A power distribution network distributed resource collaborative control method based on a graph attention network

Abstract

The application discloses a power distribution network distributed resource collaborative control method based on a graph attention network. The method constructs an intelligent control framework of GAT and RLS fusion identification, effectively overcomes the limitations of traditional methods ignoring topological information and fixed parameter control adaptive ability, combines the topological perception ability of the GAT network and the parameter recursive ability of the RLS method, realizes accurate identification and intelligent control of the complex operation state of the power distribution network, and introduces a multi-head attention mechanism and adaptive control parameter adjustment, so that the system can adaptively adjust the control strategy according to the network topology and operation state, and significantly improves the intelligent level and regulation and control precision of the control framework.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent collaborative control of power systems, and particularly relates to a distribution network distributed resource collaborative control method based on a graph attention network. BACKGROUND

[0002] With the increase of new energy installed capacity ratio, the distribution network gradually changes from the mode of "source following load" to the mode of "source-load interaction", and the load gradually begins to participate in the grid regulation. Distributed flexible resources continue to grow on the distribution network side, and how to fully tap the regulation characteristics of these flexible resources to meet the needs of new power systems has become a hot research issue.

[0003] Distributed resources (DER) in new power systems are of various types, such as air conditioning load, energy storage system, electric vehicle charging pile, and distributed photovoltaic. These distributed resources are mainly located at the end of the distribution network, enabling the user side to have certain flexible regulation capability. However, due to the characteristics of small individual capacity, scattered distribution, and large quantity of these DERs, and the complexity of their load characteristics with strong coupling and high dimensionality, traditional centralized control methods are difficult to achieve effective collaborative regulation.

[0004] Load characteristic identification is the basis for achieving accurate voltage control. In terms of parameter identification algorithms, recursive least squares (RLS) is one of the most widely used methods. Distribution network voltage control not only needs to maintain voltage within the qualified range, but also needs to coordinate various DERs to achieve optimal system operation. As a new deep learning technology, graph attention network (GAT) performs well in processing graph structure data.

[0005] However, the existing research has the following shortcomings: (1) GAT is rarely applied to load characteristic identification and collaborative control of distribution networks; (2) distributed resource collaborative control lacks intelligent decision-making mechanism, making it difficult to handle replication control problems, and the control effect needs to be improved; (3) the load characteristic identification method does not fully utilize the topology structure information of the distribution network, and the identification accuracy needs to be improved. SUMMARY

[0006] To solve the problems of load characteristic identification of distributed resources on the distribution network side of new power systems and the difficulty of traditional control methods to achieve intelligent collaboration of multiple devices, the application provides a distribution network distributed resource collaborative control method based on a graph attention network. Through the fusion identification of GAT and recursive least squares and the weight learning of the multi-head attention mechanism, accurate identification of load characteristics and intelligent control of voltage quality are achieved, effectively solving the problems of insufficient utilization of topology information and poor collaborative control effect of traditional methods.

[0007] According to an aspect of the specification of the present application, a power distribution network distributed resource collaborative control method based on a graph attention network is provided, comprising:

[0008] A load characteristic identification system integrating topological information is constructed, a ZIP load model is designed to describe the voltage sensitivity characteristics of different types of loads, and a distributed resource model is established;

[0009] Load parameter identification is performed based on a graph attention network, the graph attention network designs four types of physically meaningful attention heads, namely electrical distance attention, load similarity attention, control coordination attention and time series prediction attention, mines the correlation between nodes through a multi-head attention mechanism, and maps the feature vectors output by the multiple attention heads into ZIP coefficient prediction values;

[0010] The graph attention network output is adaptively fused with the ZIP coefficient identification result based on the recursive least squares method to obtain a graph attention enhanced ZIP coefficient identification result;

[0011] The attention weights of the graph attention network realize adaptive control of the dead zone and the control strength, and trigger the control decision when the voltage of the current node exceeds the adaptive dead zone range;

[0012] Based on the graph attention enhanced ZIP coefficient identification result, a control link for adaptive control interval adjustment, device coordination optimization and safety constraint protection is established, and based on the control decision, intelligent collaborative regulation and control of air conditioning loads, energy storage systems, electric vehicle charging piles and distributed photovoltaics are realized.

[0013] As a further technical solution, the construction of the graph attention network comprises:

[0014] The graph structure representation of the power distribution network is constructed, the nodes of the power distribution network are taken as the vertices of the graph, the branch connection relationship is taken as the edge of the graph, and an adjacency matrix is formed to describe the network topology;

[0015] The node feature vector is designed, and the active power, reactive power, voltage, load type identifier, position encoding and time encoding information of the node are integrated;

[0016] The attention coefficient calculation mechanism of the graph attention network is established, and the correlation between node pairs is calculated through an attention function and a linear transformation matrix;

[0017] The softmax function is used to normalize the attention coefficients to obtain comparable attention weights;

[0018] Based on the attention weights, the weighted aggregation of the neighbor node features is realized to generate the output feature representation of the node.

[0019] As a further technical solution, on the basis of the constructed graph attention network, multiple physical attention head designs are introduced, including:

[0020] An electrical distance attention head is designed to calculate the spatial correlation weight according to the electrical distance between nodes;

[0021] A load similarity attention head is designed to calculate the load similarity coefficient based on the load type and power characteristics;

[0022] A control coordination attention head is designed to evaluate the control coordination potential and adjustment ability matching degree between nodes;

[0023] A timing prediction attention head is designed to mine the timing correlation between nodes;

[0024] The outputs of multiple attention heads are spliced or average-pooled to form a comprehensive node feature representation.

[0025] As a further technical solution, the training of the graph attention network includes:

[0026] A two-layer graph attention network structure is constructed, the first layer performs preliminary topological feature extraction, and the second layer performs deep semantic feature learning;

[0027] An output layer structure is designed to map the graph attention network feature vector to the ZIP coefficient prediction value through a fully connected layer;

[0028] An mean square error loss function is used to optimize the graph attention network parameters on the training set to obtain a trained graph attention network.

[0029] As a further technical solution, the attention weight based on the graph attention network realizes the adjustment of the adaptive control dead zone and the control strength, including:

[0030] The trigger threshold of voltage control is dynamically adjusted according to the maximum attention weight, and the control strength is calculated based on the attention strength and voltage deviation.

[0031] As a further technical solution, after triggering the control decision, the following multi-device collaborative control signal distribution is performed:

[0032] According to the voltage deviation direction, it is judged whether the power injection needs to be increased or decreased;

[0033] Based on the device priority, adjustable capacity and attention factor, the control signal distribution proportion of each device is calculated;

[0034] Specific power regulation instructions are distributed to energy storage systems, electric vehicle charging piles, distributed photovoltaic and air conditioning loads;

[0035] Voltage correction calculation is performed to predict the voltage adjustment effect based on the attention weight.

[0036] As a further technical solution, an adaptive control interval adjustment, device coordination optimization and safety constraint protection control link is established, including:

[0037] The control interval is dynamically adjusted according to the attention weight and the system state, and the control interval is shortened when the attention weight is high, and the control interval is lengthened when the attention weight is low;

[0038] Intelligent coordination and cooperation among multiple devices are realized through the attention weight, and the power distribution among devices is optimized;

[0039] Physical constraint conditions including power limit, ramp rate limit and response delay are set to ensure that the control signal is within a safe range;

[0040] A regulation suppression mechanism is designed to suppress the control action when the voltage change trend is opposite to the control direction.

[0041] As a further technical solution, a distributed resource model is established, further comprising:

[0042] A distributed photovoltaic output model is established, which comprehensively considers the influence of solar radiation intensity, cloud influence factor and weather change factor on photovoltaic power generation power;

[0043] A state of charge dynamic model of the energy storage system is established to describe the change rule of SOC in the charging and discharging process and the time-varying characteristics of charging and discharging efficiency;

[0044] A power model of the electric vehicle charging station is established, which comprehensively considers the influence of the number of charging vehicles, the AC / DC charging ratio and the power factor on the charging power.

[0045] According to one aspect of the present application, a power distribution network distributed resource collaborative control system based on a graph attention network is provided for implementing the method, comprising:

[0046] The first main module is used to build a load characteristic identification system integrating topological information, design a ZIP load model to describe the voltage sensitivity characteristics of different types of loads, and establish a distributed resource model;

[0047] The second main module is used for load parameter identification based on a graph attention network, wherein the graph attention network designs four types of physically meaningful attention heads, i.e. electrical distance attention, load similarity attention, control coordination attention and time series prediction attention, excavates the correlation between nodes through a multi-head attention mechanism, and maps the feature vectors output by the multiple attention heads into ZIP coefficient prediction values;

[0048] The third main module is used for adaptively fusing the graph attention network output and the ZIP coefficient identification result based on the recursive least square method to obtain a graph attention enhanced ZIP coefficient identification result.

[0049] The fourth main module is used for adaptively controlling the adjustment of the dead zone and the control strength based on the attention weight of the graph attention network, and triggering a control decision when the current node voltage exceeds the adaptive dead zone range.

[0050] The fifth main module is used for establishing a control link of adaptive control interval adjustment, device coordination optimization and safety constraint protection based on the graph attention enhanced ZIP coefficient identification result, and realizing intelligent collaborative regulation and control of air conditioning load, energy storage system, electric vehicle charging pile and distributed photovoltaic based on the control decision.

[0051] According to an aspect of the present application, a non-transitory computer readable storage medium is provided, which stores computer instructions for executing the power distribution network distributed resource collaborative control method based on the graph attention network.

[0052] Compared with the prior art, the present application has the following advantages:

[0053] 1. The load characteristic identification system combining topology awareness and parameter recursion is constructed by fusing and identifying the graph attention network GAT and the recursive least square method RLS, effectively solving the problem that the traditional method ignores the network topology information and spatial correlation, and significantly improving the ZIP coefficient identification accuracy and load characteristic tracking ability.

[0054] 2. Four types of physically meaningful attention heads, i.e. electrical distance attention, load similarity attention, control coordination attention and time series prediction attention, are introduced to improve the understanding ability of the GAT network to the physical characteristics of the power distribution network. An adaptive control dead zone and control strength calculation mechanism based on attention weight is designed to enable the system to intelligently adjust the control strategy according to the network state, thereby enhancing the adaptive ability and regulation and control precision of the collaborative control.

[0055] 3. Intelligent collaborative regulation and control of the energy storage system, electric vehicle charging pile, load and distributed photovoltaic is realized, and the grid performance is effectively improved through the multi-head attention mechanism and adaptive control parameter adjustment, thereby providing an effective technical means for voltage quality control and safe and economic operation of the power distribution network in the high proportion of new energy access scenario. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings used in the embodiments or prior art description will be briefly described below. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0057] Figure 1 is the overall architecture diagram of the power distribution network distributed resource collaborative control method based on the graph attention network in the embodiment of the present application.

[0058] Figure 2 is the principle diagram of the GAT multi-head attention mechanism in the embodiment of the present application.

[0059] Figure 3 is the schematic diagram of the GAT-RLS fusion identification mechanism in the embodiment of the present application.

[0060] Figure 4 is the adaptive control decision flowchart in the embodiment of the present application. DETAILED DESCRIPTION

[0061] In view of the technical status mentioned in the foregoing background art, a new type of power distribution network intelligent regulation and control method is urgently needed, which can effectively cope with the challenges of small individual capacity of distributed resources, scattered distribution and strong coupling of load characteristics, realize topology perception and collaborative optimization of multiple types of energy resources, and improve the voltage control accuracy and system operation economic level of the power distribution network. Based on this, the present application proposes a power distribution network distributed resource collaborative control method based on a graph attention network, which realizes accurate identification of load characteristics and intelligent control of voltage quality through the fusion identification of GAT and recursive least squares and the weight learning of the multi-head attention mechanism, effectively solving the problems of insufficient utilization of topology information and poor collaborative control effect of traditional methods.

[0062] The present application mainly based on graph neural network and attention mechanism, in order to improve the intelligent level of the collaborative control of the distributed resources on the side of the new power system power distribution network. The present application fully utilizes the advantages of GAT topology perception and multi-head attention weight learning, realizes accurate identification of load characteristics through the deep fusion of the two and recursive least squares, designs a multi-device collaborative control strategy for high proportion of new energy access scenarios, and realizes intelligent coordination of energy storage, charging piles, loads and photovoltaic. The control effect obtained by the present application is more accurate and more efficient, which effectively improves the voltage control capability of the power distribution network and the power quality guarantee level under the high proportion of new energy access scenarios.

[0063] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments but not all of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application. In addition, the technical features in each embodiment or in a single embodiment provided by the present application can be combined with each other at will to form new technical solutions, and the combination is not restricted by the sequence of steps and / or the mode of structural composition, but should be based on the fact that a person of ordinary skill in the art can realize it. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist and is not within the protection scope of the present application.

[0064] The technical solution of the present application provides a power distribution network distributed resource collaborative control method based on a graph attention network. The purpose is to improve the load characteristic identification accuracy and voltage collaborative control ability, realize the topology perception regulation of energy storage systems, electric vehicle charging piles, loads and distributed photovoltaics, and guarantee the power quality and safe operation of the power distribution network in the high proportion of new energy access scenarios.

[0065] The method provided by the present application can realize the process by computer simulation software technology. The system overall architecture diagram in the embodiment of the present application is as shown in Figure 1 The power distribution network distributed resource collaborative control method based on a graph attention network provided by the present application comprises the following steps:

[0066] Step 1: Power distribution network load and distributed resource modeling.

[0067] Step 1.1: Establish a ZIP load model to describe the voltage sensitivity characteristics of the load, decompose the active power of the load into three components of constant impedance, constant current and constant power, and set the ZIP coefficient parameters of the air conditioner load, charging pile load and ordinary load.

[0068] Specifically, P and Q are active and reactive power; V is the actual voltage, V0 is the rated voltage; P0 and Q0 are the power at the rated voltage; p0, p1 and p2 are the ZIP coefficients of active power, corresponding to constant power (Z), constant current (I) and constant impedance component (P) respectively; q0, q1 and q2 are the ZIP coefficients of reactive power.

[0069] The load characteristics in the power distribution network directly affect the effect of voltage control, and accurate load modeling is the basis for realizing accurate control. Here, the ZIP model is mainly taken as the research object, and its mathematical expression is as follows:

[0070] ​

[0071] The ZIP coefficient satisfies the following constraints:

[0072] .

[0073] The application sets three adjustable loads, respectively air conditioning load, charging pile load and ordinary load. Different types of loads show different voltage sensitivity characteristics due to their equipment characteristics and working principles. The air conditioning load is affected by the compressor and the fan, and has certain constant impedance characteristics, and the ZIP parameter is set to [0.35, 0.25, 0.40]; the charging pile load mainly shows constant power characteristics, and the ZIP parameter is set to [0.75, 0.15, 0.10]; the ordinary load is a mixture of residential and commercial loads, and the ZIP parameter is set to [0.45, 0.30, 0.25].

[0074] Step 1.2: Establish a distributed photovoltaic output model, considering the influence of solar radiation intensity, cloud cover factor and weather change factor on photovoltaic power generation.

[0075] Specifically, define P r as the rated installed capacity, η s as the system efficiency, η d as the derating factor, S(t) as the solar radiation intensity, C(t) as the cloud cover factor, and W(t) as the weather change factor. As a typical clean energy, distributed photovoltaic power generation is affected by many factors such as solar radiation intensity, cloud cover and weather conditions, and has time-varying and random characteristics. The mathematical model of the photovoltaic system can be represented by the following formula:

[0076] .

[0077] Step 1.3: Build a state of charge dynamic model of the energy storage system to describe the change rule of SOC in the charging and discharging process and the time-varying characteristics of charging and discharging efficiency.

[0078] Specifically, define SOC(t) as the state of charge at time t, P s as the energy storage power, Δt as the time step, E c as the energy storage capacity, and η(t) as the charging and discharging efficiency. As a bidirectional power regulation device, the mathematical model of the energy storage system can be represented by the following formula:

[0079] .

[0080] Step 1.4: Establish a power model of the electric vehicle charging station, considering the influence of the number of charging vehicles, the AC / DC charging ratio and the power factor on the charging power.

[0081] Specifically, define N c as the number of charging vehicles; Rac , R dc is the AC / DC charging ratio; P ac , P dc is the AC / DC charging power; F ac , F dc is the power factor. The electric vehicle charging station load has the characteristics of user behavior driving, and its charging demand presents a certain randomness, so the charging station power model can be expressed as follows:

[0082] .

[0083] Step 2: Recursive least squares load parameter identification.

[0084] Step 2.1: Construct a linear regression model for ZIP model parameter identification, and express the load active power as a function of voltage;

[0085] Step 2.2: Design a parameter update algorithm for recursive least squares, realize online estimation of ZIP coefficients through recursive calculation of gain matrix and covariance matrix, and introduce forgetting factor and regularization factor to enhance the tracking ability and numerical stability of the algorithm to the time-varying load characteristics;

[0086] Step 2.3: Output the ZIP coefficient identification results of RLS method as the reference for subsequent GAT fusion identification.

[0087] Specifically, define P load (k) as the load active power measurement value at the kth sampling time, is the regression vector at k, θ is the parameter vector to be estimated, e(k) is the estimation error at k, V(k) is the node voltage value at k; θ P0 , θ P1 , θ P2 are the coefficients of constant power, constant current and constant impedance terms in the ZIP model. RLS is a classical method for load characteristic identification, and for ZIP model parameter identification, a linear regression model can be constructed, which can be represented by the following formula:

[0088] .

[0089] Define as the parameter estimation value at k; K(k) is the gain matrix at k, λ is the forgetting factor, R f is the regularization factor, P(k) is the covariance matrix at k, and I is the unit matrix. Then the RLS recursive update can be represented by the following formula:

[0090] ,

[0091] ,

[0092] .

[0093] Step 3: Graph attention network architecture design.

[0094] Step 3.1: Build a graph structure representation of the power distribution network, taking the power distribution network nodes as the graph vertices and the branch connection relationship as the graph edges to form an adjacency matrix to describe the network topology.

[0095] Step 3.2: Design a node feature vector that integrates active power, reactive power, voltage, load type identification, location encoding, and time encoding information of the node.

[0096] Step 3.3: Establish an attention coefficient calculation mechanism for the graph attention network to calculate the correlation between node pairs through an attention function and a linear transformation matrix.

[0097] Step 3.4: Normalize the attention coefficients using a softmax function to obtain comparable attention weights.

[0098] Step 3.5: Realize weighted aggregation of neighbor node features based on the attention weights to generate the output feature representation of the node.

[0099] Specifically, define H (l) as the node feature matrix of the l-th layer, W (l) as the learnable weight matrix, σ(-) as the activation function, as the normalized adjacency matrix. Graph Neural Network (GNN) is an efficient deep learning method for processing graph structure data. In the power distribution network, graph structure features can be constructed and represented as a graph G=(V, ε, A), where V is the node set, ε is the edge set, and A is the adjacency matrix. The general form of graph convolution (GCN) operation can be represented as follows:

[0100] .

[0101] GCN assigns the same weight to all neighbor nodes, but this approach cannot adaptively learn the importance of nodes, limiting its application effect in complex power distribution network scenarios.

[0102] Define h i as the input feature vector of node i, W as the shared linear transformation matrix, and a(-) as the attention function. The principle diagram of the multi-head attention mechanism of GAT in the embodiment of the present application is shown in Figure 2 GAT can adaptively learn the importance weights of neighbor nodes by introducing an attention mechanism, better capturing the complex relationships between nodes. The core of GAT is to calculate the attention coefficient eij This process can be represented by the following formula:

[0103] .

[0104] Define N i is the set of neighbor nodes of node i, a ij is the attention weight. In order to make the attention weight comparable between different nodes, the softmax function is used here for normalization processing, and this process can be represented by the following formula:

[0105] .

[0106] Based on the attention weight, the output feature of node i can be calculated by weighted aggregation of neighbor node features, and this process can be represented by the following formula:

[0107] .

[0108] Define || as concatenation operation. In order to enhance the model expression ability and stability, GAT adopts multi-head attention mechanism, and parallelly calculates k independent attention heads, and this process can be represented by the following formula:

[0109] .

[0110] Define is the sum of all values from 1 to k. And for the output layer, average pooling is used instead of concatenation operation, and this process can be represented by the following formula:

[0111] .

[0112] Define , is the active and reactive power of node i at time t, V (t) is the system voltage, type i is the load type identifier, pos i is the node position encoding, time (t) is the time encoding. The feature vector of each node in the distribution network contains electrical state information, load characteristics and space-time information. Here, the feature vector of node i at time t is defined as follows:

[0113] .

[0114] Step 4: Multi-head attention mechanism and physical attention head design. This step designs a multi-head attention mechanism, and parallelly calculates multiple independent attention heads to enhance the model expression ability and training stability, including:

[0115] Step 4.1: Design an electrical distance attention head to calculate spatial correlation weights based on the electrical distance between nodes;

[0116] Step 4.2: Design a load similarity attention head to calculate load similarity coefficients based on load types and power characteristics;

[0117] Step 4.3: Design a control coordination attention head to evaluate the control coordination potential and regulation capability matching degree between nodes;

[0118] Step 4.4: Design a timing prediction attention head to mine the timing correlation between nodes to improve prediction accuracy.

[0119] Further, the outputs of multiple attention heads are spliced or average-pooled to form comprehensive node feature representations.

[0120] Specifically, define Z ij as the electrical distance between nodes, ε Z as a small number to prevent zero values, S ij as a load similarity coefficient, C ij as a control coordination factor, T f as a timing correlation factor. For the physical characteristics of the power distribution network, four attention heads with physical meaning are designed: electrical distance attention, load similarity attention, control coordination attention, and timing prediction attention, each of which focuses on different physical characteristics of the power distribution network to improve the accuracy of load characteristic identification, as shown in the following formula:

[0121] .

[0122] Step 5: GAT network training and ZIP coefficient prediction.

[0123] Step 5.1: Construct a two-layer GAT network structure, the first layer performs preliminary topological feature extraction, and the second layer performs deep semantic feature learning;

[0124] Step 5.2: Design the output layer structure, map the GAT feature vector to the ZIP coefficient prediction value through the fully connected layer;

[0125] Step 5.3: Use the mean square error loss function to optimize the GAT network parameters on the training set;

[0126] Step 5.4: Output the ZIP coefficient prediction results of the GAT network.

[0127] Specifically, define W out , b out as the output layer weight and threshold, The feature vector of the second layer GAT. The GAT network adopts a two-layer structure, and the ZIP coefficient prediction value process of the output layer can be represented by the following formula:

[0128] .

[0129] Step 6: GAT and RLS fusion identification. This step designs an adaptive fusion weight calculation mechanism to dynamically adjust the fusion weight according to the GAT prediction confidence and the RLS identification stability, including:

[0130] Step 6.1: Weighted fusion of the ZIP coefficient output by the GAT network and the identification result of the RLS method;

[0131] Step 6.2: Output the final ZIP coefficient identification result after fusion as the basis for subsequent control decisions.

[0132] Specifically, define Z GAT as the ZIP coefficient output by the GAT network, p RLS as the identification result of the RLS method, and α GAT , α RLS as the fusion weight. The GAT-RLS fusion identification mechanism diagram in the embodiment of the application is shown in Figure 3 , the GAT network output and the RLS method result are adaptively fused to obtain the GAT enhanced load characteristic identification result, and this process can be represented by the following formula:

[0133] .

[0134] Step 6.3: The fused ZIP coefficient directly affects the voltage sensitivity characteristics of the node. According to the physical meaning of the ZIP model, the constant power load is not sensitive to voltage changes, while the constant impedance load is highly sensitive to voltage changes. Therefore, the basic voltage sensitivity of the node is corrected based on the fused ZIP coefficient.

[0135] Specifically, define as the corrected voltage sensitivity of node i, as the basic sensitivity based on the network topology, and are the constant impedance and constant power coefficients of node i, respectively, and k p is the correction coefficient, and this process can be represented by the following formula:

[0136] .

[0137] Step 7: Intelligent control decision based on attention weight.

[0138] Step 7.1: Adaptive control dead-band adjustment based on attention weight, dynamically adjust the trigger threshold of voltage control according to the maximum attention weight;

[0139] Step 7.2: Design control strength calculation formula based on attention intensity and voltage deviation, consider attention weight and voltage trend to determine control strength;

[0140] Step 7.3: Determine whether the current node voltage exceeds the adaptive dead-band range, if it does, trigger control decision.

[0141] The GAT-based collaborative control method takes the attention weight of the GAT network as an important basis for control decision, captures the importance of system state through multi-head attention mechanism, and realizes adaptive control trigger and strength adjustment.

[0142] Define delta b as the basic dead-band, alpha max as the maximum attention weight at the current time. The adaptive control decision flowchart in the embodiment of the present application is shown in Figure 4 The traditional control method uses fixed dead-band for control trigger judgment, and the GAT collaborative control method realizes adaptive dead-band adjustment based on attention weight. This process can be represented by the following formula:

[0143] .

[0144] Define I at , I v as attention intensity and voltage intensity, alpha at , alpha v as intensity weight, as average correction sensitivity. In terms of control strength, the GAT collaborative control determines according to voltage deviation, attention intensity, voltage trend and load characteristics. This process can be represented by the following formula:

[0145] .

[0146] Step 8: Multi-device collaborative control signal distribution.

[0147] Step 8.1: Determine whether to increase or decrease power injection according to the direction of voltage deviation;

[0148] Step 8.2: Calculate the control signal distribution proportion of each device based on device priority, adjustable capacity and attention factor;

[0149] Step 8.3: Distribute specific power regulation instructions to energy storage systems, electric vehicle charging piles, distributed photovoltaic and air conditioning loads;

[0150] Step 8.4: Perform voltage correction calculation, predict voltage adjustment effect based on attention weight.

[0151] Specifically, define C ess , C ev , C ac as the adjustable capacity of energy storage, charging pile, air conditioner; u GAT as the control decision output of GAT network, α avg as the average attention weight. Based on the control decision and attention weight of GAT, the control signal is allocated according to the device priority and attention factor, and when the voltage is low, the allocation of the control signal can be represented by the following formula:

[0152] .

[0153] Voltage correction is based on attention weight and voltage sensitivity corrected by fused ZIP coefficient, define P tot as the total power regulation amount, S b as the basic voltage sensitivity, F a as the attention influence factor, F t as the time factor, F d as the dynamic response factor. Voltage correction is based on attention weight, so its mathematical expression is shown in the following formula:

[0154] .

[0155] Step 9: Adaptive control parameter adjustment and constraint protection.

[0156] Step 9.1: Dynamically adjust the control interval according to the attention weight and system state, shorten the control interval when the attention weight is high, and lengthen the control interval when the attention weight is low. Specifically, define T b as the basic control interval. When the attention weight is high, the control interval is shortened, and vice versa.

[0157] GAT collaborative control can dynamically adjust the control interval according to the attention weight and system state, which can be represented by the following formula:

[0158] .

[0159] Step 9.2: Realize intelligent coordination and cooperation among multiple devices through attention weight, optimize power distribution among devices.

[0160] Define as the actual control output of device k, as the control instruction of device k, η c as the coordination coefficient, η conis the constraint coefficient. Multi-device coordination is achieved through attention weights to realize intelligent cooperation between devices, which can be represented by the following formula:

[0161] .

[0162] Step 9.3: Set the physical constraint conditions such as power limit, ramp rate limit and response delay to ensure that the control signal is within the safe range.

[0163] GAT cooperative control needs to comply with various physical constraints of devices, including power limit, ramp rate limit and response delay constraint. The control signal needs to pass the constraint check before execution, which can be represented by the following formula:

[0164] .

[0165] Step 9.4: Design a regulation inhibition mechanism to prevent control oscillation and inhibit control action when the voltage change trend is opposite to the control direction.

[0166] Define V c is the voltage change amount, is the final voltage change amount, and H(-) is a step function that outputs 1 when the independent variable is greater than 0, otherwise outputs 0. At the same time, to prevent control oscillation, a regulation inhibition mechanism is set in the GAT cooperative control, which can be represented by the following formula:

[0167] .

[0168] Finally, a simulation environment is built on the IEEE 33-node distribution network system. Performance verification is carried out.

[0169] In the above-mentioned power distribution network distributed resource cooperative control method based on graph attention network, all steps can be completed in Matlab2022b.

[0170] The application can realize accurate coordination of distributed resources through GAT enhanced intelligent cooperative control, significantly improve the voltage control effect of the power distribution network, reduce the control frequency, improve the system operation efficiency, and provide an effective technical solution for the operation of the new power system distribution network.

[0171] The implementation basis of each embodiment of the application is that the processing is realized by the programmed processing of the device with processor function. Therefore, in engineering practice, the technical scheme and function of each embodiment of the application are packaged into various modules. Based on this actual situation, on the basis of the above-mentioned embodiments, the embodiment of the application provides a power distribution network distributed resource cooperative control system based on graph attention network, which is used to execute the power distribution network distributed resource cooperative control method based on graph attention network in the above-mentioned method embodiment.

[0172] The system comprises: a first main module for constructing a load characteristic identification system fused with topological information, designing a ZIP load model to describe the voltage sensitivity characteristics of different types of loads, and establishing a distributed resource model; a second main module for load parameter identification based on a graph attention network, the graph attention network designing four types of physically meaningful attention heads of electrical distance attention, load similarity attention, control coordination attention and time series prediction attention, mining the correlation between nodes through a multi-head attention mechanism, and mapping the feature vectors output by the multiple attention heads into ZIP coefficient prediction values; a third main module for adaptively fusing the graph attention network output and the ZIP coefficient identification result based on the recursive least squares method to obtain a graph attention enhanced ZIP coefficient identification result; a fourth main module for adjusting the adaptive control dead zone and control strength based on the attention weights of the graph attention network, and triggering a control decision when the current node voltage exceeds the adaptive dead zone range; and a fifth main module for establishing a control link of adaptive control interval adjustment, device coordination optimization and safety constraint protection based on the graph attention enhanced ZIP coefficient identification result, and realizing intelligent collaborative regulation and control of air conditioning loads, energy storage systems, electric vehicle charging piles and distributed photovoltaics based on the control decision.

[0173] The power distribution network distributed resource collaborative control system based on the graph attention network provided by the embodiment of the application is used for solving the problems of difficulty in identifying distributed resource load characteristics on the power distribution network side of the new power system and difficulty in realizing intelligent collaboration of multiple devices by the traditional control method, adopts the aforementioned modules, realizes accurate identification of load characteristics and intelligent control of voltage quality through fusion identification of the GAT and the recursive least squares method and weight learning of the multi-head attention mechanism, and effectively solves the problems of insufficient utilization of topological information and poor collaborative control effect of the traditional method.

[0174] It should be noted that the system embodiments provided by the application are used to implement the methods in the method embodiments, and are also used to implement the methods in other method embodiments provided by the application, the difference is only that corresponding functional modules are set, the principle is basically the same as that of the aforementioned system embodiments provided by the application, as long as the person skilled in the art improves the modules in the aforementioned system embodiments on the basis of the aforementioned system embodiments, refers to the specific technical solutions in other method embodiments, obtains corresponding technical means through combination of technical features, and the technical solutions composed of these technical means, on the premise of ensuring the practicability of the technical solutions, the corresponding system class embodiments are obtained, which are used to implement the methods in other method class embodiments.

[0175] Based on the same inventive concept as any of the preceding embodiments, the embodiments of the application also provide a non-transitory computer-readable storage medium storing computer instructions, which cause the computer to perform the power distribution network distributed resource collaborative control method based on the graph attention network.

[0176] In summary of the embodiments, the application is applicable to solve the problems of low load characteristic identification accuracy, difficulty in coordinating distributed resources, and slow voltage regulation response in the new power system. First, a load characteristic identification system integrating topological information is constructed, a ZIP load model is designed to describe the voltage sensitivity characteristics of different types of loads, and a distributed resource model is established. Second, a graph attention network load identification method is proposed, which excavates the spatial correlation between nodes through a multi-head attention mechanism, designs four types of physically meaningful attention heads, including electrical distance attention, load similarity attention, control coordination attention, and time series prediction attention, introduces an adaptive fusion strategy to combine the GAT network output and the recursive least squares method results, and fuses the node feature vectors to realize accurate prediction of the ZIP coefficients. Third, an intelligent control strategy based on attention weight is designed, which realizes adaptive trigger judgment through dynamic adjustment of the control dead zone, determines the control strength according to the voltage deviation and attention intensity, establishes a control signal distribution mechanism based on priority and attention factor, and realizes accurate power regulation by combining voltage correction calculation and multi-device coordination optimization. Finally, a GAT-enhanced collaborative control framework is constructed to realize the deep integration of load characteristic identification and voltage control decision-making, establish a complete control link of adaptive control interval adjustment, device coordination optimization, and safety constraint protection, and realize the intelligent collaborative regulation of air conditioning load, energy storage system, electric vehicle charging pile, and distributed photovoltaic through the topological perception ability of graph neural network and the weight learning of multi-head attention mechanism. The application can effectively improve the load identification accuracy and voltage regulation effect of the power distribution network, improve the coordination ability of distributed resources and the system operation efficiency, and ensure the power quality and safe and stable operation of the new power system.

[0177] The key points of the application are:

[0178] 1. The multi-head attention mechanism of the GAT network includes electrical distance attention head to excavate spatial correlation, load similarity attention head to identify load characteristics, control coordination attention head to evaluate regulation ability, and time series prediction attention head to capture time series regularity, realizing deep utilization of the topological structure information of the power distribution network.

[0179] 2. The GAT and RLS fusion identification strategy includes an adaptive fusion weight calculation mechanism and a ZIP coefficient prediction output layer design, enabling the identification method to simultaneously utilize network topological information and time series recursive characteristics, and improving the load characteristic identification accuracy.

[0180] 3. Intelligent control decision is realized through attention weight driven adaptive control dead zone adjustment and control strength calculation, a collaborative control framework based on GAT is formed, topology perception regulation of energy storage system, charging pile, load and photovoltaic is realized, and voltage control effect and system operation efficiency are effectively improved.

[0181] The terms "including", "containing", "having" and their conjugates within the context of the specification and claims of the present application and the above-mentioned figures are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or apparatus that includes a list of steps or units not necessarily limited to the clearly listed steps or units, but can include other steps or units not clearly listed or inherent to these processes, methods, products or apparatuses.

[0182] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the present application.

Claims

1. A distributed resource cooperative control method for power distribution networks based on graph attention networks, characterized in that, include: Construct a load characteristic identification system that integrates topology information, design a ZIP load model to describe the voltage sensitivity characteristics of different types of loads, and establish a distributed resource model; Load parameter identification is based on graph attention network. The graph attention network is designed with four types of physical meaning attention heads: electrical distance attention, load similarity attention, control coordination attention, and time series prediction attention. The correlation between nodes is mined through multi-head attention mechanism, and the feature vectors output by the multi-head attention are mapped to ZIP coefficient prediction values. The graph attention network output is adaptively fused with the ZIP coefficient identification result based on recursive least squares to obtain the graph attention-enhanced ZIP coefficient identification result. The attention weights based on graph attention networks enable the adjustment of adaptive control dead zone and control strength, and trigger control decisions when the voltage of the current node exceeds the adaptive dead zone range. Based on the ZIP coefficient identification results enhanced by graph attention, a control link is established that integrates adaptive control interval adjustment, equipment coordination optimization, and safety constraint protection. Based on control decisions, intelligent coordinated regulation of air conditioning load, energy storage system, electric vehicle charging pile and distributed photovoltaic power generation is realized.

2. The distributed resource collaborative control method for power distribution networks based on graph attention networks according to claim 1, characterized in that, The construction of the graph attention network includes: A graph structure representation of the power distribution network is constructed, with power distribution network nodes as vertices and branch connection relationships as edges, forming an adjacency matrix to describe the network topology; Design node feature vectors to integrate node active power, reactive power, voltage, load type identifier, location code and time code information; Establish an attention coefficient calculation mechanism for graph attention networks, and calculate the correlation between node pairs through the attention function and linear transformation matrix; The attention coefficients are normalized using the softmax function to obtain comparable attention weights; The attention weight is used to perform weighted aggregation of neighbor node features to generate the output feature representation of the node.

3. The distributed resource collaborative control method for power distribution networks based on graph attention networks according to claim 2, characterized in that, Based on the constructed graph attention network, multiple physical attention head designs are introduced, including: Design an electrical distance attention point and calculate spatial correlation weights based on the electrical distance between nodes; The design focuses on load similarity and calculates the load similarity coefficient based on load type and power characteristics. Design a control coordination attention head to evaluate the control coordination potential and regulatory capacity matching degree between nodes; Design a temporal prediction attention head to explore the temporal correlations between nodes; The outputs of multiple attention heads are concatenated or averaged to form a comprehensive node feature representation.

4. The distributed resource collaborative control method for power distribution networks based on graph attention networks according to claim 3, characterized in that, The training of the graph attention network includes: A two-layer graph attention network structure is constructed. The first layer performs preliminary topological feature extraction, and the second layer performs deep semantic feature learning. The output layer structure is designed to map the graph attention network feature vectors to ZIP coefficient prediction values ​​through a fully connected layer. The mean squared error loss function is used to optimize the parameters of the graph attention network on the training set, resulting in a well-trained graph attention network.

5. The distributed resource collaborative control method for power distribution networks based on graph attention networks according to claim 1, characterized in that, The attention weights based on graph attention networks are used to adaptively adjust the control dead zone and control strength, including: The trigger threshold of voltage control is dynamically adjusted based on the maximum attention weight, and the control strength is calculated based on the attention intensity and voltage deviation.

6. The distributed resource cooperative control method for power distribution networks based on graph attention networks according to claim 5, characterized in that, After triggering the control decision, the following multi-device collaborative control signal allocation is performed: Determine whether to increase or decrease power injection based on the direction of voltage deviation; The control signal allocation ratio for each device is calculated based on device priority, adjustable capacity, and attention factor. Specific power regulation instructions for energy storage systems, electric vehicle charging piles, distributed photovoltaic systems, and air conditioning load allocation; Perform voltage correction calculations and predict the voltage adjustment effect based on attention weights.

7. The distributed resource collaborative control method for distribution networks based on graph attention networks according to claim 6, characterized in that, Establish a control link for adaptive control interval adjustment, equipment coordination optimization, and safety constraint protection, including: The control interval is dynamically adjusted based on the attention weight and system state, shortening the control interval when the attention weight is high and lengthening the control interval when the attention weight is low. Intelligent coordination and cooperation among multiple devices are achieved through attention weighting, optimizing power distribution among devices; Set physical constraints, including power limits, ramp rate limits, and response delays, to ensure that control signals are within safe limits; Design a regulation and suppression mechanism to suppress control action when the voltage change trend is opposite to the control direction.

8. The distributed resource collaborative control method for distribution networks based on graph attention networks according to claim 1, characterized in that, Establishing a distributed resource model also includes: Establish a distributed photovoltaic power output model that comprehensively considers the impact of solar radiation intensity, cloud layer influence factors, and weather change factors on photovoltaic power generation. Establish a dynamic model of the state of charge (SOC) of an energy storage system to describe the variation of SOC and the time-varying characteristics of charging and discharging efficiency during the charging and discharging process. Establish a power model for electric vehicle charging stations, taking into account the impact of the number of charging vehicles, the AC / DC charging ratio, and the power factor on the charging power.

9. A distributed resource collaborative control system for power distribution networks based on graph attention networks, used to implement the method described in any one of claims 1 to 8, characterized in that, include: The first main module is used to build a load characteristic identification system that integrates topology information, design a ZIP load model to describe the voltage sensitivity characteristics of different types of loads, and establish a distributed resource model. The second main module is used to identify load parameters based on a graph attention network. The graph attention network is designed with four types of physical attention heads: electrical distance attention, load similarity attention, control coordination attention, and time series prediction attention. The multi-head attention mechanism is used to mine the correlation between nodes and map the feature vectors output by the multi-head attention to the predicted ZIP coefficient values. The third main module is used to adaptively fuse the output of the graph attention network with the ZIP coefficient identification result based on the recursive least squares method to obtain the graph attention-enhanced ZIP coefficient identification result. The fourth main module is used to adjust the adaptive control dead zone and control strength based on the attention weights of the graph attention network, and to trigger control decisions when the voltage of the current node exceeds the adaptive dead zone range. The fifth main module is used to establish a control link based on the ZIP coefficient identification results of graph attention enhancement, including adaptive control interval adjustment, equipment coordination optimization, and safety constraint protection. Based on the control decisions, it enables intelligent coordinated regulation of air conditioning load, energy storage system, electric vehicle charging pile and distributed photovoltaic.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the distributed resource collaborative control method for power distribution networks based on graph attention networks as described in any one of claims 1 to 8.

Images