Lightweight cloud edge collaborative optimization scheduling method and device for virtual power plant

Abstract

The invention discloses a virtual power plant lightweight cloud edge collaborative optimization scheduling method and device, and relates to the field of virtual power plant optimization scheduling. The method comprises the following steps: dividing a virtual power plant cluster at a cloud according to the energy coupling strength of a virtual power plant; establishing a virtual power plant optimization scheduling model; modeling the virtual power plant optimization scheduling model as a Markov game model; multi-agent neural network parameters in the Markov game process are trained; configuring an edge computing unit for each cluster according to a cluster division result, and enabling each edge computing unit to receive optimal parameters of a Markov game model and a multi-agent neural network of the cloud; and each edge computing unit acquires state information of the respective cluster by using a self-adaptive data compression and acquisition method, and performs Markov game by using the received model and the optimal parameter to obtain a power scheduling result of each node in each cluster. According to the method, the virtual power plant scheduling decision accuracy under limited time and limited observation information is improved.

Description

Technical Field

[0001] The present application relates to the field of virtual power plant optimization scheduling, and in particular to a lightweight cloud-edge collaborative optimization scheduling method and device for virtual power plants. Background Art

[0002] Virtual power plants effectively integrate various centralized and distributed clean energy generation systems and various forms of energy, enabling efficient energy utilization, effective coordination of adjustable resources, and reliable supply of diverse energy needs. With the wide-area, multi-point integration of integrated energy equipment into virtual power plants, coordinated optimization of virtual power plant systems and the efficient solution of optimization models are becoming important research trends in the future. However, as the scale of virtual power plants continues to expand, ensuring the real-time performance of system scheduling decisions is becoming increasingly difficult. The application of cloud-edge collaborative technology will provide new insights into the complex real-time scheduling issues faced by virtual power plants.

[0003] Within the cloud-edge collaborative framework, the optimal scheduling of flexible resources in virtual power plants (VPPs) can achieve cost-effective decisions. However, in systems with large-scale grid-connected flexible resources, these resources are diverse and widely distributed. Defining the control area of ​​edge nodes and reducing the cross-domain transmission of adjustable flexible resource output energy during source-load fluctuations are key to improving the accuracy of VPP scheduling decisions under limited time and observation information, and are also urgent issues that need to be addressed in current VPP optimal scheduling research. Summary of the Invention

[0004] The purpose of this application is to provide a lightweight cloud-edge collaborative optimization scheduling method and device for a virtual power plant, which can improve the accuracy of virtual power plant scheduling decisions under limited time and limited observation information.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a lightweight cloud-edge collaborative optimization scheduling method for a virtual power plant, comprising:

[0007] In the cloud, the virtual power plant clusters are divided according to the energy coupling strength of the virtual power plant using the community discovery algorithm to obtain the cluster division results;

[0008] With the goal of minimizing cluster operation losses and wind and solar curtailment penalties, a virtual power plant optimization scheduling model is established;

[0009] The virtual power plant optimal dispatch model is modeled as a Markov game model;

[0010] According to the Markov game model, the multi-agent neural network parameters are trained in the Markov game process to obtain the optimal parameters of the multi-agent neural network;

[0011] According to the cluster division results, an edge computing unit is configured for each cluster, and each edge computing unit receives the optimal parameters of the Markov game model and the multi-agent neural network in the cloud;

[0012] When the edge computing units receive the power adjustment and scheduling instructions for their respective clusters from the cloud, each edge computing unit collects the status information of its respective cluster and adaptively compresses the status information to obtain lightweight status information; wherein, the edge computing units adaptively adjust the collection frequency when collecting the status information of their respective clusters;

[0013] In each edge computing unit, based on the lightweight status information of each cluster and the power adjustment scheduling instructions of each cluster, a Markov game is performed using the Markov game model and the optimal parameters of the multi-agent neural network to obtain the power scheduling results of each node in each cluster.

[0014] On the second aspect, the present application provides a lightweight cloud-edge collaborative optimization scheduling device for a virtual power plant, including: a cloud and multiple edge computing units.

[0015] The cloud is used to divide virtual power plant clusters based on the energy coupling strength of virtual power plants using a community discovery algorithm to obtain cluster division results. A virtual power plant optimization scheduling model is established with the goal of minimizing cluster operating losses and wind and solar power curtailment penalties. The virtual power plant optimization scheduling model is modeled as a Markov game model. Based on the Markov game model, the multi-agent neural network parameters in the Markov game process are trained to obtain the optimal parameters of the multi-agent neural network.

[0016] Configure an edge computing unit for each cluster in the cluster division result;

[0017] Each edge computing unit is used to receive the optimal parameters of the Markov game model and multi-agent neural network from the cloud;

[0018] Each edge computing unit is used to collect the status information of its respective cluster upon receiving the power adjustment and scheduling instructions of its respective cluster issued by the cloud, and to perform adaptive data compression on the status information to obtain lightweight status information; based on the lightweight status information of each cluster and the power adjustment and scheduling instructions of each cluster, a Markov game is performed using the Markov game model and the optimal parameters of the multi-agent neural network to obtain the power scheduling results of each node in each cluster; wherein, the edge computing unit adaptively adjusts the collection frequency when collecting the status information of its respective cluster.

[0019] According to the specific embodiments provided in this application, this application has the following technical effects:

[0020] The present application provides a lightweight cloud-edge collaborative optimization scheduling method and device for a virtual power plant. The cloud uses a community discovery algorithm to divide the virtual power plant cluster according to the energy coupling strength of the virtual power plant, and configures an edge computing unit for each cluster. Each edge computing unit is only responsible for scheduling the node power within the cluster. It can be seen that the edge computing unit, as an edge node, defines the control area range of the edge node; after the cluster is divided, according to the power adjustment scheduling instructions of each cluster issued by the cloud, power adjustment is only performed within each cluster, and there is no need to perform it between clusters, avoiding cross-domain transmission of adjustable flexible resource output energy during source and load fluctuations, thereby improving the accuracy of virtual power plant scheduling decisions under limited time and limited observation information. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0022] Figure 1 A detailed flowchart of a lightweight cloud-edge collaborative optimization scheduling method for a virtual power plant provided in one embodiment of the present application;

[0023] Figure 2 A simplified flowchart of a lightweight cloud-edge collaborative optimization scheduling method for a virtual power plant provided in one embodiment of the present application;

[0024] Figure 3 A schematic diagram of the simulation system topology provided in a specific embodiment of the present application;

[0025] Figure 4 A schematic diagram of the photovoltaic output prediction value of each node provided in a specific embodiment of this application;

[0026] Figure 5 A schematic diagram of the average reward convergence curve for optimal scheduling of a virtual power plant provided in a specific embodiment of this application;

[0027] Figure 6 This is a schematic diagram showing the comparison of daily operating losses of various clusters provided in a specific embodiment of this application;

[0028] Figure 7 A structural diagram of a lightweight cloud-edge collaborative optimization scheduling device for a virtual power plant provided in one embodiment of the present application. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0030] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0031] In an exemplary embodiment, Figure 1 As shown, a lightweight cloud-edge collaborative optimization scheduling method for a virtual power plant is provided, including the following steps 101 to 107. Among them:

[0032] Step 101: In the cloud, a community discovery algorithm is used to divide the virtual power plant clusters according to the energy coupling strength of the virtual power plant to obtain a cluster division result;

[0033] Step 102: Establish a virtual power plant optimization scheduling model with the goal of minimizing cluster operation losses and wind and solar power curtailment penalties;

[0034] Step 103: Modeling the virtual power plant optimization scheduling model as a Markov game model;

[0035] Step 104: According to the Markov game model, train the multi-agent neural network parameters in the Markov game process to obtain the optimal parameters of the multi-agent neural network;

[0036] Step 105: Based on the cluster division results, configure an edge computing unit for each cluster, and enable each edge computing unit to receive the optimal parameters of the Markov game model and the multi-agent neural network in the cloud;

[0037] Step 106: When the edge computing units receive the power adjustment scheduling instructions for their respective clusters from the cloud, each edge computing unit collects status information of its respective cluster and adaptively compresses the status information to obtain lightweight status information. The edge computing units adaptively adjust the frequency of collection when collecting the status information of their respective clusters.

[0038] Step 107: In each edge computing unit, based on the lightweight status information of each cluster and the power adjustment scheduling instructions of each cluster, a Markov game is performed using the Markov game model and the optimal parameters of the multi-agent neural network to obtain the power scheduling results of each node in each cluster.

[0039] Implement the above steps 101 to 107, adopt data adaptive fitting compression and frequency conversion acquisition algorithm to collect data from edge nodes of virtual power plants, propose a virtual power plant cluster division method based on community discovery algorithm, define cluster as the basic control unit of cloud-edge collaborative system, take the minimization of virtual power plant operating loss and renewable energy power abandonment as optimization goals, propose a virtual power plant real-time optimization method, transform the original problem into a partitioned Markov game problem, and adopt multi-agent deep reinforcement learning algorithm to solve it, which can improve the high efficiency and real-time performance of the virtual power plant control scheme under large-scale integrated energy equipment access.

[0040] The scheduling method of the present application can be understood as a "two-layer scheduling" architecture (or cloud-edge collaborative architecture). The upper layer is the cloud side, which is responsible for scheduling the overall power of each cluster, corresponding to steps 101 to 104; the lower layer is the edge side (edge ​​computing unit), which is responsible for scheduling the node power within the cluster, corresponding to steps 105 to 107.

[0041] A virtual power plant contains a large number of conversion and storage devices between different energy sources. Under the cloud-edge collaborative architecture, the basic control unit in the virtual power plant that receives cloud scheduling instructions is no longer a specific device, but an edge computing unit (also known as a cluster edge node intelligent computing device) in the area to which the equipment belongs after clustering. The cloud is responsible for model training and issuing cluster unit scheduling instructions. In addition, the cloud also has the function of collecting edge computing unit data. The edge computing unit receives the model trained by the cloud and is responsible for collecting and processing model parameters within the cluster. It performs model calculations and issues scheduling instructions based on the device status. The high degree of autonomy of the edge computing unit can improve the real-time control of large-scale equipment. In the process of collecting data, the edge computing unit has the problem of complex operation data collection redundancy, and it is necessary to improve the data compression and collection methods.

[0042] In another exemplary embodiment of this application, system clusters are divided according to the energy coupling strength index of virtual power plants, which can improve cluster autonomy and mutual assistance and provide a model foundation for cloud-edge collaborative optimization scheduling. The energy coupling strength of the virtual power plant structure and function is used as the cluster division index, and a community discovery algorithm is used to find the maximum energy coupling strength to divide the virtual power plant clusters.

[0043] Modularity is a key metric in complex network analysis and calculations, used to describe the closeness of connections within the resulting complex network communities. Therefore, modularity is used to describe the structural energy coupling strength of a virtual power plant. The functional energy coupling strength is described using supply-demand balance. Supply-demand balance reflects the reliability of electrical and thermal energy supply under source and load uncertainty. The aforementioned method for determining the energy coupling strength of a virtual power plant can be replaced by the following steps 201 to 203:

[0044] Step 201: Using the formula Calculate the structural energy coupling intensity of the virtual power plant.

[0045] Where, ρ MES is the energy coupling strength of the virtual power plant in the structure; w is the sum of the edge weights of the virtual power plant network; δ i,j A is a 0-1 indicator function, indicating whether two nodes belong to the same cluster; ij is the edge weight between node i and node j; k i is the sum of the edge weights connecting node i; k j is the sum of the edge weights connecting node j;

[0046] Step 202: Using the formula Calculate the functional energy coupling intensity of the virtual power plant.

[0047] Where, ψ MES is the functional energy coupling intensity of the virtual power plant; C is the number of cluster divisions; T is the total time period; ψ MES,c,t is the supply and demand balance degree of cluster c in period t; The power and heat source outputs of cluster c during period t; is the electrical load and thermal load in cluster c during period t; is the power line loss and heat loss on the branch line of cluster c during period t; M S,c is the total number of power and heat source nodes in cluster c; M L,c is the total number of electrical load and thermal load nodes in cluster c; is the total number of branches in cluster c;

[0048] Step 203: Based on the energy coupling strength of the virtual power plant in structure and the energy coupling strength of the virtual power plant in function, use the formula Q MES =a1ρ MES +a2ψ MES , and obtain the energy coupling intensity of the virtual power plant.

[0049] Where Q MES is the energy coupling intensity of the virtual power plant; a1 and a2 are both weighted coupling factors.

[0050] That is, according to the structural division indicators and functional division indicators in the virtual power plant cluster division model, the energy coupling intensity can be obtained through weighted coupling.

[0051] In another exemplary embodiment, in the process of calculating the structural energy coupling strength, the value of the edge weight of the virtual power plant network has a greater impact on the calculation result, and the quantization process of the edge weight also reflects the coupling strength relationship between the node energies in the system. For the edge weight calculation of the distribution network, the branch flow distribution factor is used for definition. The branch flow distribution factor is a factor that calculates the corresponding relationship between the node and branch power changes. Through this factor, p li and p lj Therefore, when the node is a distribution network node, the edge weight A between node i and node j is ij The calculation formula is:

[0052]

[0053] Where, is the edge weight between node i and node j in the distribution network; P ij is the electrical distance between node i and node j in the distribution network; p li is the impact of the power change of node i on the power change of branch l in the distribution network; p lj is the impact of the power change of node j on the power change of branch l in the distribution network; L E is the number of branches in the distribution network.

[0054] For the edge weights in the heat distribution network, the loss distribution factor along the way is used to define them. Therefore, when the node is a heat distribution network node, the edge weight A between node i and node j is ij The calculation formula is:

[0055]

[0056] Where, is the edge weight between node i and node j in the heat distribution network; Z ij is the thermal energy distance between node i and node j in the heat distribution network; Ω ij is the total connection distance between node i and node j in the heat distribution network; K is the roughness of the pipe network; G is the medium flow rate; ρ is the medium density; and D is the inner diameter of the pipe.

[0057] In another exemplary embodiment of the present application, the cluster division process in step 101 may be replaced by the following steps 301 to 305:

[0058] Step 301: Select each node in the virtual power plant as a target node in turn.

[0059] Step 302: After each target node is selected, the neighboring nodes and the target node are combined into a temporary community, and the energy coupling strength of the virtual power plant in the current state is calculated to determine the change value ΔQ of the energy coupling strength of the virtual power plant in the current state relative to the energy coupling strength in the previous state. MES .

[0060] Step 303: Obtain the maximum value of the energy coupling intensity change values ​​corresponding to all target nodes, and based on the fast bisection algorithm, determine the nodes in the temporary community corresponding to the maximum value as the same cluster to obtain a new cluster division result.

[0061] Step 304: If the number of cluster divisions in the new cluster division result is not equal to the preset number of cluster divisions, each node in the temporary community is selected as a target node in turn, and the process returns to step 302.

[0062] Step 305: If the number of cluster divisions in the new cluster division result is equal to the preset number of cluster divisions, the latest cluster division result is output as the final cluster division result.

[0063] The final clustering result includes the number of clusters and the node allocation of each cluster. After obtaining the final clustering result, the energy coupling strength at that time can also be obtained.

[0064] In another exemplary embodiment of the present application, the goal of optimizing the scheduling of a virtual power plant is to reduce system operating losses and reduce wind and solar power curtailment during the daily scheduling process. With the goal of minimizing cluster operating losses and wind and solar power curtailment penalties, the virtual power plant optimization scheduling objective function is established as:

[0065]

[0066] Where g c is the objective function for cluster c scheduling; x c C is the energy conversion and storage device in cluster c; k,i,t is the operating loss of type k equipment on node i in period t; N c is the total number of cluster nodes; is the power output by the energy conversion device at node i in time period t; is the power output by the storage device on node i in period t; C CHP,i,t is the combined heat and power electric-thermal coupling operation loss at node i in period t; a1 and b1 are the combined heat and power electric-thermal coupling operation loss coefficients; P CHPe,i,t 、P CHPh,i,t are the electricity output and heat output of the cogeneration at node i in period t respectively; C EB,i,t is the operating loss of the electric boiler at node i in period t; P EB,i,tis the electric power of the electric boiler at node i in period t; C EST,i,t is the operating loss of the power storage system at node i in period t; P EST,i,t is the power output of the storage system at node i in period t; C HST,i,t is the operating loss of the heat storage system at node i in period t; P HST,i,t is the thermal output of the heat storage system at node i in period t; C cur,i,t is the photovoltaic operation loss at node i in period t; P cpv,i,t is the photovoltaic curtailment power at node i in period t; C EX,i,t P is the operation loss of the cluster external connection line on node i in the tth period; EXe,i,t is the cluster external power tie line interaction power on node i in period t; P EXh,i,t is the interaction power of the cluster's external thermal interconnection line at node i in period t; c1, c2, c3, and c4 are the operating loss coefficients of the cogeneration, electric boiler, power storage system, and heat storage system, respectively; c5 is the penalty coefficient for abandoned solar power; c6 and c7 are the interaction loss coefficients of electric power and thermal power between clusters, respectively.

[0067] The optimal scheduling constraints of virtual power plants include equality constraints represented by power balance constraints and inequality constraints represented by cluster power interaction constraints and equipment operation constraints.

[0068] The equality constraints are:

[0069]

[0070] Where, P pv,i,t is the photovoltaic output at node i in period t; P loade,i,t is the electrical load on node i in period t; P EBh,i,t is the thermal output of the electric boiler at node i in period t; P loadh,i,t is the heat load on node i in period t.

[0071] The inequality constraints are:

[0072] Where, P EXe,max 、P EXe,min are the upper and lower limits of the cluster external power line interaction power respectively; P EXh,max 、P EXh,min The upper and lower limits of the cluster's external thermal interconnection line interaction power; P CHPe,max is the upper limit of combined heat and power output; P CHPh,max is the upper limit of thermal output of cogeneration; P EB,max is the upper limit of electric boiler power; P EST,max 、P EST,min are the upper and lower limits of the power storage system output respectively; PHST,max 、P HST,min They are the upper and lower limits of the heat storage system output respectively.

[0073] In another exemplary embodiment of this application, after virtual power plants are divided into clusters, the energy support strength within the clusters is high, and the coupling between clusters is weak. Therefore, the virtual power plant optimization scheduling problem can be described using game theory. At the same time, due to the sequential nature of the intraday scheduling problem, it can be modeled as a Markov game problem.

[0074] Among the game players, that is, within each cluster, the intelligent computing device at the edge of the cluster is responsible for collecting state information and publishing action information within the cluster. The Markov game process is as follows:

[0075] State space: S t is the virtual power plant state information, including photovoltaic output, electrical load, and thermal load. The state space of the cluster edge node agent c is:

[0076] s c,t ∈S c,t ={P c,pv,t ,P c,loade,t ,P c,loadh,t};

[0077] Where S c,t is the state space of cluster c in time period t; P c,pv,t is the photovoltaic output of cluster c in period t; P c,loade,t is the electric load of cluster c in period t; P c,loadh,t is the heat load of cluster c in period t.

[0078] Action space: a t It is the action information of virtual power plant, including the electric output and thermal output of cogeneration P CHPe,t 、P CHPh,t , the electric power P of the electric boiler EB,t , the power output P of the power storage system EST,t , the thermal output P of the heat storage system HST,t , cluster external power line interactive power P EXe,t ; Cluster external heat interconnection line interactive power P EXh,t , the action space of cluster edge node agent c is:

[0079]

[0080] Where a c,t is the action space of cluster c in time period t; P c,CHPe,t is the combined heat and power output of cluster c in period t; P c,CHPh,t is the thermal output of the combined heat and power of cluster c in period t; P c,EB,tis the electric power of the electric boiler of cluster c in period t; P c,EST,t is the power output of the storage system of cluster c in period t; P c,HST,t is the thermal output of the heat storage system of cluster c in period t; P c,EXe,t is the cluster external power tie line interaction power of cluster c in period t; P c,EXh,t is the cluster external thermal interconnection line interaction power of cluster c in period t.

[0081] Reward function: The optimal goal obtained by all C clusters is the reward value of the Markov game process. The reward value of the cluster edge node agent c is:

[0082]

[0083] Where r c,t is the reward value of cluster c in the tth period; C k,i,t N is the operating loss of type k equipment on node i in cluster c during period t; c is the total number of cluster nodes; V t is the cluster tie line limit penalty item in period t; θ is the penalty coefficient; σ is a 0-1 indicator function, which takes 1 when the tie line power exceeds the limit and takes 0 when the tie line power exceeds the limit but does not exceed the line.

[0084] In the lightweight cloud-edge collaborative framework of the virtual power plant, the cloud first initializes the network parameters, including the value network parameters and the policy network parameters. Next, hourly decision-making training is carried out in the target network. 96 decision moments a day constitute a training round. At each decision moment in each round, the cluster edge node agent uses the state observation value s c,t Get the action value a c,t , and use the objective function to calculate the reward value r c,t , proceed to the next state value s c,t+1 Iteration, forming the data in the experience reuse pool {s c,t ,a c,t ,r c,t ,s c,t+1 The value network loss function is updated using gradient descent in each iteration and is expressed as:

[0085]

[0086] Where y c is the target value; F c is the minimum sample batch for training; is the value network parameter, which is calculated using the loss function gradient descent equation and the soft update coefficient υ. θ c ←θc +η Q ▽L(θ c ); is the policy network parameter, and the calculation method uses the maximum policy gradient and υ to calculate. is the value network function, is the action of the k-th cluster at the t-th decision moment, is the action of the kth cluster at the Nth decision moment, μ is the discount factor, is the global discounted return, a c,t+1 is the action at the t+1th decision moment, is the mapping strategy from state to action, α is the action network learning factor, The target value network.

[0087] Data lightweighting includes adaptive data compression and adaptive adjustment of the acquisition frequency. Adaptive data compression and acquisition methods improve data acquisition accuracy while reducing the computational burden placed on the intelligent computing devices at the edge nodes of the cluster due to increased data volume. When fitting nonlinear data such as device power, a nonlinear adaptive fitting function is used to avoid errors introduced by linearized data fitting.

[0088] In another exemplary embodiment of the present application, adaptive data compression is performed on the state information, which specifically includes the following steps 401 to 406.

[0089] Step 401: Based on the actual value of the state information, the revolving door algorithm is used to determine the adaptive fitting function M = km α +β; where α is the coefficient of the first adaptive fitting function, k is the coefficient of the second adaptive fitting function, β is the coefficient of the third adaptive fitting function, m is the actual value of the state information, and M is the fitting value of the state information.

[0090] Step 402: Determine the change trend of the status information; the change trend includes stable, rising and falling.

[0091] Step 403: If the change trend is stable, the first adaptive fitting function coefficient α is set to 0.5.

[0092] Step 404: If the change trend is upward, the first adaptive fitting function coefficient α is set to 2.

[0093] Step 405: If the change trend is downward, the first adaptive fitting function coefficient α is set to 3.

[0094] Step 406: Obtain a fitting value of the state information using an adaptive fitting function according to the actual value of the state information and the value of the first adaptive fitting function coefficient α.

[0095] In another exemplary embodiment of the present application, during the data collection interval, it is necessary to dynamically adjust the data collection frequency according to the data validity. The edge computing unit adaptively adjusts the collection frequency when collecting the status information of each cluster, specifically including the following steps 501 to 504.

[0096] Step 501: Determine the predicted value of the state information using the least squares method based on the fitted value of the state information.

[0097] Step 502: When the predicted value of the state information is smaller than the actual value of the state information at the same time, use the formula T c ′=ΔT c +T c , determine the acquisition frequency after adaptive adjustment; T′ is the acquisition frequency after adaptive adjustment, ΔT c is the allowed range of time interval, T c is the current collection frequency.

[0098] Step 503: When the predicted value of the state information is greater than the actual value of the state information at the same time, use the formula T c ′=T c / 1.2, determines the acquisition frequency after adaptive adjustment.

[0099] Step 504: The edge computing unit collects status information of each cluster according to the adaptively adjusted collection frequency.

[0100] Adaptive data compression and collection can reduce the redundancy of complex operation data collection and provide a data basis for cloud-edge collaborative optimization scheduling of virtual power plants.

[0101] Figure 2 This is a brief flow chart of the method of this application. The brief process is: start → propose a lightweight cloud-edge collaborative framework for virtual power plants → lightweight data processing → virtual power plant cluster division → propose a virtual power plant optimization scheduling model and a multi-agent reinforcement learning solution algorithm → output virtual power plant scheduling results → end. The edge nodes of this application adopt an adaptive data compression and collection method to improve the data collection speed, and use a community discovery algorithm to cluster the virtual power plant. On the basis of the cluster architecture, a virtual power plant cloud-edge coordinated optimization scheduling method based on multi-agent reinforcement learning is proposed, which can effectively improve the model solution efficiency, reduce system losses, and improve the power balance capability and local consumption level under renewable energy output fluctuations, providing guidance for the development of virtual power plant control theory.

[0102] The effectiveness of the method of the present application is illustrated below with a specific example.

[0103] Using the actual one-year operation data of a virtual power plant in a certain area of ​​Northeast China, a virtual power plant simulation system based on IEEE 33-node power distribution system and 32-node heat distribution system was built. The simulation system topology is as follows: Figure 3 The simulation system is used as the application environment of the multi-agent deep reinforcement learning algorithm. The algorithm is implemented based on Python. The algorithm parameters are shown in Table 1. The photovoltaic output prediction value curves of PV-5, PV-11, PV-14, PV-18, PV-22, PV-28, and PV-33 connected to the simulation system are shown in Figure 4 shown.

[0104] Table 1 Multi-agent deep reinforcement learning algorithm parameters

[0105] parameter Numerical Number of hidden layer neurons 100 Discount Factor 0.001 Learning rate 0.005 Soft update coefficient 0.001

[0106] Table 2 shows the results of virtual power plant cluster division. Based on the virtual power plant cluster division, a single cluster is used as an intelligent agent, and a virtual power plant edge node intelligent computing device is set up to perform cloud-edge collaborative deep reinforcement learning algorithm training.

[0107] Table 2 Virtual power plant cluster division results

[0108] Cluster Power distribution system nodes Heat distribution system node 1 {1,2,3,4,5,19,20,21,22} {19,20,21,22,23} 2 {6,23,24,25,26,27,28} {4,6,7,8,9,10,31} 3 {7,8,9,10,11,12,13,14} {1,2,3,5,11,12,13,32} 4 {15,16,17,18} {14,15,16,17,18} 5 {29,30,31,32,33} {24,25,26,27,28,29,30}

[0109] Figure 5 The average reward convergence curve for optimal scheduling of virtual power plants under multi-cluster interconnection. Figure 5 It can be seen that in the early stages of training, the agent's decision-making fluctuates significantly, primarily due to its exploration of scheduling decisions. However, as the training step length increases, the scheduling decision information in the experience reuse pool becomes more reasonable, and the decision results better align with the actual supply and demand balance requirements of the scheduling model. Starting from the 120th round, the average reward gradually converges to the optimal value, indicating that the interactive power of each cluster in the multi-cluster interconnection meets the power flow safety range requirements, and that the clusters achieve autonomous optimal operation, achieving optimal scheduling for the entire virtual power plant.

[0110] The comparison results of daily operation losses of each cluster under the cloud-edge collaborative framework are as follows: Figure 6 As shown. Figure 6It can be seen that the changes in operating losses of each cluster are most affected by changes in load. This is mainly because the purpose of operating a virtual power plant is to ensure the reliable supply of electric and thermal loads while maximizing the absorption of renewable energy and reducing system operating losses. Therefore, changes in load directly determine changes in operating losses. When the photovoltaic output is large, the operating losses of clusters with a large proportion of photovoltaic grid-connected capacity (clusters 3 and 4) will also fluctuate significantly. This is because the fluctuation of photovoltaic output will affect the power balance of the system. To achieve local absorption of photovoltaic power, the edge node intelligent agent will issue adjustment instructions to the local controllable resources, thereby increasing the operating losses of the system. However, the operating losses of other interconnected clusters have not increased significantly, which directly verifies the effectiveness of the cloud-edge collaboration framework in improving the autonomy within the cluster.

[0111] Based on the same inventive concept, the embodiment of the present application also provides a virtual power plant lightweight cloud-edge collaborative optimization scheduling device for implementing the virtual power plant lightweight cloud-edge collaborative optimization scheduling method involved above. The implementation solution provided by the device is similar to the implementation solution described in the above method. Therefore, the specific limitations in the embodiments of one or more virtual power plant lightweight cloud-edge collaborative optimization scheduling devices provided below can be found in the above limitations of the virtual power plant lightweight cloud-edge collaborative optimization scheduling method, and will not be repeated here.

[0112] In an exemplary embodiment, Figure 7 As shown, a lightweight cloud-edge collaborative optimization scheduling device for a virtual power plant is provided, including: a cloud and multiple edge computing units.

[0113] The cloud is used to divide the virtual power plant clusters according to the energy coupling intensity of the virtual power plant using the community discovery algorithm to obtain the cluster division results; a virtual power plant optimization scheduling model is established with the goal of minimizing the cluster's operating losses and wind and solar power abandonment penalties; the virtual power plant optimization scheduling model is modeled as a Markov game model; based on the Markov game model, the multi-agent neural network parameters in the Markov game process are trained to obtain the optimal parameters of the multi-agent neural network.

[0114] An edge computing unit is configured for each cluster in the cluster division result; each edge computing unit is used to receive the optimal parameters of the Markov game model and the multi-agent neural network from the cloud; each edge computing unit is used to collect the status information of the respective cluster when receiving the power adjustment scheduling instruction of the respective cluster issued by the cloud, and adaptively compress the said status information to obtain lightweight status information; according to the lightweight status information of the respective cluster and the power adjustment scheduling instruction of the respective cluster, the Markov game model and the optimal parameters of the multi-agent neural network are used to perform a Markov game to obtain the power scheduling result of each node in each cluster; wherein, the edge computing unit adaptively adjusts the collection frequency when collecting the status information of the respective cluster.

[0115] As an optional implementation, the cloud is used to collect status information of each cluster collected by each edge computing unit.

[0116] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0117] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A lightweight cloud-edge collaborative optimization scheduling method for virtual power plants, characterized in that: include: In the cloud, the virtual power plant clusters are divided according to the energy coupling strength of the virtual power plant using the community discovery algorithm to obtain the cluster division results; With the goal of minimizing cluster operation losses and wind and solar curtailment penalties, a virtual power plant optimization scheduling model is established; The virtual power plant optimal dispatch model is modeled as a Markov game model; According to the Markov game model, the multi-agent neural network parameters are trained in the Markov game process to obtain the optimal parameters of the multi-agent neural network; According to the cluster division results, an edge computing unit is configured for each cluster, and each edge computing unit receives the optimal parameters of the Markov game model and the multi-agent neural network in the cloud; When the edge computing units receive the power adjustment and scheduling instructions for their respective clusters from the cloud, each edge computing unit collects the status information of its respective cluster and adaptively compresses the status information to obtain lightweight status information; wherein, the edge computing units adaptively adjust the collection frequency when collecting the status information of their respective clusters; In each edge computing unit, based on the lightweight status information of each cluster and the power adjustment scheduling instructions of each cluster, a Markov game is performed using the Markov game model and the optimal parameters of the multi-agent neural network to obtain the power scheduling results of each node in each cluster.

2. The lightweight cloud-edge collaborative optimization scheduling method for virtual power plants according to claim 1 is characterized in that: The method for determining the energy coupling strength of the virtual power plant includes: Using the formula Calculate the energy coupling intensity of the virtual power plant in the structure; where, ρ MES is the energy coupling strength of the virtual power plant in the structure; w is the sum of the edge weights of the virtual power plant network; δ i,j A is a 0-1 indicator function, indicating whether two nodes belong to the same cluster; ij is the edge weight between node i and node j; k i is the sum of the edge weights connecting node i; k j is the sum of the edge weights connecting node j; Using the formula Calculate the functional energy coupling intensity of the virtual power plant; where, ψ MES is the functional energy coupling intensity of the virtual power plant; C is the number of cluster divisions; T is the total time period; ψ MES,c,t is the supply and demand balance degree of cluster c in period t; The power and heat source outputs of cluster c during period t; is the electrical load and thermal load in cluster c during period t; is the power line loss and heat loss on the branch line of cluster c during period t; M S,c is the total number of power and heat source nodes in cluster c; M L,c is the total number of electrical load and thermal load nodes in cluster c; is the total number of branches in cluster c; According to the energy coupling intensity of the virtual power plant in structure and the energy coupling intensity of the virtual power plant in function, the formula Q MES =a1ρ MES +a2ψ MES , obtain the energy coupling intensity of the virtual power plant; where Q MES is the energy coupling intensity of the virtual power plant; a1 and a2 are both weighted coupling factors.

3. The lightweight cloud-edge collaborative optimization scheduling method for virtual power plants according to claim 2 is characterized in that: When the node is a distribution network node, the edge weight A between node i and node j is ij The calculation formula is: Where, is the edge weight between node i and node j in the distribution network; P ij is the electrical distance between node i and node j in the distribution network; p li is the impact of the power change of node i on the power change of branch l in the distribution network; p lj is the impact of the power change of node j on the power change of branch l in the distribution network; L E is the number of branches in the distribution network; When the node is a heat distribution network node, the edge weight A between node i and node j is ij The calculation formula is: Where, is the edge weight between node i and node j in the heat distribution network; Z ij is the thermal energy distance between node i and node j in the heat distribution network; Ω ij is the total connection distance between node i and node j in the heat distribution network; K is the roughness of the pipe network; G is the medium flow rate; ρ is the medium density; and D is the inner diameter of the pipe.

4. The lightweight cloud-edge collaborative optimization scheduling method for virtual power plants according to claim 1 is characterized in that: According to the energy coupling intensity of the virtual power plant, the community discovery algorithm is used to divide the virtual power plant clusters and obtain the cluster division results, which include: Each node in the virtual power plant is selected as the target node in turn; After each target node is selected, the neighboring nodes and the target node are combined into a temporary community, and the energy coupling strength of the virtual power plant in the current state is calculated to determine the change value of the energy coupling strength of the virtual power plant in the current state relative to the energy coupling strength in the previous state; Obtain the maximum value of the energy coupling intensity change values ​​corresponding to all target nodes, and based on the fast bisection algorithm, determine the nodes in the temporary community corresponding to the maximum value as the same cluster to obtain a new cluster division result; If the number of cluster divisions in the new cluster division result is not equal to the preset number of cluster divisions, each node in the temporary community is selected as the target node in turn, and the process returns to step "After each target node is selected, the neighboring nodes and the target node are combined into a temporary community, and the energy coupling strength of the virtual power plant in the current state is calculated to determine the change in the energy coupling strength of the virtual power plant in the current state relative to the energy coupling strength in the previous state"; If the number of cluster divisions in the new cluster division result is equal to the preset number of cluster divisions, the latest cluster division result is output as the final cluster division result.

5. The lightweight cloud-edge collaborative optimization scheduling method for virtual power plants according to claim 1 is characterized in that: The virtual power plant optimization scheduling model includes: a virtual power plant optimization scheduling objective function and a virtual power plant optimization scheduling constraint condition; the virtual power plant optimization scheduling constraint condition includes an equality constraint and an inequality constraint; The virtual power plant optimization scheduling objective function is: Where g c is the objective function for cluster c scheduling; x c C is the energy conversion and storage device in cluster c; k,i,t is the operating loss of type k equipment on node i in period t; N c is the total number of cluster nodes; is the power output by the energy conversion device at node i in time period t; is the power output by the storage device on node i in period t; C CHP,i,t is the combined heat and power electric-thermal coupling operation loss at node i in period t; a1 and b1 are the combined heat and power electric-thermal coupling operation loss coefficients; P CHPe,i,t 、P CHPh,i,t are the electricity output and heat output of the cogeneration at node i in period t respectively; C EB,i,t is the operating loss of the electric boiler at node i in period t; P EB,i,t is the electric power of the electric boiler at node i in period t; C EST,i,t is the operating loss of the power storage system at node i in period t; P EST,i,t is the power output of the storage system at node i in period t; C HST,i,t is the operating loss of the heat storage system at node i in period t; P HST,i,t is the thermal output of the heat storage system at node i in period t; C cur,i,t is the photovoltaic operation loss at node i in period t; P cpv,i,t is the photovoltaic curtailment power at node i in period t; C EX,i,t P is the operation loss of the cluster external connection line on node i in the tth period; EXe,i,t is the cluster external power tie line interaction power on node i in period t; P EXh,i,t is the cluster external thermal interconnection line interaction power at node i in period t; c1, c2, c3, c4 are the operation loss coefficients of cogeneration, electric boiler, power storage system, and heat storage system respectively; c5 is the penalty coefficient for curtailment of solar power; c6 and c7 are the interaction loss coefficients of electric power and thermal power between clusters respectively; The equality constraints are: Where, P pv,i,t is the photovoltaic output at node i in period t; P loade,i,t is the electrical load on node i in period t; P EBh,i,t is the thermal output of the electric boiler at node i in period t; P loadh,i,t is the heat load on node i in period t; The inequality constraints are: Where, P EXe,max 、P EXe,min are the upper and lower limits of the cluster external power line interaction power respectively; P EXh,max 、P EXh,min The upper and lower limits of the cluster's external thermal interconnection line interaction power; P CHPe,max is the upper limit of combined heat and power output; P CHPh,max is the upper limit of thermal output of cogeneration; P EB,max is the upper limit of electric boiler power; P EST,max 、P EST,min are the upper and lower limits of the power storage system output respectively; P HST,max 、P HST,min They are the upper and lower limits of the heat storage system output respectively.

6. The lightweight cloud-edge collaborative optimization scheduling method for virtual power plants according to claim 1 is characterized in that: In the Markov game process, each cluster is regarded as an agent; The state space of the agent is: S c,t ={P c,pv,t ,P c,loade,t ,P c,loadh,t }; Where S c,t is the state space of cluster c in time period t; P c,pv,t is the photovoltaic output of cluster c in period t; P c,loade,t is the electric load of cluster c in period t; P c,loadh,t is the heat load of cluster c in period t; The action space of the agent is: Where a c,t is the action space of cluster c in time period t; P c,CHPe,t is the combined heat and power output of cluster c in period t; P c,CHPh,t is the thermal output of the combined heat and power of cluster c in period t; P c,EB,t is the electric power of the electric boiler of cluster c in period t; P c,EST,t is the power output of the storage system of cluster c in period t; P c,HST,t is the thermal output of the heat storage system of cluster c in period t; P c,EXe,t is the cluster external power tie line interaction power of cluster c in period t; P c,EXh,t is the cluster external heat tie line interaction power of cluster c in period t; The agent's reward value is: Where r c,t is the reward value of cluster c in the tth period; C k,i,t N is the operating loss of type k equipment on node i in cluster c during period t; c is the total number of cluster nodes; V t is the cluster tie line limit penalty item in period t; θ is the penalty coefficient; σ is a 0-1 indicator function, which takes 1 when the tie line power exceeds the limit and takes 0 when the tie line power exceeds the limit but does not exceed the line.

7. The lightweight cloud-edge collaborative optimization scheduling method for virtual power plants according to claim 1 is characterized in that: Adaptively compressing the state information includes: According to the actual value of the state information, the revolving door algorithm is used to determine the adaptive fitting function M=km α +β; where α is the coefficient of the first adaptive fitting function, k is the coefficient of the second adaptive fitting function, β is the coefficient of the third adaptive fitting function, m is the actual value of the state information, and M is the fitted value of the state information; Determining a change trend of the status information; the change trend includes stable, rising and falling; If the change trend is stable, the first adaptive fitting function coefficient α is set to 0.5; If the change trend is upward, the first adaptive fitting function coefficient α takes a value of 2; If the change trend is downward, the first adaptive fitting function coefficient α is set to 3; According to the actual value of the state information and the value of the first adaptive fitting function coefficient α, the adaptive fitting function is used to obtain the fitting value of the state information.

8. The lightweight cloud-edge collaborative optimization scheduling method for virtual power plants according to claim 7 is characterized in that: The edge computing units adaptively adjust the frequency of collection when collecting status information of their respective clusters, specifically including: According to the fitted value of the state information, the predicted value of the state information is determined by using the least square method; When the predicted value of the state information is smaller than the actual value of the state information at the same time, the formula T c ′=ΔT c +T c , determine the acquisition frequency after adaptive adjustment; T′ is the acquisition frequency after adaptive adjustment, ΔT c is the allowed range of time interval, T c is the current acquisition frequency; When the predicted value of the state information is greater than the actual value of the state information at the same time, the formula T c ′=T c / 1.2, determines the acquisition frequency after adaptive adjustment; The edge computing units collect status information of their respective clusters according to the adaptively adjusted collection frequency.

9. A lightweight cloud-edge collaborative optimization scheduling device for a virtual power plant, characterized in that: include: Cloud and multiple edge computing units; The cloud is used to divide virtual power plant clusters based on the energy coupling strength of virtual power plants using a community discovery algorithm to obtain cluster division results. A virtual power plant optimization scheduling model is established with the goal of minimizing cluster operating losses and wind and solar power curtailment penalties. The virtual power plant optimization scheduling model is modeled as a Markov game model. Based on the Markov game model, the multi-agent neural network parameters in the Markov game process are trained to obtain the optimal parameters of the multi-agent neural network. Configure an edge computing unit for each cluster in the cluster division result; Each edge computing unit is used to receive the optimal parameters of the Markov game model and multi-agent neural network from the cloud; Each edge computing unit is used to collect the status information of its respective cluster upon receiving the power adjustment and scheduling instructions of its respective cluster issued by the cloud, and to perform adaptive data compression on the status information to obtain lightweight status information; Based on the lightweight status information of each cluster and the power adjustment scheduling instructions of each cluster, a Markov game is performed using the Markov game model and the optimal parameters of the multi-agent neural network to obtain the power scheduling results of each node in each cluster; among them, the edge computing unit adaptively adjusts the collection frequency when collecting the status information of each cluster.

10. The lightweight cloud-edge collaborative optimization scheduling device for virtual power plants according to claim 9 is characterized in that: The cloud is used to collect the status information of each cluster collected by each edge computing unit.

Images