Hierarchical cooperative control method for multiple aggregated distributed resource clusters in source-load storage

Abstract

The invention discloses a hierarchical cooperative control method for multiple aggregation distributed resource clusters of source-load storage. The method comprises the following steps: dividing control levels of the multiple aggregation distributed resources of the source-load storage; based on the control level division result, aggregating the resources of the same control level into the same resource cluster, and constructing a plurality of aggregated distributed resource clusters of source-load storage; establishing a distributed resource cluster cooperative control model; and introducing a graph neural network model to solve the distributed resource cluster cooperative control model. According to the method, an improved fuzzy clustering algorithm is designed to classify resources, a mechanism of weighted information propagation and dynamic adjustment of propagation strength is introduced to improve an information transmission mechanism of the graph neural network, and the graph neural network is innovatively applied to the solution of a distributed resource cluster cooperative control model; and the modeling capability and the solving efficiency of a complex system are improved. And the calculation efficiency is kept in a high-dimensional and nonlinear scene, and the operation requirements of a large-scale distributed power grid are met.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and in particular to a hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads and storage. Background Art

[0002] Driven by the global energy transition and the "dual carbon" goals, distributed resources (including distributed photovoltaics, energy storage devices, and adjustable loads) have become an essential component of modern power systems. However, due to the randomness, uncertainty, and spatial dispersion of distributed resources, their high proportion of access poses numerous challenges to the voltage and frequency stability and operational economics of the power grid.

[0003] To address these challenges, traditional centralized control solutions have gradually exposed the following shortcomings: Response delay: Centralized control relies on a central control system for data collection and decision-making. Information transmission and decision-making can lead to response delays, making it difficult to cope with the rapid fluctuations of distributed resources; Single point of failure: Centralized control systems rely on a single central node for operation. If a node fails, the entire system will fail; Poor scalability: As the scale of distributed resource access increases, centralized systems face bottlenecks in computing power and communication bandwidth. To address these issues, distributed collaborative control technology has gradually become a research hotspot. It aims to achieve efficient coordination of multiple resources such as sources, loads, and storage through hierarchical resource aggregation and distributed optimization.

[0004] Existing technologies primarily achieve coordinated control of source, load, and storage resources through hierarchical distributed optimization and communication coordination. Distributed resources are divided into multiple subclusters based on their response speed, regulation capabilities, and geographic location. Distributed optimization algorithms (such as the alternating direction method of multipliers, ADMM) are then used within each subcluster for local control. Simultaneously, global resource allocation is achieved through a central coordination node to ensure system energy balance and operational efficiency. However, practical applications have exposed the following issues:

[0005] 1. Current distributed resource control schemes often rely on centralized control methods, which have poor adaptability to real-time dynamic environments and cannot quickly respond to changes in distributed resources (such as load fluctuations and distributed energy access). Control strategies also struggle to effectively handle the heterogeneity of different resources (such as response time, power range, and constraints), resulting in inefficient resource utilization.

[0006] 2. The lack of effective coordination mechanisms between resource clusters of different control levels (fast response, medium response, and slow response) leads to response lags and coordination failures during resource scheduling. The complex interactions between resources make traditional optimization methods unable to capture the dependencies and collaborative potential between clusters.

[0007] 3. Resource clusters are subject to multiple dynamic constraints during operation (such as voltage limits, power balance, and economic constraints). Traditional methods struggle to efficiently integrate these constraints into control models, resulting in control results that fail to meet actual requirements. The lack of adaptive processing capabilities for real-time constraint changes can lead to system instability or inefficiency.

[0008] 4. Distributed power grid systems have complex topologies and diverse interactions between resources (such as geographic proximity and power transmission relationships). Traditional methods have limited ability to model these complex relationships, resulting in deviations between optimization results and actual operations. When the number of resources is large, traditional optimization methods are computationally complex and difficult to meet real-time requirements. Summary of the Invention

[0009] To solve the above technical problems, the present invention provides a hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads and storage, comprising the following steps:

[0010] S1, divides the control levels of various aggregated distributed resources of source, load and storage;

[0011] S2, based on the control level classification result, aggregate resources of the same control level into the same resource cluster, and build a distributed resource cluster with multiple aggregations of source, load and storage;

[0012] S3, establish a distributed resource cluster collaborative control model;

[0013] S4, introduce a graph neural network model to solve the distributed resource cluster collaborative control model.

[0014] Wherein, the step S1 specifically includes the following steps:

[0015] S101, performing data collection and preprocessing on the multiple aggregated distributed resources of the source, load and storage;

[0016] S102, performing characteristic analysis on the multiple aggregated distributed resources of source, load and storage to extract important characteristics that affect resource classification; the characteristic analysis includes regulation capability analysis and economic analysis, and integrating the important characteristics into a unified characteristic vector: X i =[ΔP max ,T resp ,R 2 ,e i ,C unit ,R unit ,E ratio ]

[0017] Where ΔP max is the resource adjustable power range, T resp To measure the time from when the resource receives the control instruction to when it reaches the target state, R2 The coefficient of determination for determining whether the regulatory capacity changes linearly with time, e i To determine whether the regulatory ability changes linearly over time, C unit is the unit adjustment cost, R unit E is the benefit of computing resource per unit power regulation. ratio R unit with C unit The ratio of

[0018] S103, using an improved fuzzy clustering algorithm to classify the multiple aggregated distributed resources of source, load and storage;

[0019] S104, classifying the multiple aggregated distributed resources of source, load and storage into three levels: level one fast response, level two medium response and level three slow response.

[0020] The improved fuzzy clustering algorithm is based on the fuzzy clustering algorithm and introduces weighted characteristic measurement and characteristic constraint conditions for improvement, specifically including:

[0021] According to the different characteristics of the various aggregated distributed resources of source, load and storage, the importance of resource classification is different, and a weight w is assigned to each characteristic. j , the improved distance calculation formula is:

[0022]

[0023] Where D ik is the weighted distance from resource i to cluster center k, w j is the weight of feature j, x ij is the value of resource i on feature j, μ kj is the value of cluster center k on feature j;

[0024] Use the entropy method to assign weights to each feature: Among them H j is the information entropy of feature j;

[0025] Add constraints to the optimization objective:

[0026]

[0027] Where, f con is the characteristic constraint term, λ is the constraint weight, is the square of the distance from the resource to the cluster center on feature j.

[0028] Preferably, the distributed resource cluster collaborative control model includes an objective function and constraints, specifically including:

[0029] With the safe operation of the distribution network as the core goal, the following three goals are optimized:

[0030] 1) Minimize node voltage deviation:

[0031]

[0032] Where J1 is the square sum of node voltage deviation, V i is the voltage of the i-th node, V ref is the reference voltage, N is the total number of nodes in the distribution network;

[0033] 2) Minimize the active power loss of the line:

[0034]

[0035] Where J2 is the total power loss of the line, R ij is the resistance of line ij, I ij is the current of line ij, It is the collection of all lines in the distribution network;

[0036] 3) Optimal load balancing:

[0037]

[0038] Where, J3 is the sum of the absolute values of the power deviations of each phase, P φ is the total active power of phase φ, P total is the three-phase total active power, P total =∑ φ∈ { A,B,C}P φ ;

[0039] According to the three objectives, the objective function is established as:

[0040] minJ=α1J1+α2J2+α3J3

[0041] Where α1, α2, and α3 are target weight coefficients;

[0042] The constraints include:

[0043] 1) Power Constraints:

[0044]

[0045] Where, P i With Q i are the active power output and reactive power output of resource i, and are the upper and lower limits of active power output of resource i, and are the upper and lower limits of reactive power of resource i, and G is the set of all resources;

[0046] 2) Voltage constraint:

[0047]

[0048] Where V i is the voltage amplitude at node i, and are the upper and lower limits of the voltage of node i, and N is the set of all nodes;

[0049] 3) Line flow constraints:

[0050]

[0051] Where S ij is the apparent power of line ij, is the rated capacity of line ij, is the set of all transmission lines in the network;

[0052] 4) Cluster dynamic constraints:

[0053] Fast response cluster response time constraints:

[0054]

[0055] Where, T resp,i is the response time of fast response resource i, T req,1 is the maximum response time of the fast response cluster, G i A collection of resources for rapid response clusters;

[0056] Medium-speed response cluster dynamic power allocation constraints:

[0057]

[0058] Where, P req,2 is the target active power of the medium-speed response cluster, ΔP i is the adjustable power range of resource i G2 is a resource collection of a medium-speed response cluster;

[0059] Slow cluster planning adjustment constraints:

[0060]

[0061] Where, P plan With Q plan are the planned active and reactive powers of the slow-response cluster, respectively.

[0062] Preferably, establishing the graph neural network model includes the following steps:

[0063] S401, constructing a graph structure according to the distributed resource cluster collaborative control model: Graph G = (V, E), where V represents a set of nodes in the graph, and E represents a set of edges in the graph;

[0064] Each node v k ∈V represents a resource cluster G k , each of the resource clusters G k The state of is represented by the feature vector:

[0065]

[0066] Where, For resource cluster G k The initial characteristic vector, P k With Q k They are resource cluster G k Active power output and reactive power output, T resp,k For resource cluster G k The response time, C unit,k For resource cluster G k Unit adjustment cost, R unit,k is the computing resource cluster G k The benefit of each unit of power regulation, E ratio,k R unit,k with C unit,k The ratio of

[0067] Each edge e ij Represents the resource cluster G k Relationship or interaction between:

[0068] e ij =[ΔP ij ,ΔQ ij ,T flow,ij ]

[0069] Where ΔP ij and ΔQ ij Represents node v i and v j The power difference or reactive power flow, T flow,ij represents the response time of power flow;

[0070] Use the adjacency matrix A to represent the structural relationship of the graph, A ij Represents node v i and v j Is there a connection between A ij =1 means there is a connection, A ij =0 means no connection;

[0071] S402: Based on the graph structure, a new information dissemination mechanism is introduced; the new information dissemination mechanism is based on the traditional information dissemination mechanism and introduces a mechanism of weighted information dissemination and dynamic adjustment of dissemination intensity;

[0072] S403, a training data set, taking the initial characteristic vector of each node as input, and outputting the final adjustment strategy and scheduling result of each resource cluster through collaborative optimization of the nodes;

[0073] S404, defining a loss function to optimize the objective function and ensure that the model effectively learns the coordination between the resource clusters;

[0074] S405, using backpropagation and gradient descent to optimize weights and node feature vectors, and using the Adam optimizer to accelerate the training process;

[0075] S406, completing the training of the graph neural network model.

[0076] The weighted information propagation is specifically as follows:

[0077] At each information dissemination, the connection weight λ between the resource clusters ij Can be dynamically adjusted to:

[0078] λ ij＝ α·f(r i ,r j )+β·g(c i ,c j )+γ·h(d ij )

[0079] Where r i ,r j represent the response characteristics of clusters i and j, respectively, c i ,c j represents the constraints between clusters i and j, d ij represents the topological adjacency relationship between clusters i and j, f, g, and h are mapping functions used to measure the similarity between different features, and α, β, and γ are hyperparameters that control the importance of each feature in the weight calculation.

[0080] The dynamic adjustment of the dissemination intensity is specifically as follows:

[0081] According to the control level of the resource cluster and the constraints between the resource clusters, the intensity and frequency of information dissemination of each resource cluster are dynamically adjusted. Assume that in the information dissemination of the kth layer, node v i Need to get from its adjacent node v jReceive information and re-update its own status. The node information update formula is:

[0082]

[0083] Where, is node v i In the feature representation of the kth layer, N i is node v i The set of adjacent nodes, w ij For edge e ij The weighted propagation coefficient of node v i Its adjacent node v j The information propagation intensity between them is represented by , and f is the ReLU function.

[0084] Preferably, the weighted propagation coefficient w ij Adjust according to the following rules:

[0085] 1) Responsiveness Weighting

[0086]

[0087] Where, T resp,i and T resp,j is node v i and v j response time, α is the adjustment coefficient; if v i and v j Belong to resource clusters with similar control levels, then w ij Larger, indicating greater intensity of information transmission;

[0088] 2) Constrained matching weighting

[0089] For constrained resource clusters, information dissemination between nodes will dynamically adjust weights:

[0090]

[0091] Where, P max,i and P max,j Node v i and v j The maximum power of , β is the adjustment coefficient, which is used to control the influence of power difference on weight.

[0092] Preferably, the loss function is specifically:

[0093] Loss of voltage stability:

[0094]

[0095] Power balance loss:

[0096]

[0097] To ensure that the resource cluster operates within the control capability, the constraint satisfaction loss is defined:

[0098]

[0099] The total loss function is:

[0100] L total＝ L V +L P +L C .

[0101] Preferably, the solution step in step S4 is specifically:

[0102] S407, initializing the characteristic vector of each node Set the edge weighted propagation coefficient w of the graph ij ;

[0103] S408, information propagation and node update: performing multiple rounds of iterative updates of the graph neural network model, and calculating the characteristic vectors of the nodes through the weighted information propagation;

[0104] S409, constraint optimization: after each round of iteration, the constraint optimization is performed to ensure that the scheduling result of the resource cluster meets the safety and economic requirements of the power grid;

[0105] S410, minimizing the loss function: performing training according to the objective function and the loss function, and gradually optimizing the coordinated control among the resource clusters;

[0106] S411, output result: output the final scheduling strategy of each resource cluster.

[0107] Compared with the prior art, the beneficial technical effects of the present invention are:

[0108] This application proposes a new information propagation mechanism based on graph neural networks (GNNs). Through weighted information transfer between nodes and dynamic constraint embedding, it can rapidly propagate information and dynamically adjust control strategies within distributed resource clusters. Without relying on centralized computing, it achieves real-time dynamic optimization of resource clusters, improving resource utilization efficiency and responsiveness.

[0109] 2. A new information dissemination mechanism dynamically adjusts the intensity and frequency of dissemination, enabling collaboration between clusters of varying response levels and improving the overall efficiency of the control strategy. The introduction of multi-level control signals (fast response, medium response, and slow response) dynamically adjusts cluster priorities based on different task requirements, enabling efficient collaboration between resource clusters.

[0110] 3. A dynamic constraint information dissemination mechanism is introduced to integrate the dynamic constraints of each resource cluster (such as power caps and response times) into the node update process. Dynamic adjustments are made through the constraint matrix to ensure that the control results meet actual operational requirements. Constraints are gradually optimized over multiple iterations, ensuring that each resource cluster operates within its capabilities, ensuring the security and economic efficiency of the power grid.

[0111] 4. A graph neural network is used to model the topology of the power grid system, fully leveraging the adjacency matrix and node characteristic vectors to capture the complex interactions between resources. A dynamic information propagation mechanism based on weighted propagation coefficients is proposed, which dynamically adjusts the information propagation path and intensity based on the physical location and characteristics of resources, improving the modeling capabilities and solution efficiency of complex systems. This approach maintains computational efficiency in high-dimensional and nonlinear scenarios, adapting to the operational requirements of large-scale distributed power grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0112] Figure 1 is a flow chart of the overall steps of the present invention;

[0113] Figure 2 This is a flow chart of step S1 of the present invention;

[0114] Figure 3 Solve the flow chart for step S4 of the present invention;

[0115] Figure 4 This is an example diagram of the output method of the classification results of the present invention;

[0116] Figure 5 This is an example diagram of the global optimization power grid dispatching strategy of the present invention. DETAILED DESCRIPTION

[0117] In the description of the present application, the meaning of "multiple" is at least two, such as two, three, etc., unless otherwise clearly and specifically limited. In the embodiments of the present application, all directional indications (such as up, down, left, right, front, back, top, bottom ...) are only used to explain the relative position relationship, movement, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication also changes accordingly. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or Internet of Things terminal that includes a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or Internet of Things terminals.

[0118] In addition, references to "embodiments" herein mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of such phrases in various places in the specification does not necessarily refer to the same embodiment, nor does it necessarily refer to independent or alternative embodiments that are mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0119] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. It should be noted that the numerical values described in this application should not be fixed values. The numerical values are only example values and should not be limited to the scope of protection of this application. The relevant numerical data of this application can be determined according to the specific situation of the power grid.

[0120] like Figure 1 As shown, the present invention provides a hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads and storages, including the following steps:

[0121] S1, divides the control levels of various aggregated distributed resources of source, load and storage;

[0122] S2, based on the control level classification result, aggregate resources of the same control level into the same resource cluster, and build a distributed resource cluster with multiple aggregations of source, load and storage;

[0123] S3, establish a distributed resource cluster collaborative control model;

[0124] S4, introduce a graph neural network model to solve the distributed resource cluster collaborative control model.

[0125] Furthermore, if Figure 2 As shown, step S1 specifically includes the following steps:

[0126] In the source-load-storage resource control level division, data collection and preprocessing are performed on the various aggregated distributed resources of the source-load-storage; then, the characteristics of the various aggregated distributed resources of the source-load-storage are analyzed to extract important characteristics that affect resource classification; finally, an improved fuzzy clustering algorithm (Fuzzy C-Means, FCM) is used to classify the various aggregated distributed resources of the source-load-storage into three levels: level one (fast response), level two (medium response) and level three (slow response) resources.

[0127] Among them, in data collection and preprocessing, operating data is collected from distributed sources (photovoltaic, wind power), loads (industrial, commercial, residential) and energy storage devices (battery energy storage) to provide a basis for subsequent classification.

[0128] The collection method is as follows:

[0129] Device access: Collect resource operation data through IoT terminals and adopt unified communication protocols (such as Modbus and LoRa).

[0130] Data types collected: power data, real-time output power P(t) and reactive power Q(t); response speed, the time T from receiving the instruction to completing the action; economy, unit adjustment cost C unit .

[0131] The data preprocessing process is as follows:

[0132] 1. Outlier processing: Eliminate outliers based on the 3σ principle of mean and standard deviation:

[0133] x∈[μ-3σ,μ+3σ]

[0134] Among them, μ is the mean, σ is the standard deviation, and x is the normal data.

[0135] 2. Missing value filling: use linear interpolation to fill missing values

[0136]

[0137] Standardization processing: Perform min-max normalization processing on numerical data to convert it into standard data.

[0138]

[0139] Then, resource characteristics analysis is performed to extract important characteristics that affect resource classification and provide input for subsequent algorithms. The characteristics analysis includes regulation capacity analysis and economic analysis. The process is as follows:

[0140] 1. Adjustment capacity analysis is used to evaluate the adjustability of resources and determine their ability to play a role in rapid response.

[0141] The adjustable power range of computing resources within the normal operating range is:

[0142] ΔP max =P rated -P min

[0143] Where: P rated is the rated power of the resource, P min The minimum operating power of the resource.

[0144] Measures the time it takes for a resource to reach its target state after receiving a control instruction:

[0145] T resp =t end -t start

[0146] Where: t start is the control signal sending time, t end Time to complete adjustments for resources

[0147] According to T resp The response speed is divided into 3 levels, where T resp ≤1s is fast response, 1s <T resp ≤10s is medium speed response, T resp >10s is a slow response.

[0148] Based on the collected dynamic curve P(t) of power change over time, we can determine whether the regulation capability changes linearly over time:

[0149] 1) According to the change of battery charging and discharging power P(t) over time and load power P load (t) Over time, the energy storage charge and discharge curve and the load response curve are constructed. The energy storage charge and discharge curve shows the change of the battery's charge and discharge power over time, with time on the horizontal axis and power on the vertical axis; the load response curve shows the change of load power over time, with time on the horizontal axis and power on the vertical axis.

[0150] 2) Assuming the adjustment curve is a linear relationship:

[0151] P(t)=a·t+b

[0152] Where a is the power change rate and b is the initial power.

[0153] The fitting parameters a and b are calculated using the least squares method to obtain the model fitting value P fit (t).

[0154] 3) Calculate the coefficient of determination R 2 :

[0155]

[0156] Where, P i is the observed value, i.e. the power value actually measured at the i-th time point; P fit,i is the model fitting value, that is, the power value at the i-th time point predicted by the linear model; is the mean of the power observations.

[0157] If R 2 ≥0.95, indicating that the power shows a linear trend.

[0158] 4) Calculate the residual e i =P i -P fit,i ,If there is no obvious pattern in the residuals, it means that the ,adjustment ability changes linearly over time.

[0159] 2. Economic analysis: determine the economic priority of resource regulation by evaluating the regulation costs and benefits.

[0160] The average cost of computing resource per unit power regulation C unit :

[0161]

[0162] Where ΔP max is the adjustable power range of the resource, and the operating cost is the energy consumption or operation and maintenance fee during the resource adjustment process.

[0163] The benefit of computing resource per unit power regulation R unit :

[0164]

[0165] The load response benefit is the compensation obtained by resources participating in grid ancillary services (such as load peak shaving and valley filling).

[0166] Calculate the adjusted return R unit and adjustment cost C unit The ratio E ratio , used to evaluate the economic feasibility of resources.

[0167]

[0168] E ratio >1, indicating high resource economy and priority adjustment; E ratio ≤1 indicates low resource economy and low priority. The extracted important features are integrated into a unified feature vector to provide input for the subsequent classification algorithm:

[0169] X i =[ΔP max ,T resp ,R 2 ,e i ,C unit ,R unit ,E ratio ]

[0170] Where ΔP max is the resource adjustable power range, T resp To measure the time from when the resource receives the control instruction to when it reaches the target state, R 2The coefficient of determination for determining whether the regulatory capacity changes linearly with time, e i To determine whether the regulatory ability changes linearly over time, C unit is the unit adjustment cost, R unit E is the benefit of computing resource per unit power regulation. ratio R unit with C unit The ratio of

[0171] Finally, an improved fuzzy clustering algorithm is used to classify resources. Based on the fuzzy clustering algorithm, weighted characteristic measurements and characteristic constraints are introduced to achieve accurate classification of source, load and storage resources. Specifically, the following are included:

[0172] According to the different characteristics of the various aggregated distributed resources (such as power range, response speed, and economy) of the source, load, and storage, the importance of resource classification is different, and a weight w is assigned to each characteristic. j , the improved distance calculation formula is:

[0173]

[0174] Where D ik is the weighted distance from resource i to cluster center k, w j is the weight of feature j, x ij is the value of resource i on feature j, μ kj is the value of cluster center k on feature j;

[0175] Use the entropy method to assign weights to each feature: Among them H j is the information entropy of feature j;

[0176] Add constraints to the optimization objective:

[0177]

[0178] Where, f con is the characteristic constraint item, which represents the restriction on resource economy and response speed, λ is the constraint weight, is the square of the distance from the resource to the cluster center on feature j.

[0179] The process of improving the fuzzy clustering algorithm to classify resources is as follows:

[0180] 1. Initialize parameters, set the number of clusters to 3, and initialize the membership matrix U (0) , set the fuzzy index to 2 and define the feature constraints.

[0181] 2. Iterative optimization:

[0182] Calculate weighted distance:

[0183] According to the current membership matrix U, update each cluster center:

[0184] According to the new cluster center, update the membership matrix:

[0185] Update the objective function:

[0186] If the change of the objective function J or the membership matrix U is less than the threshold ∈, the iteration is stopped; otherwise, return to step 1 to continue.

[0187] 3. Output classification results

[0188] like Figure 4 As shown, according to the membership matrix, each resource is assigned to the category with the largest membership.

[0189] Furthermore, in step S2, according to the control level division results (fast response, medium response, slow response), resources of the same control level are aggregated into the same cluster, and three distributed resource clusters, namely, fast response cluster, medium response cluster, and slow response cluster, are obtained to ensure efficient coordination and overall optimization of resources.

[0190] Furthermore, in step S3, based on the aggregation results, a distributed resource cluster collaborative control model is established. The model mainly consists of two parts: one is the objective function with the safe operation of the distribution network as the core, and the other is related constraints including three types of response cluster dynamic constraints, specifically including:

[0191] The first is the objective function, which takes the safe operation of the distribution network as the core goal and optimizes the following three goals:

[0192] 1) Minimize node voltage deviation:

[0193]

[0194] Where J1 is the square sum of node voltage deviation, V i is the voltage of the i-th node, V ref is the reference voltage, set to 1.0pu, and N is the total number of nodes in the distribution network;

[0195] 2) Minimize the active power loss of the line:

[0196]

[0197] Where J2 is the total power loss of the line, R ij is the resistance of line ij, I ij is the current of line ij, It is the collection of all lines in the distribution network;

[0198] 3) Optimal load balancing:

[0199]

[0200] Where, J3 is the sum of the absolute values of the power deviations of each phase, P φ is the total active power of phase φ, P total is the three-phase total active power, P total =∑ φ∈ { A,B,C}P φ ;

[0201] According to the above three objectives, the comprehensive objective function is established as follows:

[0202] minJ=α1J1+α2J2+α3J3

[0203] Where α1, α2, and α3 are target weight coefficients, which are used to balance the importance of voltage stability, line loss, and load balance.

[0204] The model constraints are set as follows:

[0205] 1) Power Constraints:

[0206]

[0207] Where, P i With Q i are the active power output and reactive power output of resource i, and are the upper and lower limits of active power output of resource i, and are the upper and lower limits of reactive power of resource i, and G is the set of all resources;

[0208] 2) Voltage constraint:

[0209]

[0210] Where V i is the voltage amplitude at node i, and are the upper and lower limits of the voltage of node i, and N is the set of all nodes;

[0211] 3) Line flow constraints:

[0212]

[0213] Where S ij is the apparent power of line ij, is the rated capacity of line ij, is the set of all transmission lines in the network;

[0214] 4) Cluster dynamic constraints:

[0215] Fast response cluster response time constraints:

[0216]

[0217] Where, T resp,i is the response time of fast response resource i, T req,1 is the maximum response time of the fast response cluster, G i A collection of resources for rapid response clusters;

[0218] Medium-speed response cluster dynamic power allocation constraints:

[0219]

[0220] Where, P req,2 is the target active power of the medium-speed response cluster, ΔP i is the adjustable power range of resource i

[0221] G2 is a resource collection of a medium-speed response cluster;

[0222] Slow cluster planning adjustment constraints:

[0223]

[0224] Where, P plan With Q plan are the planned active and reactive powers of the slow-response cluster, respectively.

[0225] In summary, combining the objective function and constraints, the final form of the coordinated control model is:

[0226] minJ=α1J1+α2J2+α3J3

[0227]

[0228] Furthermore, in step S4, a method for solving a distributed resource cluster collaborative control model based on a graph neural network model is proposed.

[0229] Among them, the graph neural network model construction method is as follows:

[0230] First, based on the distributed resource cluster collaborative control model, a graph structure is constructed to represent the relationship between resource clusters. The graph G = (V, E), where V represents the node set in the graph and E represents the edge set in the graph.

[0231] Each node v k ∈V represents a resource cluster G k , where the state of each resource cluster is represented by a feature vector: in For cluster G k The initial feature vector contains information such as the power, response time, control unit capacity, and resource type of the cluster.

[0232] Each edge e ij Represents the relationship or interaction between resource clusters. It can represent power flow, resource sharing, or scheduling collaboration between clusters. Each edge has additional features, such as power flow, regulation amount, response speed, etc.

[0233] e ij =[ΔP ij ,ΔQ ij ,T flow,ij ]

[0234] where ΔP ij and ΔQ ij Represents node v i and v j The power difference or reactive power flow, T flow,ij Represents the response time of power flow.

[0235] Use the adjacency matrix A to represent the structural relationship of the graph, A ij Represents node v i and v j Is there a connection between A ij =1 means there is a connection, A ij =0 means no connection.

[0236] Then, based on the definition of the graph structure, a new information dissemination mechanism was designed:

[0237] To accommodate the varying levels of control and response characteristics among resource clusters, we introduced weighted information dissemination and a dynamic adjustment mechanism for dissemination intensity, building upon the traditional information dissemination mechanism. Inter-cluster information dissemination not only relies on adjacency but also enhances its expressiveness through the following two mechanisms:

[0238] 1) Weighted propagation:

[0239] At each information propagation, the connection weight λ between clusters ij It will be dynamically adjusted according to its response characteristics, constraints, etc. This weighted propagation mechanism ensures that the weights of information between different clusters are tailored according to their importance and characteristics. The specific definitions are as follows:

[0240] λij =α·f(r i ,r j )+β·g(c i ,c j )+γ·h(d ij )

[0241] Where: r i ,r j Represent the response characteristics of clusters i and j (corresponding to the three categories obtained by clustering above), c i ,c j represents the constraints between clusters i and j, d ij represents the topological adjacency relationship between clusters i and j, f, g, and h are mapping functions used to measure the similarity between different features, and α, β, and γ are hyperparameters that control the importance of each feature in the weight calculation.

[0242] 2) Dynamically adjust communication intensity:

[0243] According to the control level of the resource cluster (fast response, medium response, slow response) and the constraints between clusters, the intensity and frequency of information dissemination of each cluster are dynamically adjusted to ensure that the information dissemination between clusters matches their regulation capabilities.

[0244] Assume that in the information propagation of the kth layer, node v i Need to get from its adjacent node v j Receive information and re-update its own status, then the node information update formula is:

[0245]

[0246] in: is node v i In the feature representation of the kth layer, N i is node v i The set of adjacent nodes, w ij For edge e ij The weight coefficient of node v i Its adjacent node v j The weighting coefficient is dynamically adjusted based on the cluster's responsiveness, control level, and other constraints (such as power constraints, response time, etc.), and f is the ReLU function.

[0247] Furthermore, the weighted propagation coefficient w ij Adjust according to the following rules:

[0248] 1) Responsiveness Weighting

[0249] If v i and v jBelong to clusters with similar control levels, then the weight w ij A larger value indicates a stronger intensity of information transmission.

[0250]

[0251] Where T resp,i and T resp,j is node v i and v j The response time is α, and α is the adjustment coefficient.

[0252] 2) Constrained matching weighting

[0253] For resource clusters with constraints (such as power and economic constraints), information propagation between nodes will dynamically adjust the weights based on their power range and constraint similarity:

[0254]

[0255] Among them, P max,i and P max,j Node v i and v j The maximum power of , β is the adjustment coefficient, which is used to control the influence of power difference on weight.

[0256] Furthermore, the graph neural network model is trained and optimized:

[0257] 1) Training Dataset

[0258] Taking the initial characteristic vector of each node as input (including response time, power range, etc.), the final adjustment strategy and scheduling results of each resource cluster are output through collaborative optimization of the nodes.

[0259] 2) Loss Function

[0260] To optimize the objective function and ensure that the model effectively learns the coordination between resource clusters, the following loss function is defined:

[0261] Loss of voltage stability:

[0262]

[0263] Power balance loss:

[0264]

[0265] To ensure that the cluster operates within the control capability, the constraint satisfaction loss is defined:

[0266]

[0267] The total loss function is:

[0268] L total＝ L V +L P +L C

[0269] 3) Training and optimization

[0270] Backpropagation and gradient descent are used to optimize weights and node feature vectors. The Adam optimizer is used to accelerate the training process. Backpropagation and gradient descent are commonly used in neural network training, and the Adam optimizer is a commonly used tool in neural network optimization.

[0271] Furthermore, after completing the training of the graph neural network model, the distributed resource cluster collaborative control model is solved, such as Figure 3 As shown, the specific steps for solving are:

[0272] 1. Initialize the feature vector of each node Set the edge weighted propagation coefficient w of the graph ij ;

[0273] 2. Information propagation and node update: Perform multiple rounds of iterative updates of the graph neural network model and calculate the node feature vectors through weighted information propagation;

[0274] 3. Constraint optimization: After each iteration, constraint optimization is performed to ensure that the resource cluster scheduling results meet the safety and economic requirements of the power grid;

[0275] 4. Loss function minimization: Training is performed based on the objective function and loss function to gradually optimize the coordinated control between resource clusters;

[0276] 5. Output results: Output the final dispatching strategy of each resource cluster to ensure the safe, stable and efficient operation of the entire power grid system. Among them, the output final dispatching strategy mainly includes the following information: the fast response cluster adjusts power at a high frequency (seconds) to respond to load changes or power grid events to meet high dynamic demands in a short period of time and avoid frequency and voltage collapse; the medium response cluster adjusts power at the minute level to dynamically match load demands and reduce dispatching costs; the slow response cluster adjusts power output at a low frequency to ensure a smooth match of long-term power grid load demands. Figure 5 As shown in the figure, a globally optimized grid dispatching strategy is ultimately formed through the coordinated control of fast, medium and slow response clusters.

[0277] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in this application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are necessary for each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, and are not limitations. The above details do not limit this application to being implemented by adopting the above specific details. The above description disclosed is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be very obvious to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but to the widest scope consistent with the principles and novel features of this invention.

[0278] The above is only a preferred embodiment of the invention of this application and is not intended to limit the invention of this application. Any modifications, equivalent replacements, etc. made within the spirit and principles of the invention of this application should be included in the scope of protection of the invention of this application.

[0279] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads and storage, characterized in that: The following steps are involved: S1, divides the control levels of various aggregated distributed resources of source, load and storage; S2, based on the control level classification result, aggregate resources of the same control level into the same resource cluster, and build a distributed resource cluster with multiple aggregations of source, load and storage; S3, establish a distributed resource cluster collaborative control model; S4, introduce a graph neural network model to solve the distributed resource cluster collaborative control model.

2. The hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads, and storages according to claim 1 is characterized in that: The step S1 specifically includes the following steps: S101, performing data collection and preprocessing on the multiple aggregated distributed resources of the source, load and storage; S102, performing characteristic analysis on the multiple aggregated distributed resources of source, load and storage to extract important characteristics that affect resource classification; the characteristic analysis includes regulation capability analysis and economic analysis, and integrating the important characteristics into a unified characteristic vector: X i =[ΔP max ,T resp ,R 2 ,e i ,C unit ,R unit ,E ratio ] Where ΔP max is the resource adjustable power range, T resp To measure the time from when the resource receives the control instruction to when it reaches the target state, R 2 The coefficient of determination for determining whether the regulatory capacity changes linearly with time, e i To determine whether the regulatory ability changes linearly over time, C unit is the unit adjustment cost, R unit E is the benefit of computing resource per unit power regulation. ratio R unit with C unit The ratio of S103, using an improved fuzzy clustering algorithm to classify the multiple aggregated distributed resources of source, load and storage; S104, classifying the multiple aggregated distributed resources of source, load and storage into three levels: level one fast response, level two medium response and level three slow response.

3. The hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads and storages according to claim 2 is characterized in that: The improved fuzzy clustering algorithm is based on the fuzzy clustering algorithm and introduces weighted feature measurement and feature constraint conditions to improve it. Specifically, it includes: According to the different characteristics of the various aggregated distributed resources of source, load and storage, the importance of resource classification is different, and a weight w is assigned to each characteristic. j , the improved distance calculation formula is: Where D ik is the weighted distance from resource i to cluster center k, w j is the weight of feature j, x ij is the value of resource i on feature j, μ kj is the value of cluster center k on feature j; Use the entropy method to assign weights to each feature: Among them H j is the information entropy of feature j; Add constraints to the optimization objective: Where, f con is the characteristic constraint term, λ is the constraint weight, is the square of the distance from the resource to the cluster center on feature j.

4. The hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads, and storages according to claim 1 is characterized in that: The distributed resource cluster collaborative control model includes an objective function and constraints, specifically including: With the safe operation of the distribution network as the core goal, the following three goals are optimized: 1) Minimize node voltage deviation: Where J1 is the square sum of node voltage deviation, V i is the voltage of the i-th node, V ref is the reference voltage, N is the total number of nodes in the distribution network; 2) Minimize the active power loss of the line: Where J2 is the total power loss of the line, R ij is the resistance of line ij, I ij is the current of line ij, It is the collection of all lines in the distribution network; 3) Optimal load balancing: Where, J3 is the sum of the absolute values of the power deviations of each phase, P φ is the total active power of phase φ, P total is the three-phase total active power, P total =∑ φ∈{A,B,C} P φ ; According to the three objectives, the objective function is established as: minJ=α1J1+α2J2+α3J3 Where α1, α2, and α3 are target weight coefficients; The constraints include: 1) Power Constraints: Where, P i With Q i are the active power output and reactive power output of resource i, and are the upper and lower limits of active power output of resource i, and are the upper and lower limits of reactive power of resource i, and G is the set of all resources; 2) Voltage constraint: Where V i is the voltage amplitude at node i, and are the upper and lower limits of the voltage of node i, and N is the set of all nodes; 3) Line flow constraints: Where S ij is the apparent power of line ij, is the rated capacity of line ij, is the set of all transmission lines in the network; 4) Cluster dynamic constraints: Fast response cluster response time constraints: Where, T resp,i is the response time of fast response resource i, T req,1 is the maximum response time of the fast response cluster, G i A collection of resources for rapid response clusters; Medium-speed response cluster dynamic power allocation constraints: Where, P req,2 is the target active power of the medium-speed response cluster, ΔP i is the adjustable power range of resource i G2 is a resource collection of a medium-speed response cluster; Slow cluster planning adjustment constraints: Where, P plan With Q plan are the planned active and reactive powers of the slow-response cluster, respectively.

5. The hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads, and storages according to claim 4 is characterized in that: Establishing the graph neural network model includes the following steps: S401, constructing a graph structure according to the distributed resource cluster collaborative control model: Graph G = (V, E), where V represents a set of nodes in the graph, and E represents a set of edges in the graph; Each node v k ∈V represents a resource cluster G k , each of the resource clusters G k The state of is represented by the feature vector: Where, For resource cluster G k The initial characteristic vector, P k With Q k They are resource cluster G k Active power output and reactive power output, T resp,k For resource cluster G k The response time, C unit,k For resource cluster G k Unit adjustment cost, R unit,k is the computing resource cluster G k The benefit of each unit of power regulation, E ratio,k R unit,k with C unit,k The ratio of Each edge e ij Represents the resource cluster G k Relationship or interaction between: e ij =[ΔP ij ,ΔQ ij ,T flow,ij ] Where ΔP ij and ΔQ ij Represents node v i and v j The power difference or reactive power flow, T flow,ij represents the response time of power flow; Use the adjacency matrix A to represent the structural relationship of the graph, A ij Represents node v i and v j Is there a connection between A ij =1 means there is a connection, A ij =0 means no connection; S402: Based on the graph structure, a new information dissemination mechanism is introduced; the new information dissemination mechanism is based on the traditional information dissemination mechanism and introduces a mechanism of weighted information dissemination and dynamic adjustment of dissemination intensity; S403, a training data set, taking the initial characteristic vector of each node as input, and outputting the final adjustment strategy and scheduling result of each resource cluster through collaborative optimization of the nodes; S404, defining a loss function to optimize the objective function and ensure that the model effectively learns the coordination between the resource clusters; S405, using backpropagation and gradient descent to optimize weights and node feature vectors, and using the Adam optimizer to accelerate the training process; S406, completing the training of the graph neural network model.

6. The hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads, and storages according to claim 5 is characterized in that: The weighted information propagation is specifically as follows: At each information dissemination, the connection weight λ between the resource clusters ij Can be dynamically adjusted to: l ij =α·f(r i ,r j )+β·g(c i ,c j )+γ·h(d ij ) Where r i ,r j represent the response characteristics of clusters i and j, respectively, c i ,c j represents the constraints between clusters i and j, d ij represents the topological adjacency relationship between clusters i and j, f, g, and h are mapping functions used to measure the similarity between different features, and α, β, and γ are hyperparameters that control the importance of each feature in the weight calculation.

7. The hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads, and storages according to claim 5 is characterized in that: The dynamic adjustment of the dissemination intensity is specifically as follows: According to the control level of the resource cluster and the constraints between the resource clusters, the intensity and frequency of information dissemination of each resource cluster are dynamically adjusted. Assume that in the information dissemination of the kth layer, node v i Need to get from its adjacent node v j Receive information and re-update its own status. The node information update formula is: Where, is node v i In the feature representation of the kth layer, N i is node v i The set of adjacent nodes, w ij For edge e ij The weighted propagation coefficient of node v i Its adjacent node v j The information propagation intensity between them is represented by , and f is the ReLU function.

8. The hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads, and storages according to claim 7 is characterized in that: The weighted propagation coefficient w ij Adjust according to the following rules: 1) Responsiveness Weighting Where, T resp,i and T resp,j is node v i and v j response time, α is the adjustment coefficient; if v i and v j Belong to resource clusters with similar control levels, then w ij Larger, indicating greater intensity of information transmission; 2) Constrained matching weighting For constrained resource clusters, information dissemination between nodes will dynamically adjust weights: Where, P max,i and P max,j Node v i and v j The maximum power of , β is the adjustment coefficient, which is used to control the influence of power difference on weight.

9. The hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads, and storages according to claim 5 is characterized in that: The loss function is specifically: Loss of voltage stability: Power balance loss: To ensure that the resource cluster operates within the control capability, the constraint satisfaction loss is defined: The total loss function is: L total＝ L V +L P +L C 。 10. The hierarchical collaborative control method for a distributed resource cluster with multiple aggregations of sources, loads, and storages according to claim 5, characterized in that: The solution steps in step S4 are specifically: S407, initializing the characteristic vector of each node Set the edge weighted propagation coefficient w of the graph ij ; S408, information propagation and node update: performing multiple rounds of iterative updates of the graph neural network model, and calculating the characteristic vectors of the nodes through the weighted information propagation; S409, constraint optimization: after each round of iteration, the constraint optimization is performed to ensure that the scheduling result of the resource cluster meets the safety and economic requirements of the power grid; S410, minimizing the loss function: performing training according to the objective function and the loss function, and gradually optimizing the coordinated control among the resource clusters; S411, output result: output the final scheduling strategy of each resource cluster.

Images