.

[0059] Step 1: Uncertainty modeling and scenario generation

[0060] 1.1 Construct joint probability distribution

[0061] To accurately describe the complex dependence relationship among wind power, photovoltaic and load, the Copula theory is first introduced. According to the Sklar theorem, any multi-dimensional joint distribution function can be represented as a combination of its marginal distribution function and a Copula function. Let the wind power, photovoltaic and load be denoted as , the joint probability density function can be decomposed as:

[0062]

[0063] where, is the marginal probability density function of the th variable; is the the edge cumulative distribution function (CDF) of a variable transformed to be uniformly distributed over the interval [0, 1]; the edge cumulative distribution function (CDF) of a variable is the Copula density function, which describes the dependence structure among variables.

[0064] To flexibly capture asymmetric and tail dependence, the present application employs R-vine Copula. R-vine Copula decomposes the high-dimensional Copula density function into a product of a series of bivariate pair-Copula through a multi-layer tree structure (as shown in FIG. 1). For three-dimensional variables, the R-vine Copula density function can be decomposed as: Figure 2

[0065]

[0066] where, and are the pair-Copula of the first layer tree structure, which respectively describe the direct dependence between variable pairs and is the pair-Copula of the second layer tree structure, which describes the conditional dependence between variables and and are the conditional distribution functions. The tree structure is determined in a data-driven manner, and the optimal Copula function family (such as Frank, Clayton, etc.) and parameters are selected for each pair.

[0067] 1.2 Scenario generation and reduction

[0068] Based on the constructed R-vine Copula model, a large number of high-fidelity wind-solar-load joint scenarios are generated through times (for example, ) Monte Carlo sampling. To reduce the computational complexity of subsequent optimization, the K-means clustering algorithm is used to reduce the generated scenarios to (typically ) representative typical scenarios , where is the scenario index, and is the time period index. Each typical scenario represents a cluster center, and the initial probability is determined by the proportion of the number of samples contained in the cluster to the total number of samples.

[0069] ​​​​Step 2: Two-stage distribution robust optimization (DRO) model

[0070] 2.1 Model formulation

[0071] The two-stage DRO model is established, and its compact form is as follows:

[0072] where, is the first-stage decision variable, representing the day-ahead determined, scenario-independent scheduling decision, such as the start-stop state of the generator unit, and its feasible region is ; is the first-stage cost. is the true probability of scenario , and all constitute the probability vector , which is constrained within a probability fuzzy set .

[0073] is the second-stage cost function, representing the minimum operating cost generated by the optimal real-time scheduling when the first-stage decision is determined and scenario occurs, and its expression is: .

[0074] where, is the second-stage decision variable, also known as the recourse variable, which is the real-time adjustment decision under scenario , such as the actual grid-connected power of renewable energy, the charging and discharging power of energy storage, and the exchange power with the grid, etc.; is the second-stage cost coefficient vector; is the feasible region of the second-stage decision, defined by a series of operating constraints, mainly including:

[0075] Power balance constraint:

[0076] where, is the output of the conventional generator unit in period; and are the actual grid-connected powers of wind and photovoltaic, respectively; and are the discharging and charging powers of shared energy storage, respectively; and are the power purchase and power sale to the grid, respectively; is the actual load in period.

[0077] Energy storage operating constraint:

[0078]

[0079] where, is the state of charge at the end of the period; and are the charging and discharging efficiencies, respectively; is the dispatch time interval. Meanwhile, the state of charge and the charging / discharging power of the energy storage should satisfy its capacity and power upper and lower bound constraints.

[0080] Other operational constraints: including the upper and lower bounds of the conventional generators' output, the exchange power limit with the grid, the upper bound of the renewable energy output, and the regulation constraint of the demand response load.

[0081] Probability fuzzy set The definition of the probability fuzzy set is the key of the present application, which characterizes the deviation range between the true probability distribution and the empirical (initial) probability distribution The fuzzy set is jointly constrained by the 1-norm and the ∞-norm:

[0082]

[0083] where, and are the deviation radii determined according to the preset confidence level, which are used to limit the fluctuation range of the probability distribution, so as to ensure the robustness of the scheduling scheme to the error of the probability distribution.

[0084] 2.2 Model solving

[0085] The above three-layer (min-max-min) optimization problem is solved iteratively by using the column and constraint generation (C&CG) algorithm, which is decomposed into a master problem (Master Problem) and a subproblem (Subproblem) to be solved alternately until the upper and lower bounds converge.

[0086] Second stage: lower-layer distributed optimization

[0087] The upper-layer DRO model gives the global day-ahead scheduling plan, and the lower layer executes the plan in a distributed manner that protects privacy through the ADMM algorithm.

[0088] Step 3: ADMM distributed coordination

[0089] 3.1 Problem decomposition and consistency constraints

[0090] The entire system is decomposed into multiple independent aggregation units (participants), such as VPPs, SESs, etc. Each unit k The objective of each unit is to minimize its local cost function Each unit is connected to others through energy exchange, denoted as unit The interaction power between unit and unit is denoted as To ensure the overall power balance of the system, the consistency constraint needs to be satisfied:

[0091] 3.2 ADMM iteration process

[0092] To solve the distributed optimization problem, an augmented Lagrangian function is constructed and iterated through ADMM algorithm. In the q+1 th iteration, the following steps are performed:

[0093] Local optimization step (update): Each aggregation unit , given the interaction power of other units and the current Lagrangian multiplier (i.e., the internal transaction price) , solves a local optimization problem to update its internal decision variables and planned interaction power . The optimization objective function is:

[0094] where is the coefficient of the quadratic penalty term. In this step, each unit only uses its own private information to make decisions.

[0095] Price update step (update): After all units complete the local optimization, the coordination center collects the planned interaction power of each unit and updates the Lagrangian multiplier according to the degree of violation of the consistency constraint (i.e., the residual ):

[0096] The updated multiplier will be sent to each unit as the price signal for the next iteration.

[0097] The above steps are repeated until the primal residual and dual residual are both less than the preset threshold, indicating that the algorithm converges and all parties reach a consensus on the energy exchange plan. In this way, the invention realizes the distributed collaborative execution of the upper robust scheduling plan while protecting the commercial privacy of each participant.

[0098] Simulation results and analysis

[0099] ​​To verify the effectiveness of the method, a simulation model is built using MATLAB R2021a software and Gurobi solver. The example system includes a virtual power plant (VPP) operator and an independent shared energy storage (SES) operator. The VPP aggregates a 20MW wind farm, a 15MW photovoltaic power station, a 10MW controllable distributed generator and various loads. The SES operator manages a 5MW / 20MWh battery energy storage system. The dispatching period is 24 hours, with a resolution of 1 hour. The interaction of the system with the external power grid follows the peak-valley time-of-use price mechanism.

[0100] As shown in Figure 2 , the application first analyzes historical data using the R Vine Copula model to obtain a tree structure of the dependence relationship between wind power (node 1), photovoltaic power (node 2) and load (node 3). This structure can accurately capture the complex nonlinearity and tail correlation between variables, providing more accurate assessment of extreme event risk than traditional models, and providing high-fidelity scenario input for subsequent upper-level distribution robust optimization.

[0101] As shown in Figure 3 , the figure shows the convergence process of the ADMM algorithm in the lower-level distributed dispatching of the application. As can be seen from the figure, the operating cost of the VPP operator and the profit of the SES operator both converge quickly and reach stability within 30 iterations. This result shows that the distributed collaborative mechanism used in the application has high computational efficiency, and can guide multiple independent subjects to quickly reach a globally near-optimal and mutually beneficial dispatching agreement without exchanging their internal private information.

[0102] As shown in Figure 4 , the figure shows the final optimization dispatching result of the system under a typical worst-case scenario identified by the upper-level distribution robust optimization model. During the low-load, low-price and high-renewable-energy-output period (e.g. 0:00-7:00), the shared energy storage is charged to absorb the surplus green power. During the peak-load, high-price period (e.g. 8:00-11:00 and 17:00-21:00), the shared energy storage is discharged to meet the peak load demand, thereby reducing the demand for the VPP to purchase high-priced power from the grid or start high-cost generator sets. This charging and discharging behavior clearly demonstrates the synergistic complementary relationship between the VPP and the SES achieved by the method of the application, i.e. the SES provides critical flexibility support for the VPP, while the SES itself also achieves profitability through energy time-shifting arbitrage.

[0103] The simulation data comparison and analysis show that, compared with the traditional robust optimization scheduling scheme without cooperation with the SES, the total operation cost of the VPP can be reduced by 12.6% by using the method of the application, and the curtailment rate of renewable energy can be greatly reduced from 6.7% to 2.3%, and a new energy consumption rate of more than 95% is achieved. This proves that the method proposed in the application can perfectly balance the operation economy of the system and the robustness when facing uncertainty, and a significant effect better than the prior art is achieved.

[0104] It can be understood that the above specific description of the application is only used to illustrate the application and is not limited to the technical solutions described in the embodiments of the application. Those skilled in the art should understand that the application can still be modified or replaced equivalently to achieve the same technical effect; as long as the use needs are met, it is within the protection scope of the application.

Claims

1. A two-layer distributed bar scheduling method for virtual power plants considering shared energy storage, characterized in that, Includes the following steps: Step 1: Determine the global scheduling plan based on upper-level global robust scheduling using R-vine Copula and distributed robust optimization: Step 1.1: Construct a source-load joint probability distribution model: Collect historical data of wind power, photovoltaic power and load, and use R-Teng Copula theory to construct a high-dimensional joint probability distribution model that can capture the nonlinearity, asymmetry and tail correlation characteristics between variables; Step 1.2: Generate and reduce typical scenarios: Based on the constructed R-vine Copula model, Monte Carlo sampling is performed to generate random scenarios that can reflect the true characteristics of the source load. Then, the K-means clustering algorithm is used to reduce the massive number of scenarios into a group of representative typical scenarios and initial probabilities. Step 1.3: Construct and solve a two-stage robust DRO model: Establish a DRO model with the goal of minimizing the total upper-level scheduling cost. This model searches for the worst-case scenario probability distribution within a probability fuzzy set defined by the 1-norm and the ∞-norm, and minimizes the expected operating cost based on this distribution to obtain a global scheduling plan that balances economy and robustness. The global scheduling plan includes the globally optimal energy interaction power between the virtual power plant and the shared energy storage. Step 1.3.1, Model Construction: A two-stage DRO model is established, in the following form: in, The first-stage decision variable represents the scheduling decision determined previously and independent of the specific scenario; the feasible region is... ; For the cost of the first phase, For the scene The true probability, all The constructed probability vector Constrained in a probabilistic fuzzy set Inside; The cost function for the second stage represents the decision made in the first stage. The scene is already set. When this occurs, the minimum operating cost generated through optimal real-time scheduling is expressed as: in, These are the decision variables for the second stage, also known as follow-up variables, and are used in the scenario. Real-time adjustment of decisions; This represents the cost coefficient vector for the second stage. The feasible region for the second-stage decision is defined by operational constraints and includes: Power balance constraints: in, For conventional generator sets exist Efforts during a specific time period; and These are the actual grid-connected power of wind power and solar power, respectively; and These are the discharge and charging power of the shared energy storage, respectively. and These represent the power purchased from and sold to the power grid, respectively. for Actual load during the time period; Energy storage operation constraints: in, for State of charge of energy storage at the end of the period; and These are the charging and discharging efficiencies, respectively. For the scheduling time interval, the state of charge and charging / discharging power of energy storage must meet the upper and lower limits of capacity and power. Other operational constraints include the upper and lower limits of output and ramping constraints of conventional generator sets, the power exchange limit with the grid, the constraint that the output of renewable energy should not exceed its predicted value, and the adjustment constraints of demand response load. probabilistic fuzzy sets It depicts the true probability distribution Compared with empirical probability distribution The range of deviations between them is constrained by both the 1-norm and the ∞-norm of the fuzzy set. in, and It is a deviation radius determined according to a preset confidence level, used to limit the fluctuation range of the probability distribution; Step 1.3.2, Model Solving: The column and constraint generation C&CG algorithm is used to iteratively solve the above two-stage sub-Bruker optimization problem. It is decomposed into a main problem M and a sub-problem to be solved alternately until the upper and lower bounds converge. Step 2: Based on the alternating direction multiplier method (ADMM), the lower-level distributed cooperative scheduling executes and decomposes the global scheduling plan: Step 2.1: Construct local optimization models for each participant: Establish a local optimization model for each independent aggregation unit of the operator with the goal of optimizing its own operating cost, including its internal equipment constraints and energy interaction variables with other units; The specific method in step 2.1 is as follows: The entire system is decomposed into multiple independent aggregation units, each unit The goal is to minimize its local cost function The units are interconnected through energy exchange, and are denoted as units. With unit The interaction power between them is And it needs to satisfy consistency constraints: ; Step 2.2: Establish the augmented Lagrangian function and solve iteratively: Using the global optimal energy interaction power determined in Step 1.3 as the consistency constraint for the lower-level distributed optimization, construct the augmented Lagrangian function of the system, and solve iteratively using the ADMM algorithm to decompose the global scheduling plan into the optimal operating strategy of each unit until the algorithm converges.

2. The virtual power plant two-layer distributed bar scheduling method considering shared energy storage as described in claim 1, characterized in that: The specific method in step 1.1 is as follows: To describe the complex dependency among wind power, solar power, and load, we first introduce Copula theory. According to Sklar's theorem, any multidimensional joint distribution function is expressed as a combination of its marginal distribution functions and a Copula function. Let the three random variables—wind power, solar power, and load—be denoted as follows: Joint probability density function Decomposed into: in, For the first Marginal probability density function of each variable; For the first Marginal cumulative distribution function of variables The result of the conversion is A variable that is uniformly distributed over an interval; It is the Copula density function, used to describe the dependency structure between variables; To capture asymmetry and tail correlation, R-vine Copula is employed. R-vine Copula decomposes the high-dimensional Copula density function into a product of a series of paired two-dimensional Copulas through a multi-layered tree structure. For a three-dimensional variable, its R-vine Copula density function decomposes as follows: in, and It is a pairing copula of the first-level tree structure, which describes the pairs of variables. and Direct dependencies between them; It is a pairing copula in a second-level tree structure, describing the variables Given conditions, variables and The conditional dependencies between them; and The tree structure is determined using a data-driven approach, with the conditional distribution function as the basis, and the optimal family of Copula functions and parameters are selected for each pair.

3. The virtual power plant two-layer distributed bar scheduling method considering shared energy storage as described in claim 1, characterized in that: The specific method in step 1.2 is as follows: Based on the constructed R-vine Copula model A second Monte Carlo sampling is performed to generate a combined wind-solar-load scene. To reduce the computational complexity of subsequent optimization, the K-means clustering algorithm is used to cluster the generated scenes. The scenario was reduced to A representative typical scenario ,in For scene indexing, Indexed by time period, for each typical scenario Represents a cluster center with its initial probability It is determined by the proportion of the number of samples included in the cluster to the total number of samples.

4. The virtual power plant two-layer distributed bar scheduling method considering shared energy storage as described in claim 1, characterized in that: The specific method in step 2.2 is as follows: To solve the distributed optimization problem, an augmented Lagrangian function is constructed, and the ADMM algorithm is used for iteration. In the (q+1)th iteration: Local optimization steps, namely Update: Each aggregation unit Given the interaction power of other units and the current Lagrange multipliers, In this case, solve the local optimization problem to update its internal decision variables. and planned interaction power The objective function for optimization is: in, It is the coefficient of the second-order penalty term. In this step, each unit only uses its own private information to make decisions. Price update steps, namely Update: After all units have completed local optimization, the coordination center collects the planned interaction power of each unit and updates it according to the degree of violation of consistency constraints, i.e., the residual. To update the Lagrange multipliers : The updated multipliers will be sent to each unit as the price signal for the next iteration; The above steps are repeated until both the primal residual and the dual residual are less than the preset threshold, indicating that the algorithm has converged and all parties have reached a consensus on the energy exchange plan.

Images