Power system state estimation method considering unmanned aerial vehicle scheduling

Abstract

The invention aims to solve the problem of insufficient robustness of a traditional power grid state estimation method in post-disaster recovery and emergency communication scenes, and provides a power system emergency state estimation method based on an unmanned aerial vehicle collaborative graph neural network. According to the method, unmanned aerial vehicle dynamic communication scheduling and a space-time diagram neural network are fused, and the problems of data missing, topology time varying, communication interruption and the like are solved. A node voltage estimation and confidence thermodynamic diagram is generated by constructing a simulation data set, designing a space-time diagram convolutional network and combining a diagram feature reconstruction module, and a priority recovery strategy is formulated. Meanwhile, an unmanned aerial vehicle emergency communication optimization model is established, and efficient distribution of communication resources is realized. The method has the advantages that an information-physical system collaborative optimization method is provided, a rapid and reliable communication alternative scheme is provided, and the problems of measurement missing and topology time varying are solved. By establishing a closed-loop feedback model, dynamic optimization of communication resource allocation scheduling is realized, and the method adapts to a real-time dynamic mode of a power system.

Description

Technical Field

[0001] The present invention relates to the field of power system state estimation, and specifically to a power system state estimation method considering UAV scheduling, aiming to improve the robustness of power system state estimation in communication-constrained scenarios. Background Art

[0002] Modern power systems have gradually evolved into information-physical coupled systems. The information network needs to sense information such as power grid topology and node power in real time, realize the circulation of node information, and achieve the estimation of operating states through state estimation. Extreme events can damage both the information network and the physical network of the power system simultaneously, resulting in inaccurate state estimation. UAVs equipped with wireless communication modules can act as temporary communication base stations through the dual functions of dynamic networking and real-time data collection. Therefore, in the problem of power system state estimation based on information-physical coupling, the impact of UAV-based information transmission on the physical system should be considered.

[0003] The application of traditional state estimation methods has limitations in extreme events. On the one hand, they rely on the data information of fixed measurement devices, and when some node data is missing due to extreme disasters or equipment failures, the estimation accuracy decreases. On the other hand, existing research on information-physical coupled systems usually assumes that the topology is stable and reliable, without considering the topology changes caused by faults, and it is difficult to cope with the state estimation errors caused by topology changes. Therefore, the existing general recovery strategies for communication networks are not directly applicable, and it is necessary to study the power system state estimation technology for UAV emergency communication in emergency situations to improve the robustness of state estimation algorithms under measurement loss and time-varying topologies. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies mentioned in the above background art, and provide a state estimation model that can be based on UAV communication collaboration in emergency situations, which helps to perform state estimation in the case of data loss.

[0005] The technical solution adopted by the present invention is: A power grid state estimation technology based on graph neural network under UAV emergency communication, including a graph neural network model for state estimation and a UAV emergency communication optimization model, specifically including the following steps:

[0006] First, collect measurement data and electrical parameters and process them into a data set. The specific process is as follows: construct a set of feasible lines based on a standard power system example, construct a set of time-varying topologies based on the set of feasible lines, calculate the power flow results on the set of time-varying topologies, add Gaussian white noise to the node injection power, and embed random missing values based on the Gaussian distribution; finally, use the node injection power, voltage values, branch admittances, branch currents, and load powers with added noise and missing values as the data set.

[0007] Construct a feature reconstruction model under spatio-temporal scales, including a time feature extraction module and a spatial feature module. On the time scale, by introducing an exponential time decay function and a memory gating mechanism to extract time-dependent relationships, dynamically adjust the weight coefficients of historical measurement data, and achieve the extraction of dynamic features of the power system; on the spatial scale, construct a spatial feature extraction model based on a graph convolutional neural network, define the spatial distance-dependent relationship with branch admittance values, and use spatially decaying weights to weight the features of its adjacent nodes to achieve adaptive feature fusion of spatially adjacent nodes; perform linear regression on the time features and spatial features to obtain the feature reconstruction value.

[0008] Construct the backbone of a state estimation model based on a spatio-temporal graph convolutional neural network. On the time dimension, construct a gated recurrent unit to achieve the learning of dynamic laws such as voltage fluctuations, load changes, and topological adjustments by retaining a certain amount of memory for the processed measurement data; on the spatial dimension, improve the spatial graph convolutional kernel based on branch admittance values. Each node exchanges features with its connected neighbor nodes, and the feature information of different nodes will decay to different degrees according to their electrical distances. After fusing all the information, perform a non-linear transformation to achieve the extraction of spatial features; after extracting spatio-temporal features, embed the power flow equation and the Kalman filter loss into the graph neural network loss function part to constrain the predicted value, and add a confidence coupling mechanism to ensure that the model can synchronously output the state estimation value and the corresponding confidence.

[0009] Construct a UAV scheduling model. First, use an interpretable machine learning model to analyze the parameters of the graph neural network and calculate the node recovery priority accordingly. Subsequently, construct a mixed integer programming model, where the objective function is to minimize the weighted sum of the UAV scheduling distance, the node recovery priority, and the communication noise; the constraints include UAV working height constraint, UAV energy consumption constraint, UAV coverage range constraint, UAV inter-distance constraint, turning distance constraint when UAVs avoid obstacles, and communication quality threshold constraint.

[0010] In the state estimation model based on a graph neural network, construct an online learning mechanism. When the confidence of the estimation result is lower than the predefined threshold, enable the online learning mechanism, use the Kalman filter to perform state estimation on the measurement data during this period, and construct a small dataset based on the state estimation result to fine-tune the state estimation result. After the model is fine-tuned, the recovery priority will be recalculated.

[0011] The beneficial effects of the present invention are as follows: A method for state estimation of an information - physical coupled power system based on unmanned aerial vehicles (UAVs) is proposed. By integrating information network transmission and physical power grid state estimation, robust state estimation in extreme disaster scenarios is achieved. This method considers the mutual dependence of the information - physical network and enhances the robustness of the state estimation system through the optimal scheduling of the information network. The UAV emergency mobile base station is an aerial wireless base station that can move autonomously in the affected area and form communication links, providing a fast and reliable alternative for wireless communication. The present invention considers the measurement missing and topology time - varying problems existing in the scheduling process in case of emergency, and dynamically reconstructs features based on graph feature propagation. When local data is missing, reliable estimated values can still be generated through the graph structure under the constraint of the power flow equation. Secondly, by adopting a graph convolutional neural network in the spatio - temporal scale to aggregate spatio - temporal features, the influence caused by topology time - variation is suppressed. Finally, the graph neural network is dynamically adjusted through an online learning mechanism to adapt to the dynamic mode of the power system. The present invention realizes the dynamic optimization of communication resource allocation scheduling and adapts to the real - time dynamic mode of the power system by establishing a closed - loop feedback model of node recovery priority - online learning mechanism. Brief Description of the Drawings

[0012] Figure 1 It is a flow chart of a power system state estimation method considering UAV scheduling described in this application. Detailed Embodiment

[0013] The following further elaborates on the present invention in detail with reference to the drawings and specific examples, which is convenient for clearly understanding the present invention, but they do not constitute a limitation to the present invention.

[0014] The present invention provides a power system state estimation method for collaborative optimization of an information - physical system based on UAVs, including the following steps:

[0015] S1: Obtain the power grid topology structure, electrical parameters, fault location, system parameter communication fault area, and determine the UAV base station position coordinates; establish a topology reconstruction model based on the IEEE standard example, perform topology reconstruction on the example as a time - varying topology set, sample the load based on the time - varying topology set and calculate the power flow, and add Gaussian white noise and random missing values to the power flow calculation results to obtain a training dataset.

[0016] S2: Build a graph feature reconstruction model in the spatio - temporal scale. Based on the spatio - temporal feature dependence relationship, reconstruct the missing values and outliers. Aggregate its time - neighborhood information and space - neighborhood information as spatio - temporal expressions to replace the outliers and missing values, and embed the graph feature reconstruction layer into the graph neural network.

[0017] S3: Construct the backbone of the state estimation model based on the graph neural network. Improve the graph convolutional neural network through electrical distance, consider the influence of admittance values during the propagation of spatial information, and extract temporal features based on the gated recurrent unit to construct the backbone of the spatio-temporal graph convolutional network. Establish a physical information regularization term based on the power flow equation and the Kalman filter loss, embed it into the graph neural network loss, and train the graph neural network based on the training dataset. After training, use an interpretable machine learning model to analyze the graph neural network and formulate the node recovery priority.

[0018] S4: Establish a UAV scheduling model. The objective function is to minimize the weighted sum of the UAV scheduling distance, node recovery priority, and communication noise. The constraints are the UAV working altitude and obstacle avoidance altitude, coverage range constraint, endurance constraint, turning distance constraint during obstacle avoidance, and communication quality threshold constraint.

[0019] S5: Construct an online learning mechanism. When the confidence level of the prediction results within a certain period is lower than the preset threshold, enable the online learning mechanism. First, use the Kalman filter to calculate the exact solution. Subsequently, based on the Kalman filter calculation results and measurement data, construct a small dataset, use this dataset to fine-tune the model, and recalculate the node recovery priority.

[0020] In S1, define the UAV set in the UAV planning problem as M ∈ {0, 1, 2....j}, the power system node set as N ∈ {0, 1, 2...i}, and the coordinates corresponding to the power system nodes as (x i , y i , z i ). At the same time, there are circular obstacles within the coverage area of the entire power system. Define A ∈ {0, 1, 2....k} as the obstacle area set, and the corresponding radius of the obstacle area as r k .

[0021] In S1, define the line set as L. The standard test case performs a single-line disconnection operation on all lines, and uses the Matpower tool to verify the power flow convergence, voltage stability, and power balance constraints after disconnection one by one, and screen out the feasible disconnection set that meets the N-1 criterion. Subsequently, generate a time-varying topology sequence based on this set, select multiple combined lines from the feasible set to disconnect, update the topology state after each disconnection operation, and record the line on-off state, node electrical measurements, and power flow data. Finally, perform an AC power flow calculation for each topology state, dynamically adjust the generator output and load distribution to maintain power balance, and at the same time verify whether the line load rate and node voltage meet the upper and lower limit constraints. Finally, output a time-varying dataset containing topology dynamic evolution, measurement data, and physical constraints. The process is as follows:

[0022]

[0023] ∑Pg = ∑(P load - P cut ); (3)

[0024] Equation (1) constructs the set of feasible lines L F , and filters the line set by quantifying the feasibility conditions for single-line disconnection. In the equation, V i max and V i min represent the upper and lower limits of the node voltage respectively; Equation (2) defines the recurrence logic for topological dynamic changes, constraining that each line disconnection operation only selects lines from the pre-screened feasible set and inherits the topological state of the previous time step. In the equation, F t represents the variable topology constructed at time t; Equation (3) satisfies power conservation under any topology by integrating generator output and load shedding. P g is the generator output, P load is the load, and P cut is the load shedding value.

[0025] After obtaining the time-varying topology set F, the load is randomly sampled, and Matpower is used to calculate the power flow results. Based on the power flow calculation results, Gaussian white noise is added to simulate measurement errors, and random missing values are added to simulate measurement omissions. The expressions are as follows:

[0026]

[0027] Equations (4)-(5) are the processes of adding random noise and missing values, where M i,j ∈ {0, 1} m×n is the mask matrix, used to represent the positions of random missing places. X is the original value of the electrical parameter, Y is a random variable subject to a Gaussian distribution, and p is the data missing probability subject to a Gaussian distribution. When the mask matrix is 1, the electrical parameter is added with noise, otherwise it is set to NaN. P N , Q N represent the active power injection and reactive power injection of the node after adding noise and random missing respectively. The finally constructed data features are D = {P N , Q N , a, L a , G, B}, where a is the branch current, L a is the load power, and G and B are the real and imaginary parts of the admittance matrix respectively.

[0028] In S2, the graph neural network reads the node feature map with missing features and uses spatio-temporal feature aggregation to reconstruct the feature map to handle missing values. Define the power system topology graph as G ∈ (N, E), where N is the set of power system topology nodes and E is the set of power system topology edges. Define the node information feature map as H. The feature reconstruction expression under the spatio-temporal scale is:

[0029]

[0030] Equation (6) is the expression of spatial local features and temporal local features. Spatio-temporal information is obtained by aggregating the features of adjacent nodes. In the equation, and represent the spatial mean and temporal mean respectively, and represent the spatial decay weight and temporal decay weight respectively. Equations (7) and (8) are the spatial decay model and temporal decay model respectively. In the spatial decay model, the electrical distance between nodes is defined based on admittance. In the temporal decay model, the temporal node distance is defined through topological similarity and timestamp. A feature decay model is constructed through spatio-temporal distance. In the equation, E ij is the reciprocal of admittance, sim is the topological similarity, and w and b are learnable weights and biases. Equation (9) is the definition of topological similarity. In the equation, A i and A j are two adjacency matrices representing topologies, and Vec represents the matrix vector operation. Equation (10) is the expression of the final spatio-temporal feature embedding value. Linear regression is performed through a fully connected layer to fit the weighted value of the time-space features.

[0031] In S3, an improved graph convolutional neural network is proposed as the backbone of the state estimation model. The spatial convolutional kernel is improved using electrical distance. By aggregating spatio-temporal scale information, the spatial topology information in the power system topology is extracted, and historical information is aggregated through the GRU gating mechanism to suppress the impact caused by topological time variation. The structure expression of the improved spatio-temporal scale graph convolutional neural network proposed by the present invention is:

[0032] H A =σ(D -1 / 2 GD -1 / 2 XW + b); (11)

[0033]

[0034] Equation (11) is the spatial convolution expression. For the node features at each time step, the information of its spatial neighbors is weighted and aggregated to extract local spatial patterns. In the equation, H AFor the extracted spatial local features, D is the degree matrix, G is the electrical distance represented by the admittance matrix, X is the graph feature, W and b represent the learnable weights and biases respectively, and σ is the activation function. Equation (12) is the forgetting gate expression, which determines how much historical information to retain or forget based on the current spatial feature and the historical state. In the equation represents the output feature of the forgetting gate, and concat represents the graph feature concatenation operation. Equation (13) is the update gate expression, which dynamically adjusts the fusion weight of the current input feature and the historical hidden state. In the equation is the output feature of the update gate. Equation (14) is the candidate hidden state, which generates a new candidate state using the historical information filtered by the reset gate and the current spatial feature. In the equation represents the temporary update value at the current time step. Equation (15) is the spatio-temporal joint output expression, which summarizes the spatio-temporal features of all nodes and serves as the input for downstream tasks. In the equation represents the spatio-temporal joint output feature.

[0035] The definition of the node recovery priority in S3 depends on the analysis of the trained graph neural network parameters to calculate the importance of nodes in the power system topology. The expression for calculating the importance weight of power system topology nodes based on neural network parameter analysis is:

[0036]

[0037] In the present invention, an interpretable machine learning model is used to decompose the output of the graph neural network model into the contributions of each feature, and the recovery priority is calculated based on the marginal contribution value of the nodes. In equation (16), the marginal contribution value of the graph neural network for calculating nodes is W S , where F represents the set of all features, and f(S) represents the predicted value estimated using the feature set. In equation (17), the marginal contribution value is normalized to obtain In equation (18), the reciprocal of the marginal contribution value is taken to obtain the node recovery priority W i .

[0038] In S3, a graph neural network architecture integrating a physical information neural network is adopted. Based on the graph neural network, a physical information regularization term weighted by physical equations is proposed, and a hybrid physical model is integrated into the loss function, and the confidence output is coupled, mainly including the mean square error term, the physical power flow equation constraint term, the penalty term of the traditional numerical method, and the confidence coupling term. The expression of the graph neural network loss function in the present invention is:

[0039] Loss = β1Loss d + β2Loss p + β3Loss v + β4Loss c ; (19)

[0040]

[0041]

[0042] Q ij = -V i 2 B ij -V i V j (G ij sinθ ij -B ij cosθ ij ); (24)

[0043]

[0044] Equation (19) is the total expression of the loss function, and equations (20), (21), (25), and (26) represent the mean square error loss, the physical information term loss, the numerical calculation loss, and the confidence coupling term, respectively.

[0045] Equation (20) measures the basic loss between the graph neural network and the training data based on the mean square error, where V represents the predicted value and the true value of the voltage amplitude, respectively, θ represents the predicted value and the true value of the voltage phase angle, respectively, and α1, α2 are the corresponding weight coefficients.

[0046] Equation (21) then uses the power flow equation to impose certain constraints on the predicted values of the neural network, ensuring that the output values of the graph neural network can satisfy the power flow equation as much as possible. (22)-(25) are the node active power balance equation, the node reactive power balance equation, the line active power injection balance, and the line reactive power injection balance, respectively. Among them, represents the node active power calculated based on the voltage predicted value and the power flow equation, and P represents the true value of the node active power, represents the node reactive power calculated based on the voltage predicted value and the power flow equation, and Q represents the true value of the node active power. represents the line active power calculated based on the voltage predicted value and the power flow equation, and P' represents the true value of the line active power, represents the line reactive power calculated based on the voltage predicted value and the power flow equation, and Q' represents the true value of the line active power.

[0047] As the numerical loss term, equation (25) measures the error between the output of the graph neural network and the calculated value of the Kalman filter, where and They respectively represent the numerical solutions of the voltage amplitude and voltage phase angle calculated based on the Kalman filter algorithm under the same input. By adding a numerical loss term, the neural network will make the output as close as possible to the output value of the Kalman filter during the training process.

[0048] As the confidence coupling term, Equation (26) defines the inverse relationship between the confidence output and the loss of the predicted value. When the confidence of the voltage amplitude or voltage phase angle is relatively large, its corresponding confidence will decrease.

[0049] In S4, the decision variable of the UAV scheduling model is the coordinate of the UAV working position, and the constraint conditions of the proposed UAV scheduling model are:

[0050]

[0051]

[0052] Equation (27) is the altitude constraint of the UAV. Since the UAV must work at a certain altitude, the altitude constraint ensures that the UAV can work properly while restricting the minimum altitude of the UAV not to be lower than the obstacle altitude to prevent collisions. Where h j represents the altitude decision variable of the final coordinate of the j-th UAV, represents the upper limit altitude of the j-th UAV, represents the obstacle avoidance altitude of the j-th UAV, that is, the minimum altitude.

[0053] Equations (28), (29), (30), and (31) are the endurance constraints of the UAV, which need to ensure that the endurance of the UAV when it departs can reach the target node and contains enough remaining power to support hovering work. represents the hovering power of the j-th UAV, that is, the power required for the UAV to counteract gravity. represents the forward power of the j-th UAV, d i,j and c i,j respectively represent the straight-line distance and the curved distance when the j-th UAV reaches the i-th node of the final destination. v j represents the speed of the j-th UAV when it travels at a constant speed. Therefore, the time can be estimated based on the distance and speed, and then multiplied by the total power to estimate the required endurance energy consumption. represents the maximum endurance energy consumption of the j-th UAV, and it is necessary to ensure that the UAV can reach the destination within this energy consumption. Equations (33) and (34) are the expressions of the hovering power and the forward power of the UAV, where g represents the acceleration due to gravity, m represents the mass of the UAV, S UAV represents the surface area of the UAV rotor, ρ is the air density, and λ is the air resistance coefficient.

[0054] Equation (32) is the total distance when the UAV circles around the fault area, and α j represents the maximum turning angle of the j-th UAV. The last term in the constraint expression represents the curve distance when the UAV circles around the circular area at the maximum turning angle.

[0055] Equations (35) and (36) are the communication coverage constraints of the UAVs, which ensure that at least one power system node is within the sensing range of the final working position of each UAV, ensuring that the UAVs can normally sense the node data information nearby. Secondly, it restricts the maximum number of nodes within the coverage radius of the UAVs to avoid problems such as decreased communication quality and communication failure caused by too many communication nodes. In the formula, R j represents the maximum information sensing radius of the j-th UAV. In Equation (37), D j,i represents the Euclidean distance between the j-th UAV and the i-th power system node, and (x i , y i , h i ) and (x j , y j , h j ) represent the three-dimensional coordinates of the UAV and the power grid node respectively.

[0056] In Equation (38), it is stipulated that the communication coverage of the UAV is related to the hovering height. In the formula, θ UAV is the elevation angle of the UAV with respect to the horizontal plane when hovering. In the present invention, the elevation angle is stipulated to be a fixed value to ensure that the hovering height of the UAV is within a certain range.

[0057] Equation (39) is the mutual exclusion constraint of the UAVs. In the formula, represents the two-dimensional Euclidean distance between any two UAVs, and D min represents the shortest distance between any two UAVs, ensuring that multiple UAVs will not work at the same node.

[0058] Equations (40), (41), (42), and (43) are the UAV communication loss model and communication quality constraints. Equation (40) is the expression of the effective signal power of the UAV. In the formula, represents the effective signal power of the j-th UAV, represents the signal transmission power of the UAV, represents the signal loss power of the UAV, represents the noise power of the UAV. Equation (41) gives the expression of the UAV logarithmic distance path loss model. γ represents the wavelength of the transmitted signal, and P LF (h) represents the path loss at the reference distance, and X σ represents a random variable used to describe the normal shadow effect. Equation (42) is the communication quality constraint used to restrict the minimum communication quality of the model. In the formula, Represents the minimum communication quality required for the \(i\)-th power system node (using signal-to-noise ratio as the representation index), ensuring that when the UAV communicates, the communication quality is within the available range. Equation (43) is the expression of the signal-to-noise ratio of the UAV.

[0059] In step 4, the objective function expression of the UAV scheduling model is:

[0060]

[0061] Equation (44) is the objective function of the optimization model. The first two objectives are to minimize the sum of the movement distances of all UAVs to ensure the optimization of the overall energy consumption of the UAVs, enabling the UAVs to quickly reach the destination and start working. The third term is to minimize the sum of the power system node recovery priorities, where \(D\) j,i represents the distance from a certain UAV to a certain power system node, and \(W\) i represents the weight of the corresponding power system node. When the importance of the node is greater, its corresponding weight is lower. The fourth term is to minimize the communication noise of the UAVs, thereby maximizing the overall communication quality as much as possible. In the formula, represents the signal-to-noise ratio in the channel during UAV communication.

[0062] In S5, an intelligent online learning mechanism is constructed by presetting a confidence threshold. When the average confidence of the state estimation output within a certain period of time is lower than the preset threshold, the online learning mechanism is activated. The Kalman filter is used to calculate the state estimation value for the measured values during this period. A small dataset is constructed based on the measured values and the Kalman filter estimation results to fine-tune the model parameters to ensure the model's learning of the dynamic changes of the power system.

[0063]

[0064] \(D\) on =\(\{z\) τ , \(V\) KF,τ , \(θ\) KF,τ \}; (47)

[0065]

[0066] Equation (45) is the activation condition of the online learning mechanism. When the average confidence of the state estimation model is low within a certain period of time, the online learning mechanism is activated to adjust the model, where respectively represent the average confidence of the voltage amplitude and voltage phase angle, \(c\) V,th , \(c\) θ,th respectively represent the preset confidence thresholds of the voltage amplitude and voltage phase angle, \(T\) represents the moment when the online learning is activated, and \(T_0\) represents the time window. Equation (46) is the calculation process of the Kalman filter, \(V\) KF and \(θ\) KFrespectively represent the voltage amplitude and phase angle estimated using the Kalman filter, z τ , represents the measurement data within the time window, ∑(τ - ) represents the covariance matrix. Equation (47) is the process of constructing the online learning dataset, and an online learning dataset D is constructed based on the measurement values and the Kalman filter results on . Equation (48) is the model fine-tuning process, and the state estimation model is retrained using the online learning dataset, and θ th are the parameters after model fine-tuning.

Claims

1. A power system state estimation model considering UAV scheduling, characterized in that: A state estimation model and a UAV emergency communication scheduling model based on graph neural network are constructed. In the graph neural network model, first, measurement data in the scenarios of time-varying topology and measurement missing are collected as an offline training data set; a spatio-temporal graph convolutional neural network improved based on electrical distance is constructed and repeatedly stacked as the backbone network of the state estimation model; abnormal features are reconstructed based on spatio-temporal feature dependencies, and the output of the graph neural network is the state variables and result confidence. After training, an interpretable machine learning model is used to explain the node importance; In the UAV scheduling model, the restoration priority is formulated based on the node importance of the graph neural network, and further a UAV scheduling model is constructed to calculate the optimal UAV working positions. After collecting measurement data, it is used as the input of the state estimation model; when the confidence of the state estimation result is lower than the preset threshold within a certain period of time, an online learning mechanism is enabled to fine-tune the model and recalculate the node restoration priority to form a closed-loop iteration. In the data set construction stage of the model, a set of feasible lines is constructed to simulate the time-varying power grid topology scenario, the power flow results are calculated using the time-varying topology sequence, and Gaussian white noise and random missing are added to the power flow calculation results based on the Gaussian distribution to simulate the measurement noise and partial data missing scenarios, and finally a data set for training the graph neural network model is constructed. In the state estimation model, in the graph neural network, the admittance matrix of the power grid, node fault information, active power, reactive power, load power, branch admittance and branch current of each bus node are used as the input of the feature graph, and the node voltage amplitude and phase angle, as well as the confidence, are used as the output; After training, the importance of the graph neural network nodes is analyzed based on the interpretable machine learning model as the node importance criterion, and the restoration priority of the power system nodes is formulated. The graph neural network includes a graph feature reconstruction module, which performs graph feature reconstruction operations on the missing features based on the time-space scale features, embeds the reconstructed values into the missing measurements, and can perform power system state estimation in the case of missing some feature information; the backbone of the graph neural network is an improved spatio-temporal graph convolutional neural network, which aggregates the features at the spatio-temporal scale to improve the model prediction accuracy, and embeds the power flow equation, Kalman filter loss and confidence mechanism into the graph loss function to constrain the model output. The objective function of the UAV emergency communication optimization model is to minimize the weighted sum of the flight distance of the UAV, node restoration priority and communication noise; the constraints of the UAV emergency communication optimization model include the working height of the UAV, coverage range, turning distance of the UAV, communication quality threshold constraint and UAV endurance constraint; In the state estimation model based on graph neural network, an online learning mechanism is constructed. When the confidence of the estimation result is lower than the predefined threshold, the online learning mechanism is enabled, the measurement data during this period is used for state estimation using Kalman filter, and a small data set is constructed based on the state estimation result to fine-tune the state estimation result. After the model is fine-tuned, the restoration priority will be recalculated.

2. The graph neural network state estimation method according to claim 1, characterized in that: The expressions for constructing the time-varying topology set and the measurement data in the missing value and noise scenarios are: ∑P g = ∑(P a - P cut ) ； Among them, L F is the set of feasible lines, V i max and V i min represent the upper and lower limits of the node voltage respectively, F t represents the variable topology constructed at time t, P g is the generator output, P a is the load power, P cut is the load shedding value, F is the set of time-varying topologies constructed, M i,j ∈{0, 1} m×n is the mask matrix, used to represent the positions of random missing places, P and Q are the active and reactive injection powers at the node ends respectively, Y1 and Y2 are random variables subject to Gaussian distribution, p is the data missing probability subject to Gaussian distribution, NaN represents the missing value, P N , Q N represent the active and reactive injection powers at the nodes after adding noise and random missing respectively, and the finally formed data features are D = {P N , Q N , a, L a , G, B}, where a is the branch current, L a is the load power, and G and B are the real and imaginary parts of the admittance matrix respectively.

3. The graph neural network state estimation method according to claim 1, characterized in that: The expression for reconstructing abnormal features based on the spatio-temporal feature dependency relationship based on spatio-temporal features is as follows: Among them and represent the spatial mean and the temporal mean respectively, and represent the spatial decay weight and the temporal decay weight respectively, E ij is the reciprocal of admittance, sim is the topological similarity, w and b are learnable weights and biases, A i and A j are two adjacency matrices representing the topology, Vec represents the matrix-vector operation.

4. The graph neural network state estimation method according to claim 1, characterized in that the expression of the spatio-temporal graph convolutional neural network improved based on electrical distance is: H A = σ(D -1 / 2 GD -1 / 2 XW + b)); Among them, H A is the spatially local feature after extraction, D is the degree matrix, G is the electrical distance represented by the admittance matrix, X is the graph feature, W and b represent the learnable weights and biases respectively, σ is the activation function, r i out represents the output feature of the forgetting gate, concat represents the graph feature concatenation operation, is the output feature of the update gate, represents the temporary update value at the current time step, represents the spatio-temporal joint output feature.

5. The method for constructing a graph neural network according to claim 4, characterized in that: The expression of the loss function in the graph neural network is: Loss = β1Loss d + β2Loss p + β3Loss v + β4Loss c ; Q ij = -V i 2 B ij -V i V j (G ij sinθ ij -B ij cosθ ij ); Among them, Loss d 、Loss p 、Loss v 、Loss c represent mean square error, physical information loss, numerical loss, and confidence loss respectively, V represents the predicted value and the true value of the voltage amplitude respectively, θ represents the predicted value and the true value of the voltage phase angle respectively, represents the active power of the node calculated based on the predicted voltage value and the power flow equation, and P represents the true value of the active power of the node, represents the reactive power of the node calculated based on the predicted voltage value and the power flow equation, and Q represents the true value of the active power of the node, represents the active power of the line calculated based on the predicted voltage value and the power flow equation, and P′ represents the true value of the active power of the line, represents based on the predicted voltage value and the power flow equation, Q′ represents the true value of the active power of the line, where and represent the numerical solutions of the voltage amplitude and the voltage phase angle calculated based on the Kalman filter algorithm under the same input respectively, c V,i , c θ,i represent the confidence of the voltage amplitude and the confidence of the voltage phase angle in the model output respectively.

6. The method for constructing a drone scheduling model according to claim 1, characterized in that the expression of the drone scheduling model is: Among them, Let \(M\in\{0,1,2,\ldots,j\}\) be the set of UAVs, \(N\in\{0,1,2,\ldots,i\}\) be the set of power system nodes, and the coordinates of the power system nodes be \((x i ,y i ). Let \(A\in\{0,1,2,\ldots,k\}\) be the set of obstacle regions, and the radius of the obstacle region be \(r k represents the upper limit height of the \(j\)-th UAV, represents the obstacle avoidance height of the \(j\)-th UAV, represents the hovering power of the \(j\)-th UAV, that is, the power required for the UAV to counteract gravity, represents the forward power of the \(j\)-th UAV, \(d i,j and \(c i,j respectively represent the straight-line distance and the curved distance when the \(j\)-th UAV reaches the \(i\)-th node of the final destination, \(v j represents the speed of the \(j\)-th UAV when moving at a constant speed, represents the maximum endurance energy consumption of the \(j\)-th UAV. It is necessary to ensure that the UAV can reach the destination and work within this energy consumption. \(g\) represents the acceleration due to gravity, \(m\) represents the mass of the UAV, \(S UAV represents the surface area of the UAV rotor, \(\rho\) is the air density, \(\lambda\) is the air resistance coefficient, \(\alpha j represents the maximum steering angle of the \(j\)-th UAV, \(\theta\) is the elevation angle of the UAV when hovering with respect to the horizontal plane, \(R j represents the maximum information perception radius of the \(j\)-th UAV, \(D j,i represents the Euclidean distance between the \(j\)-th UAV and the \(i\)-th power system node, \((x i ,y i ,h i ) and \((x j ,y j ,h j ) respectively represent the three-dimensional coordinates of the UAV and the power grid node, \(D j1,j2 represents the two-dimensional Euclidean distance between any two UAVs, \(D min represents the shortest distance between any two UAVs, represents the effective signal power of the \(j\)-th UAV, represents the signal transmission power of the UAV, represents the signal loss power of the UAV, represents the noise power of the UAV, \(\gamma\) represents the wavelength of the transmitted signal, \(P LF (h)\) represents the path loss at the reference distance, \(X σ represents a random variable used to describe the normal shadowing effect, represents the minimum communication quality required for the \(i\)-th power system node.​ 7. The method for constructing a drone scheduling model according to claim 6, characterized in that in the drone scheduling model, the expression of the restoration priority of different nodes is: Where F represents the set of all features, S represents the feature set, f(S) represents the predicted value estimated using the feature subset, and W S represents the marginal contribution value calculated based on the interpretable machine learning model, represents the normalized marginal contribution value, and W i represents the recovery priority of the i-th node.

8. The graph neural network state estimation method according to claim 1, characterized in that: The expression of the online learning mechanism of the state estimation model is: D on = {z τ , V KF,τ , θ KF,τ}; where represent the confidence levels of the voltage amplitude and voltage phase angle, respectively, c V,th , c θ,th represent the preset confidence level thresholds of the voltage amplitude and voltage phase angle, respectively, T represents the moment to start online learning, T0 represents the time window, V KF and θ KF represent the voltage amplitude and phase angle estimated using the Kalman filter, respectively, z τ, represents the measurement data within the time window, ∑(τ - ) represents the covariance matrix, D on is the data set for online learning, θ th are the parameters after model fine-tuning.

Images