Virtual power plant distributed cooperative regulation and control method and system based on consensus ADMM and graph convolutional network

Abstract

The invention discloses a virtual power plant distributed cooperative regulation and control method and system based on a consensus ADMM and a graph convolutional network, and the method comprises the following steps: constructing a global optimization problem of aggregation regulation and control of a virtual power plant, the target being to minimize the total operation cost, and constraints comprising total power balance and individual power limitation; decomposing the global optimization problem into local optimization sub-problems of the distributed resource intelligent terminals by adopting a consensus ADMM framework; each intelligent terminal generates a local power decision function through a pre-trained graph convolutional network GCN based on the local state information, the dual variables and the neighbor information; the intelligent terminals exchange dual variables and consistency variables through a communication coordination layer, and iterative updating is carried out until convergence; and the output execution layer converts the final decision into an actual control instruction and issues and executes the actual control instruction. By fusing the consensus ADMM and the graph convolutional network, the method shows significant performance advantages in multiple aspects in virtual power plant distributed cooperative regulation and control.

Description

TECHNICAL FIELD

[0001] The present application relates to a virtual power plant, in particular to a virtual power plant distributed collaborative regulation method and system based on consensus ADMM and graph convolution network. BACKGROUND

[0002] The existing virtual power plant (VPP) aggregation regulation technology mainly relies on centralized optimization architecture, which has significant technical defects. First, the central server needs to process massive distributed resource data, resulting in huge computing and communication pressure and poor system scalability. Second, centralized control has a single point of failure risk. Once the central node fails, the entire system will be paralyzed. In addition, the existing optimization algorithm relies heavily on accurate distributed resource cost function models, but in reality, user-side load, energy storage loss, etc. The model is difficult to accurately obtain, resulting in model mismatch and decreased regulation performance. Finally, centralized data collection poses a risk of user data privacy leakage. Globally, the research on virtual power plants has developed into an intersection of power systems, information and communication technology, and artificial intelligence. With the increasing penetration of distributed energy, it is particularly important to achieve efficient and reliable aggregation regulation. Accurate regulation not only improves the power grid's ability to absorb renewable energy and enhances system stability, but also provides flexible regulation resources for the electricity market. Current research not only involves the development and optimization of regulation models, but also includes the exploration of communication topology adaptability, algorithm convergence, and system practicality. Through these studies, scholars hope to build a more intelligent and reliable distributed energy management system. SUMMARY

[0003] The purpose of the present application is to provide a virtual power plant distributed collaborative regulation method and system based on consensus ADMM and graph convolution network to solve the problems raised in the background.

[0004] To achieve the above purpose, the present application provides the following technical solutions: The virtual power plant distributed collaborative regulation method based on consensus ADMM and graph convolution network comprises the following steps: Construct a global optimization problem for virtual power plant aggregation regulation, with the goal of minimizing total operating cost and constraints including total power balance and individual power limits; Use the consensus ADMM framework to decompose the global optimization problem into local optimization sub-problems for each distributed resource intelligent terminal; Each intelligent terminal generates a local power decision function based on local state information, dual variables, and neighbor information through a pre-trained graph convolution network GCN; Intelligent terminals exchange dual variables and consistency variables through the communication coordination layer and iteratively update until convergence; The output execution layer converts the final decision into actual control instructions and issues them for execution.

[0005] As a preferred scheme of the present application: the global optimization problem objective function is expressed as: Wherein represents the power adjustment amount of the i-th resource, represents the cost function of the resource; The constraint condition is: represents the sum of the power adjustment amounts of all resources must be equal to the total power regulation instruction issued by the power grid dispatching center ; represents the power adjustment amount of each resource must be within its physical limit range, wherein and respectively represent the minimum and maximum adjustable power of the i-th resource, and N represents the total number of distributed resources participating in regulation.

[0006] As a preferred scheme of the present application: the method for decomposing the global problem into local optimization sub-problems of each distributed resource intelligent terminal by using the consensus ADMM framework is as follows: Introducing auxiliary variables , the global optimization problem is converted into an equivalent consensus form, and the global optimization problem is re-expressed as: The corresponding augmented Lagrangian function is constructed as: Wherein is the dual variable, is the penalty parameter, used to control the penalty strength of constraint violation; The ADMM algorithm is solved by the following iteration steps (1) updating the power adjustment amount : (2) updating the auxiliary variable : (3) updating the dual variable : .

[0007] As a preferred scheme of the present application: the local power decision function generated by the graph convolution network GCN is: Wherein a decision function represented by the G network representation, is a current dual variable, is state information of the resource itself, represents information of a neighbor node, is a network parameter of a graph convolution network GCN.

[0008] As a preferred scheme of the present application: the graph convolution network GCN network structure satisfies: wherein is an adjacency matrix with a self-loop, is a corresponding degree matrix.

[0009] As a preferred scheme of the present application, the graph convolution network GCN model is optimized in a targeted manner, a new dynamic graph structure learning model is established by introducing an attention mechanism, and the solving details of the core problem are optimized: The graph convolution network GCN introduces a graph attention network (GAT) layer on the basis of a standard architecture, dynamically learns the coupling weight between nodes, replaces the pre-defined static adjacency matrix, and a new decision function is represented as: ** wherein represents a dynamic coupling coefficient between node i and neighbor node j calculated by the attention mechanism, is a parameter shared by the attention mechanism, and represents a splicing operation; the mechanism enables the model to adaptively learn the complex nonlinear coupling relationship between distributed resources, improves the representation ability of the system dynamic characteristics, and thus optimizes the accuracy and efficiency of the collaborative regulation at the model level.

[0010] As a preferred scheme of the present application: the communication coordination layer supports a distributed communication topology, and the communication frequency is synchronized with the ADMM iteration period.

[0011] As a preferred scheme of the present application: the distributed resources include one or more of a photovoltaic inverter, a wind power converter, an energy storage converter, and a load controller.

[0012] Further, the present application provides a virtual power plant distributed collaborative regulation system based on consensus ADMM and a graph convolution network, comprising: a plurality of intelligent terminals respectively connected to the distributed resources, for local state perception and decision generation; a communication coordination module for managing data exchange between the intelligent terminals; a distributed consensus module for updating global consistency variables through neighbor communication; An output execution module is configured to convert the optimization result into a control instruction and issue the control instruction for execution.

[0013] As a preferred scheme of the present application, the intelligent terminal is provided with a graph convolution network (GCN) model, which is configured to generate a power decision based on local state information, dual variables and neighbor information.

[0014] Compared with the prior art, the present application has the following beneficial effects: The present application has the following advantages: 1. The present application combines consensus ADMM and graph convolution network, and has significant performance advantages in virtual power plant distributed collaborative regulation. In terms of computing efficiency, the distributed architecture decomposes the global optimization problem into local sub-problems for parallel solving, greatly reducing the computing burden of the center node, improving the scalability of the system, and enabling it to adapt to the access requirements of large-scale distributed energy resources. In terms of communication overhead, the method only needs to exchange a small amount of consistency variables and dual variables between neighbor nodes, with low communication frequency and small data volume, effectively reducing the network bandwidth pressure and enhancing the practicability of the system in a communication-limited environment. In terms of model adaptability, the introduction of GCN overcomes the dependence on accurate analytical models, enabling the system to learn optimal decision strategies from data, significantly improving the robustness and regulation accuracy of the system in the presence of model uncertainty or mismatch. In terms of reliability, the distributed control structure completely eliminates the risk of single-point failure, and even if some nodes fail, the remaining nodes can continue to operate collaboratively through local communication, ensuring the high availability and resilience of the system. In addition, the localized decision mechanism effectively protects the privacy of user data, meeting the current regulatory requirements for data security and privacy protection. Simulation and actual deployment results show that the method can quickly converge to the global optimal solution or approximate optimal solution, minimizing the total operating cost while meeting the grid regulation instructions, providing a feasible technical path for efficient, reliable and intelligent regulation of virtual power plants with high proportion of renewable energy access. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The present application is a method flowchart. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0017] The method of the application utilizes a distributed optimization layer, decomposes the global target into local optimization sub-problems of each agent through a consensus ADMM algorithm framework; utilizes a local decision layer, each agent is built-in with a GCN model, and solves the local sub-problem according to the local state and neighbor information; utilizes a communication coordination layer, transmits the dual variables and consistency variables between each agent, and realizes global cooperation; utilizes an output execution layer, converts the final decision into actual control instructions and executes.

[0018] The data processing and optimization solving specifically include: Global problem construction: The aggregation regulation problem of a virtual power plant (VPP) is essentially a large-scale constrained optimization problem. The core goal is to coordinate and organize the operation states of various types of distributed energy resources (DERs) within the VPP to minimize the overall operation cost or maximize the comprehensive benefits under the premise of meeting the grid dispatching requirements.

[0019] The aggregation regulation problem of the VPP is constructed as an optimization problem with global constraints.

[0020] The objective function represents the minimization of the total operation cost of all distributed resources, where represents the power adjustment amount of the i-th resource, represents the cost function of the resource.

[0021] This equality constraint represents that the sum of the power adjustment amounts of all resources must be equal to the total power regulation instruction issued by the grid dispatching center .

[0022] This inequality constraint represents that the power adjustment amount of each resource must be within its physical limit range, where and represent the minimum and maximum adjustable power of the i-th resource, respectively.

[0023] In this mathematical framework, N represents the total number of distributed resources participating in regulation, and these resources are diverse, including but not limited to: photovoltaic power generation systems, wind turbine generators, battery energy storage systems, adjustable loads, electric vehicle charging piles, etc. Each resource has its unique operating characteristics and cost structure, which makes the optimization problem highly heterogeneous and complex.

[0024] 2. Consensus ADMM solving framework Alternating Direction Method of Multipliers (ADMM) is an algorithmic framework that is well-suited for solving distributed optimization problems. It combines the advantages of dual decomposition and the strong convergence properties of the augmented Lagrangian method, making it effective for large-scale distributed optimization problems.

[0025] To apply the consensus ADMM algorithm, we first introduce auxiliary variables , transforming the original problem into an equivalent consensus form. Through this transformation, the original problem can be reformulated as: The corresponding augmented Lagrangian function is constructed as: where is the Lagrange multiplier (dual variable), is the penalty parameter, controlling the strength of the penalty for constraint violations. In this augmented Lagrangian function, two important components are introduced: first, the standard Lagrange multiplier term , which enforces the consistency constraint; second, the augmentation term , a quadratic penalty term that primarily serves to improve the convergence properties of the algorithm, especially in non-strongly convex cases. The penalty parameter plays a crucial role in the algorithm: on one hand, a larger ρ value can accelerate the penalty for constraint violations, benefiting the satisfaction of constraints; on the other hand, an excessively large ρ value can lead to the dominance of the quadratic term in the objective function, slowing down the convergence speed. In practical applications, adaptive strategies are often employed to adjust the ρ value, balancing the convergence speed and constraint satisfaction.

[0026] is the Lagrange multiplier vector, also known as the dual variable. In the dual ascent method, these multipliers record the historical information of constraint violations and adjust accordingly at each iteration. In the ADMM framework, the update of these dual variables has a clear physical meaning: they actually reflect the gap between the current solution and the feasible solution, and this gap is gradually reduced through the iterative process.

[0027] The construction of the augmented Lagrangian function has several important advantages: first, by introducing the quadratic penalty term, the algorithm's convergence is guaranteed even if the original problem is not strictly convex; second, this function form allows the problem to be naturally decomposed into multiple subproblems that can be solved in parallel; finally, the presence of the augmentation term makes the algorithm relatively robust to parameter selection, reducing the difficulty of parameter tuning.

[0028] From a computational perspective, this function form is particularly suitable for distributed optimization. Notice that all terms in the function except the objective function term are separable, which means the entire optimization problem can be completely decomposed into multiple subproblems, each involving only a single variable and the corresponding auxiliary variable. This decomposition property is the key to the effectiveness of the ADMM algorithm in handling large-scale distributed optimization problems.

[0029] In practical implementation, each term in the augmented Lagrangian function has a specific computational meaning: The term ensures that the optimization process moves towards reducing the total cost; The term serves as a guide for constraint satisfaction; and The term ensures the smoothness of the solution and numerical stability of the algorithm.

[0030] Through this carefully designed function form, the consensus ADMM algorithm can transform a complex global optimization problem into a series of relatively simple local optimization problems, which can be solved in parallel, and then through the coordination step to ensure the satisfaction of global constraints. This decomposition-coordination solving strategy not only improves computational efficiency, but also well maintains the physical meaning and mathematical properties of the problem.

[0031] The ADMM algorithm is solved through the following iterative steps: (1) Update (solved in parallel): This step allows each resource to independently solve a local optimization subproblem involving only its own variable, fully embodying the distributed nature of the algorithm.

[0032] (2) Update (with consensus constraints): This step usually requires a central coordinator or through a distributed negotiation mechanism to ensure that the global constraints are satisfied.

[0033] (3) Dual variable update: This step penalizes the deviation between the current solution and the feasible solution, ensuring that the algorithm eventually converges to an optimal solution that satisfies all constraints. The advantage of the ADMM algorithm lies in its ability to decompose a complex global optimization problem into multiple simpler subproblems, which can be solved in parallel, greatly improving computational efficiency. At the same time, the algorithm has good convergence properties, ensuring convergence to the optimal solution even in non-strongly convex cases.

[0034] 3. GCN-assisted decision making In practical applications, the cost function of distributed resources is often difficult to model accurately. This can be due to the complex nonlinear nature of the function itself or because the specific form is not disclosed due to business confidentiality. To overcome this challenge, the present invention employs a graph convolutional network to learn the optimal decision-making strategy.

[0035] A graph convolutional network is a deep learning model specifically designed to handle graph-structured data by aggregating information from a node and its neighbors through a message passing mechanism. In the context of VPP, each distributed resource can be considered a node in the graph, and the communication connections between resources form the edges of the graph. To solve the optimization problem depends on the update , a GCN is employed to learn this mapping relationship: where is the adjacency matrix with self-loops, is the corresponding degree matrix.

[0036] In each layer, the GCN performs two main operations: 1) Neighbor information aggregation through ; and 2) Feature transformation through .

[0037] Through multiple layers of graph convolution operations, each node can ultimately obtain information about its multi-hop neighbors, forming a global perception of the entire graph structure. This mechanism is particularly suitable for VPP environments because each resource's decision depends not only on its own state but also on the states of other resources in the network.

[0038] Finally, the GCN output layer gives the local decision: where denotes the decision function represented by the G network, is the current dual variable, is the resource's own state information, denotes the information of neighboring nodes, is the network parameter of the GCN.

[0039] This GCN-based decision-making method has important advantages: first, it does not require an explicit cost function model but learns the optimal decision-making strategy directly from data; second, through the information transmission of the graph structure, each decision considers global state information, ensuring the coordination of decisions; finally, once trained, the GCN can quickly perform forward propagation calculations, meeting the requirements of real-time regulation.

[0040] In the embodiment of the present application, the overall implementation process of the technical solution, the system adopts a layered distributed architecture, mainly including four core layers: a distributed optimization layer, a local decision-making layer, a communication coordination layer and an output execution layer. The layers are cooperatively optimized through close information interaction, and finally the distributed collaborative regulation and control of the virtual power plant (VPP) is completed.

[0041] In the system initialization phase, each distributed energy resource (DER) intelligent terminal (Agent) completes communication topology construction, local state initialization and algorithm parameter setting. The communication topology is represented in a graph structure, where the nodes correspond to DERs and the edges represent communication connections. Each Agent obtains its adjustable power boundary, operating state and other local information, and receives the global power regulation instructions issued by the dispatch center, as well as initializes the ADMM algorithm parameters (including the penalty parameter ρ, the dual variable λ, etc.), laying a foundation for subsequent distributed optimization iteration.

[0042] The distributed optimization layer uses the consensus ADMM framework to decompose the global optimization problem into local sub-problems for each Agent. As shown in Figure 1 , the algorithm achieves collaborative solution through the following iterative steps: first, each Agent solves the local optimization sub-problem in parallel, updating the local variable ; second, the consensus variable is updated through inter-neighbor communication or light central coordinator; finally, the dual variable is updated to penalize the constraint deviation. The entire process fully embodies the decomposition-coordination characteristics of the ADMM algorithm, ensuring the satisfaction of global constraints while fully leveraging the efficiency advantages of distributed computing.

[0043] In the local decision-making layer, each Agent is embedded with a graph convolution network (GCN) model to assist decision-making to address problems where the cost function is unknown or difficult to accurately model. The input of the GCN model structure includes the local state , the dual variable and the neighbor information , which realizes feature extraction and information aggregation through multi-layer graph convolution operations, and finally outputs the local decision variable . The training of the GCN can be done using historical data supervised learning or online reinforcement learning, with the goal of gradually approximating the optimal operating strategy. The introduction of this layer greatly enhances the system's adaptability to heterogeneous resources and the level of decision-making intelligence.

[0044] The communication coordination layer is responsible for managing information exchange between agents, only a small amount of data such as local decision variables, dual variables and consensus variables needs to be transmitted. The communication frequency is synchronized with the ADMM iteration, usually set to once every 5-10 seconds, which meets the requirements of real-time power system regulation and significantly reduces the communication load. This layer supports multiple communication topologies, including star, ring, mesh, etc., with good network adaptability.

[0045] The output execution layer converts the optimization results into actual control instructions and issues them to the local controllers of each resource for execution. This layer is also responsible for monitoring the satisfaction of global power balance constraints. If deviations are found, additional iterations or the activation of backup resources can be used to compensate, ensuring that the system always operates within a safe and controllable range.

[0046] The overall system operation process is a closed-loop optimization process: after initialization, each agent performs local decision-making and variable updating in parallel, coordinates through neighbor communication, and finally outputs and executes control instructions. The entire process iterates until the convergence condition is met or the maximum iteration number is reached, thereby achieving efficient, reliable and distributed regulation of virtual power plant resources.

[0047] This technical solution has many significant performance advantages in distributed collaborative regulation of virtual power plants by integrating consensus ADMM and graph convolutional networks. In terms of computational efficiency, the distributed architecture decomposes the global optimization problem into parallel local sub-problems, greatly reducing the computational burden of the central node, improving the scalability of the system, and enabling it to adapt to the access requirements of large-scale distributed energy resources. In terms of communication overhead, this method only needs to exchange a small amount of consensus variables and dual variables between neighbor nodes, with low communication frequency and small data volume, effectively reducing network bandwidth pressure and enhancing the practicality of the system in communication-limited environments. In terms of model adaptability, the introduction of GCN overcomes the dependence on precise analytical models, enabling the system to learn optimal decision-making strategies from data, significantly improving the robustness and regulation accuracy of the system in the presence of model uncertainty or mismatches. In terms of reliability, the distributed control structure completely eliminates the risk of single-point failure, and even if some nodes fail, the remaining nodes can still continue to operate collaboratively through local communication, ensuring the high availability and resilience of the system. In addition, the localized decision-making mechanism effectively protects user data privacy, meeting current data security and privacy protection regulatory requirements. Simulation and actual deployment results show that this method can quickly converge to the global optimal solution or near-optimal solution, minimizing total operating costs while meeting grid regulation instructions, providing a feasible technical path for efficient, reliable and intelligent regulation of virtual power plants with high proportions of renewable energy access.

[0048] It will be obvious to a person skilled in the art that the application is not limited to the details of the foregoing exemplary embodiments and can be implemented in other concrete forms without departing from the spirit or essential characteristics of the application. The embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. No reference signs in the claims should be considered as limiting the scope of the claims to the identity of the reference signs therein.

[0049] Furthermore, it should be understood that although the description is made on the basis of the embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that those skilled in the art can understand.

Claims

1. A virtual power plant distributed collaborative regulation method based on consensus ADMM and graph convolution network, characterized in that, The method comprises the following steps: A global optimization problem of virtual power plant aggregation regulation is constructed, with the objective of minimizing total operation cost, and constraints including total power balance and individual power limit; A consensus ADMM framework is used to decompose the global optimization problem into local optimization sub-problems of each distributed resource intelligent terminal; Each intelligent terminal generates a local power decision function based on local state information, dual variables and neighbor information through a pre-trained graph convolution network (GCN); Intelligent terminals exchange dual variables and consistency variables through a communication coordination layer, and iteratively update until convergence; An output execution layer converts the final decision into actual control instructions and executes them.

2. The virtual power plant distributed collaborative regulation method based on consensus ADMM and graph convolution network according to claim 1, wherein, The objective function of the global optimization problem is expressed as: wherein denotes the power adjustment amount of the i-th resource, denotes the cost function of the resource; The constraint condition is: The sum of power adjustment amounts of all resources must be equal to the total power control instruction issued by the grid dispatch center ; The power adjustment amount of each resource must be within its physical limit range, where and respectively represent the minimum and maximum adjustable power of the ith resource, and N represents the total number of distributed resources participating in the regulation.

3. The virtual power plant distributed collaborative regulation method based on consensus ADMM and graph convolution network according to claim 1, characterized in that, The method of decomposing the global problem into local optimization sub-problems of each distributed resource intelligent terminal using the consensus ADMM framework is as follows: Introducing auxiliary variables transforming the global optimization problem into an equivalent consensus form, the global optimization problem reformulated as: The corresponding augmented Lagrangian function is constructed as: wherein is a dual variable, is a penalty parameter for controlling the strength of the penalty for constraint violations; The ADMM algorithm is solved by the following iteration steps (1) updating the power adjustment amount : (2) update the auxiliary variables : (3) updating dual variables : 。 4. The virtual power plant distributed collaborative regulation method based on consensus ADMM and graph convolution network according to claim 1, characterized in that, The local power decision function generated by the graph convolution network (GCN) is: wherein denotes a decision function represented by a G network, is a current dual variable, is state information of the resource itself, denotes information of a neighbor node, is a network parameter of a graph convolution network GCN.

5. The virtual power plant distributed collaborative regulation method based on consensus ADMM and graph convolution network according to claim 4, characterized in that, The graph convolution network (GCN) network structure satisfies: wherein is the adjacency matrix with self-loops, is the corresponding degree matrix.

6. The virtual power plant distributed collaborative regulation method based on consensus ADMM and graph convolution network according to claim 5, characterized in that, The graph convolution network (GCN) model is optimized, and a new dynamic graph structure learning model is established by introducing an attention mechanism: The graph convolution network (GCN) model introduces a graph attention network (GAT) layer based on the standard architecture, dynamically learns the coupling weight between nodes, replaces the pre-defined static adjacency matrix, and the new decision function is expressed as: wherein represents the dynamic coupling coefficient between node i and neighbor node j computed by the attention mechanism; is a parameter that shares the attention mechanism, representing the concatenation operation.

7. The consensus ADMM and graph convolution network-based virtual power plant distributed collaborative regulation method according to claim 1, characterized in that, The communication coordination layer supports a distributed communication topology, and the communication frequency is synchronized with the ADMM iteration period.

8. The virtual power plant distributed collaborative regulation method based on consensus ADMM and graph convolution network according to claim 1, characterized in that, The distributed resources include one or more of photovoltaic inverters, wind power converters, energy storage converters, and load controllers.

9. The system for implementing the consensus ADMM and graph convolution network based distributed collaborative regulation method of a virtual power plant according to any one of claims 1-8, characterized in that, It comprises: A plurality of intelligent terminals are respectively connected to each distributed resource for local state sensing and decision making; A communication coordination module is used to manage data exchange between intelligent terminals; A distributed consensus module updates global consistency variables through neighbor communication; An output execution module is used to convert optimization results into control instructions and execute them.

10. The virtual power plant distributed collaborative regulation system based on consensus ADMM and graph convolution network according to claim 8, wherein, The intelligent terminal is built-in with a graph convolution network (GCN) model for generating power decisions based on local state information, dual variables and neighbor information.

Images