Energy consumption network joint estimation method and system based on physical mask autoencoder

Abstract

The application discloses a power consumption network joint estimation method and system based on a physical mask autoencoder, relates to the field of power systems and their automation technology, and comprises the following steps: acquiring the topological structure and measurement data of a regional power consumption network; mapping the topological structure and measurement data of the regional power consumption network to obtain a undirected graph model, performing double random mask operations on the undirected graph model to obtain mask graph data, wherein the mask graph data comprises a mask node set and a mask edge set; and inputting the mask graph data into a pre-trained physical information graph mask autoencoder model to output regional power consumption network topological structure and operation state estimation results satisfying physical constraints.

Description

Technical Field

[0001] This invention relates to the field of power systems and their automation technology, specifically to a joint estimation method and system for energy-consuming networks based on physical mask autoencoders. Background Technology

[0002] With the large-scale integration of distributed energy resources, especially photovoltaics, traditional regional energy networks are rapidly transforming into proactive regional energy networks, exhibiting significant randomness and volatility in system operation. This transformation makes the power flow of regional energy networks complex and variable, requiring frequent reconfiguration of the network topology to adapt to operational needs. However, limited by investment costs, the coverage of high-precision measurement equipment in regional energy networks remains low, and dispatch centers often only obtain fragmented measurement data. This coexistence of topological uncertainty and incomplete measurement data has plunged regional energy networks into a double-blind dilemma: high-precision state estimation is difficult without accurate topological information, while the lack of accurate state information, in turn, hinders the timely identification of topological changes, severely impacting the real-time observability and safe dispatch of regional energy networks.

[0003] Existing solutions to this problem all have significant shortcomings in practical engineering applications. The traditional sequential approach of identifying the topology first and then estimating the state severs the inherent physical coupling between the topology and electrical state in a power system. If there is a deviation in the upstream topology identification, the error will be amplified in the state estimation stage, leading to severely distorted final results. Furthermore, the computational process is cumbersome and cannot meet the millisecond-level response requirements of proactive regional energy networks to sudden events. On the other hand, although data-driven intelligent algorithms have advantages in computational speed, existing end-to-end models often lack an inherent adherence to the physical laws of power systems, causing their outputs to frequently violate basic physical principles. When faced with complex faults or operating modes not seen in the training set, the reliability and generalization ability of the models drop significantly, making them unreliable as a basis for grid dispatching decisions. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a joint estimation method and system for energy-consuming networks based on physical mask autoencoders.

[0005] Firstly, the objective of this invention can be achieved through the following technical solution: a joint estimation method for energy-consuming networks based on a physical mask autoencoder, the method comprising the following steps: The topology and measurement data of the regional energy consumption network are obtained; the topology and measurement data of the regional energy consumption network are mapped to obtain an undirected graph model; a double random masking operation is performed on the undirected graph model to obtain masked graph data, wherein the masked graph data includes a set of masked nodes and a set of masked edges. The mask image data is input into a pre-trained physical information graph-based mask autoencoder model, and the output is the estimation results of the regional energy network topology and operating status that satisfy physical constraints. The physical information graph mask autoencoder model is trained using pre-received sample mask graph data and a multi-objective joint loss function that includes physical residual loss; the physical residual loss is generated based on the soft admittance matrix and the reconstructed node voltage state.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the double random masking operation on the undirected graph model includes randomly replacing node feature vectors to generate a masked node set, and randomly removing edge connections in the graph to generate a masked edge set, as follows: An undirected graph model for establishing an operational snapshot of a regional energy consumption network at a certain moment. G =( A,X,E ),in A Let be the adjacency matrix, representing the adjacency matrix that reflects the connection relationships of the edge set; X This is a node feature matrix, which contains active power, reactive power, and voltage measurement data for each node in the regional energy network. E This is the set of edges, corresponding to the line connections and switching states in the regional energy network. Randomly sample a subset from the node set. X mask ⊂ X As a set of mask nodes; the original feature vector of each node in the set of mask nodes is replaced with a unified and learnable mask identifier vector, which is automatically updated through backpropagation during training; Randomly select a subset of edges E mask ⊂ E As a set of masked edges; from the adjacency matrix A Removing the corresponding connections results in an incomplete mask diagram structure. G̃ =( X̃, Ẽ ), as mask image data.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the physical information graph mask autoencoder model includes a graph attention encoder, a state reconstruction decoder, and a topology identification decoder; The process of obtaining the physical residual loss is as follows: The sample mask image data is acquired and input into the graph attention encoder, which outputs a latent feature representation matrix. The sample mask image data and the latent feature representation matrix are input into the state reconstruction decoder, which outputs the reconstructed node voltage state. The reconstructed node voltage state includes the voltage amplitude and phase angle of the node. The latent feature representation matrix is ​​input into the topology identification decoder, and the output is the probability adjacency matrix. Based on the predicted branch existence probability in the probability adjacency matrix, the physical admittance parameters of the candidate branches are weighted and aggregated to generate the soft admittance matrix.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of inputting the sample mask image data into the graph attention encoder and outputting the latent feature representation matrix, as follows: For the encoder's first l Layer, only aggregate nodes i The set of visible neighbor nodes Information.

[0009] compute nodes i Its visible neighbor nodes Attention coefficient between The calculation formula is: in, For nodes i In the l Features of the layer W (l) For a learnable weight matrix, || denotes the concatenation operation; The node features are updated by weighted aggregation and non-linear activation of neighbor node features based on the attention coefficient: in, σ It is a non-linear activation function; after multi-layer aggregation, the latent feature representation is obtained. Z .

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of inputting the sample mask image data and the latent feature representation matrix into the state reconstruction decoder and outputting the reconstructed node voltage state, as follows: The output latent feature representation matrix ZThe incomplete adjacency matrix of the sample mask image data is simultaneously input into the state reconstruction decoder; the state reconstruction decoder includes at least one layer of auxiliary graph attention network, which uses the partial topological information provided by the incomplete adjacency matrix to perform secondary aggregation and feature refinement on the latent feature representation matrix to obtain the refined state feature matrix. Z' ; The refined state feature matrix Z' The input is fed into the linear mapping layer, which projects the feature dimensions of each node onto the power system state space and directly outputs the voltage magnitude vector and voltage phase angle vector of all nodes in the network at the current moment, as the reconstructed node voltage state. The process of inputting the latent feature representation matrix into the topological identification decoder and outputting the probability adjacency matrix is ​​as follows: For any two nodes in the regional energy consumption network i and j From the latent feature representation matrix Z Extract the corresponding feature vector z i , z j Perform feature concatenation operation to obtain the joint feature vector of node pairs; The joint feature vector of the node pairs is input into the topology identification decoder, which adopts a two-layer multilayer perceptron structure. First, the hidden layer features are calculated through a first-layer linear transformation and a nonlinear activation function. Then, the connection discrimination logic value is output through a second-layer linear transformation, and its calculation formula is as follows: in, W 1 , b 1 The first layer weight matrix and bias vector are... w 2 , b 2 For the second layer weight vector and bias scalar, σ a For activation functions; The Sigmoid activation function is used to map the connection discrimination logic values ​​to branch connection probabilities in the interval (0,1). pᵢ : Using multilayer perceptrons to process node pairs ( i , j The latent feature concatenation vector outputs the connection probability. pᵢ The calculation formula is as follows: The connection probabilities of all node pairs are summed to form a probabilistic adjacency matrix.  , where matrix elements ij Representation Nodes i With nodes j There is confidence that there is a physical connection between them.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of weighted aggregation of the physical admittance parameters of the candidate branches to generate a soft admittance matrix, as follows: Branch connection probabilities from the topology decoding branch output p k Admittance parameters of candidate branches Y k Perform weighted summation to construct the soft admittance matrix. Y soft : in, Let be the set of all existing branches; based on the soft admittance matrix, make the probability of topology prediction differentiable with respect to the physical power flow residual.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the physical residual loss is constructed based on the soft admittance matrix and the reconstructed node voltage state, and the power flow physical residual equation is constructed according to Kirchhoff's laws, as follows: Based on the node voltage magnitude vector within the reconstructed node voltage state |V̂| With voltage phase vector θ By combining the data, a complex voltage vector is constructed for all nodes in the network. V̂ , of which i The complex voltage at each node is expressed as: ; Based on the fully readable admittance matrix Y soft Using Kirchhoff's current law and power equations, the estimated complex injected power vector of all nodes in the network is calculated. Ŝ Its matrix operation formula is: in, diag ( V̂ ) represents a diagonal matrix with complex voltage vectors as its diagonal elements; Obtain the actual measured power vector from the regional energy network measurement snapshot. S ,vector S It consists of measured active power and measured reactive power; the complex injected power vector is calculated and estimated. Ŝ Compared with the actual measured complex power vector S The Euclidean distance between them defines the physical power flow residual loss. Lphy : By minimizing the physical power flow residual loss, the voltage state reconstructed by the constraint model and the predicted topology satisfy the power flow physical equations of the power system.

[0013] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: training the pre-trained physical information graph mask autoencoder model as follows: A multi-objective joint loss function is constructed, which includes mask node reconstruction loss, topology classification loss, and physical residual loss. The encoder and dual-branch decoder are then trained end-to-end by minimizing the joint loss function, as follows. Construct a multi-objective joint loss function that includes masked node reconstruction loss, topology classification loss, physical power flow residual loss, and non-masked node anchor point loss: in, L recon This represents the reconstruction error of the mask nodes. L topo For the binary cross-entropy loss of the topology, L phy The physical power flow residual is calculated based on the soft admittance matrix. L unmask For anchor point loss of non-masked nodes; λ T , λ Φ , β Hyperparameters are used to balance the weights of each loss term; Using physical power flow residual loss L phy As a physical consistency regularization term; due to the fully compliant microadmittance matrix Y soft Enforcing Kirchhoff's laws, the physical information graph mask autoencoder model minimizes... L phy During the process, the actual measured power vector will be automatically filtered. S Random measurement noise that does not satisfy the physical conservation law of the power system is used to guide the system state manifold based on the physical information graph mask autoencoder model to converge to the physical truth value, thereby realizing physical denoising of noisy measurement data.

[0014] Secondly, in order to achieve the above objectives, this invention discloses a joint estimation system for energy-consuming networks based on a physical mask autoencoder, comprising: A random masking module is used to acquire the topology and measurement data of the regional energy consumption network; to map the topology and measurement data of the regional energy consumption network to obtain an undirected graph model; and to perform a double random masking operation on the undirected graph model to obtain masked graph data, wherein the masked graph data includes a set of masked nodes and a set of masked edges. The joint estimation module is used to input the mask map data into a pre-trained physical information graph-based mask autoencoder model and output the estimation results of the regional energy network topology and operating status that meet the physical constraints. The physical information graph mask autoencoder model is trained using pre-received sample mask graph data and a multi-objective joint loss function that includes physical residual loss; the physical residual loss is generated based on the soft admittance matrix and the reconstructed node voltage state.

[0015] In another aspect of the present invention, in order to achieve the above-mentioned objective, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor. When the processor loads and executes the computer program, it employs the energy network joint estimation method based on physical mask autoencoder as described above.

[0016] The beneficial effects of this invention are: This invention constructs a graph model based on regional energy consumption network measurement data and uses a graph mask autoencoder to extract latent features from sparse measurement data. Then, it employs a dual-branch parallel decoding and fully differentiable physical constraint mechanism to synchronously restore the topology and state, improving the physical consistency of the model. Finally, it uses physical power flow residuals to guide model convergence, achieving accurate perception of the regional energy consumption network under the double-blind dilemma. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the workflow of the present invention; Figure 3 This is a schematic diagram of the overall architecture of the PI-GMAE deep neural network in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the node voltage estimation error between the present invention embodiment and the benchmark method; Figure 5This is a schematic diagram comparing real-time tracking performance under topological change scenarios in this embodiment of the invention; Figure 6 This is a schematic diagram of the real-time joint estimation performance index curve during continuous dynamic topology switching according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the deep neural network training results in an embodiment of the present invention; Figure 8 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: like Figure 1 As shown, a joint estimation method for energy-consuming networks based on physical mask autoencoders includes the following steps: S101: Obtain the topology and measurement data of the regional energy consumption network; map the topology and measurement data of the regional energy consumption network to obtain an undirected graph model; perform a double random masking operation on the undirected graph model to obtain masked graph data, wherein the masked graph data includes a set of masked nodes and a set of masked edges. Performing a double random masking operation on an undirected graph model involves randomly replacing node feature vectors to generate a masked node set, and randomly removing edge connections in the graph to generate a masked edge set. The process is as follows: An undirected graph model for establishing an operational snapshot of a regional energy consumption network at a certain moment. G =( A,X,E ),in A Let be the adjacency matrix, representing the adjacency matrix that reflects the connection relationships of the edge set; X This is a node feature matrix, which contains active power, reactive power, and voltage measurement data for each node in the regional energy network. E This is the set of edges, corresponding to the line connections and switching states in the regional energy network. Randomly sample a subset from the node set. X mask ⊂ X As a set of mask nodes; the original feature vector of each node in the set of mask nodes is replaced with a unified and learnable mask identifier vector, which is automatically updated through backpropagation during training; Randomly select a subset of edgesE mask ⊂ E As a set of masked edges; from the adjacency matrix A Removing the corresponding connections results in an incomplete mask diagram structure. G̃ =( X̃, Ẽ ), as mask image data.

[0020] S102: Input the mask image data into a pre-trained physical information graph-based mask autoencoder model, and output the estimation results of the regional energy network topology and operating status that meet the physical constraints; The physical information graph mask autoencoder model is trained using pre-received sample mask graph data and a multi-objective joint loss function that includes physical residual loss; the physical residual loss is generated based on the soft admittance matrix and the reconstructed node voltage state.

[0021] The physical information graph mask autoencoder model includes a graph attention encoder, a state reconstruction decoder, and a topology identification decoder. The process of obtaining the physical residual loss is as follows: The sample mask image data is acquired and input into the graph attention encoder, which outputs a latent feature representation matrix. The sample mask image data and the latent feature representation matrix are input into the state reconstruction decoder, which outputs the reconstructed node voltage state. The reconstructed node voltage state includes the voltage amplitude and phase angle of the node. The latent feature representation matrix is ​​input into the topology identification decoder, and the output is the probability adjacency matrix. Based on the predicted branch existence probability in the probability adjacency matrix, the physical admittance parameters of the candidate branches are weighted and aggregated to generate the soft admittance matrix.

[0022] The process of inputting sample mask image data into a graph attention encoder and outputting the latent feature representation matrix is ​​as follows: For the encoder's first l Layer, only aggregate nodes i The set of visible neighbor nodes Information.

[0023] compute nodes i Its visible neighbor nodes Attention coefficient between The calculation formula is: in, For nodes i In the l Features of the layer W(l) For a learnable weight matrix, || denotes the concatenation operation; The node features are updated by weighted aggregation and non-linear activation of neighbor node features based on the attention coefficient: in, σ It is a non-linear activation function; after multi-layer aggregation, the latent feature representation is obtained. Z .

[0024] The process of inputting sample mask image data and latent feature representation matrix into the state reconstruction decoder and outputting the reconstructed node voltage state is as follows: The output latent feature representation matrix Z The incomplete adjacency matrix of the sample mask image data is simultaneously input into the state reconstruction decoder; the state reconstruction decoder includes at least one layer of auxiliary graph attention network, which uses the partial topological information provided by the incomplete adjacency matrix to perform secondary aggregation and feature refinement on the latent feature representation matrix to obtain the refined state feature matrix. Z' ; The refined state feature matrix Z' The input is fed into the linear mapping layer, which projects the feature dimensions of each node onto the power system state space and directly outputs the voltage magnitude vector and voltage phase angle vector of all nodes in the network at the current moment, as the reconstructed node voltage state. The process of inputting the latent feature representation matrix into the topological identification decoder and outputting the probability adjacency matrix is ​​as follows: For any two nodes in the regional energy consumption network i and j From the latent feature representation matrix Z Extract the corresponding feature vector z i , z j Perform feature concatenation operation to obtain the joint feature vector of node pairs; The joint feature vector of the node pairs is input into the topology identification decoder, which adopts a two-layer multilayer perceptron structure. First, the hidden layer features are calculated through a first-layer linear transformation and a nonlinear activation function. Then, the connection discrimination logic value is output through a second-layer linear transformation, and its calculation formula is as follows: in, W 1 , b 1 The first layer weight matrix and bias vector are... w 2 , b 2For the second layer weight vector and bias scalar, σ a For activation functions; The Sigmoid activation function is used to map the connection discrimination logic values ​​to branch connection probabilities in the interval (0,1). pᵢ : Using multilayer perceptrons to process node pairs ( i , j The latent feature concatenation vector outputs the connection probability. pᵢ The calculation formula is as follows: The connection probabilities of all node pairs are summed to form a probabilistic adjacency matrix.  , where matrix elements  ij Representation Nodes i With nodes j There is confidence that there is a physical connection between them.

[0025] The process of weighted aggregation of the physical admittance parameters of candidate branches to generate the soft admittance matrix is ​​as follows: Branch connection probabilities from the topology decoding branch output p k Admittance parameters of candidate branches Y k Perform weighted summation to construct the soft admittance matrix. Y soft : in, Let be the set of all existing branches; based on the soft admittance matrix, make the probability of topology prediction differentiable with respect to the physical power flow residual.

[0026] The physical residual loss is constructed based on the soft admittance matrix and the reconstructed node voltage state, and the power flow physical residual equation is constructed according to Kirchhoff's laws. The process is as follows: Based on the node voltage magnitude vector within the reconstructed node voltage state |V̂| With voltage phase vector θ By combining the data, a complex voltage vector is constructed for all nodes in the network. V̂ , of which i The complex voltage at each node is expressed as: ; Based on the fully readable admittance matrix Y soft Using Kirchhoff's current law and power equations, the estimated complex injected power vector of all nodes in the network is calculated. Ŝ Its matrix operation formula is: in, diag ( V̂ ) represents a diagonal matrix with complex voltage vectors as its diagonal elements; Obtain the actual measured power vector from the regional energy network measurement snapshot. S ,vector S It consists of measured active power and measured reactive power; the complex injected power vector is calculated and estimated. Ŝ Compared with the actual measured complex power vector S The Euclidean distance between them defines the physical power flow residual loss. L phy : By minimizing the physical power flow residual loss, the voltage state reconstructed by the constraint model and the predicted topology satisfy the power flow physical equations of the power system.

[0027] The pre-trained model based on the physical information graph mask autoencoder is trained as follows: A multi-objective joint loss function is constructed, which includes mask node reconstruction loss, topology classification loss, and physical residual loss. The encoder and dual-branch decoder are then trained end-to-end by minimizing the joint loss function, as follows. Construct a multi-objective joint loss function that includes masked node reconstruction loss, topology classification loss, physical power flow residual loss, and non-masked node anchor point loss: in, L recon This represents the reconstruction error of the mask nodes. L topo For the binary cross-entropy loss of the topology, L phy The physical power flow residual is calculated based on the soft admittance matrix. L unmask For anchor point loss of non-masked nodes; λ T , λ Φ , β Hyperparameters are used to balance the weights of each loss term; Using physical power flow residual loss L phy As a physical consistency regularization term; due to the fully compliant microadmittance matrix Y soft Enforcing Kirchhoff's laws, the physical information graph mask autoencoder model minimizes... L phy During the process, the actual measured power vector will be automatically filtered. SRandom measurement noise that does not satisfy the physical conservation law of the power system is used to guide the system state manifold based on the physical information graph mask autoencoder model to converge to the physical truth value, thereby realizing physical denoising of noisy measurement data.

[0028] Specifically, a multi-objective joint loss function is constructed, comprising masked node reconstruction loss, topology classification loss, physical power flow residual loss, and non-masked node anchor point loss. A self-supervised iterative training strategy is adopted, and a mixed scenario sample library containing different historical moments, load levels, and topologies of the regional energy consumption network is established during the training process. A rolling sampling mechanism is used to continuously extract changing data snapshots as input.

[0029] A dynamically changing random mask pattern is applied to each set of input data, forcing the model to learn the topological logic of regional energy consumption networks with diverse and universal applicability. Simultaneously, physical power flow residual loss is utilized. L phy As a physical consistency regularization term, it automatically filters actual measurement data by utilizing the physical constraint properties of the soft admittance matrix. Ŝ Random measurement noise that does not satisfy Kirchhoff's laws guides the model to converge to a system state manifold that approximates the physical truth. The joint loss function is minimized using stochastic gradient descent, and the encoder and decoder parameters are updated via backpropagation until the model converges.

[0030] Specifically, the present invention will be further illustrated below through embodiments: Step A: Construct a simulation dataset of a regional energy consumption network hybrid scenario with a high proportion of distributed photovoltaic access.

[0031] First, a hybrid scenario simulation dataset based on the IEEE 33-node regional energy network system is constructed. To simulate the operating characteristics of active regional energy networks and avoid security and privacy issues in real data, measurement and topology data of regional energy networks containing multiple types of topologies and photovoltaic access are generated based on the actual power grid operation rules. By connecting distributed power sources and designing various typical topology operation modes, a hybrid scenario simulation dataset containing measurement noise and random missing data is generated as the basis for model validation. Step B: Perform node voltage estimation performance analysis under static scenarios.

[0032] The generated regional energy network measurement data is mapped to a graph data model, and a double random masking operation is performed on node features and edge connections to simulate the scenarios of missing edge-side measurements and incomplete topology information recording in actual regional energy network operation, and to analyze the node voltage estimation performance under static scenarios. The trained model is used to quantitatively test the state estimation accuracy under a single topology scenario. The error distribution of node voltage estimation under missing measurement conditions is compared with that of traditional weighted least squares method and lag reference method, especially the error suppression capability for densely loaded areas and photovoltaic access points, to prove the effectiveness of the model in eliminating cascade errors. Step C: Perform real-time tracking performance analysis under dynamic topology mutation.

[0033] A deep neural network model based on a physical information graph mask autoencoder is constructed. A graph attention encoder is used to extract latent features from the mask graph data, and a parallel state reconstruction decoder and topology identification decoder output the overall network voltage state and topology connection probability, respectively. Real-time tracking performance under dynamic topology changes is analyzed. A topology switching event caused by switching actions during continuous operation of a regional energy network is simulated. The model's identification response speed at the moment of topology change and the stability of state estimation during transient processes with unknown topology are tested to verify whether the model has zero-latency real-time sensing capabilities. Step D: Perform convergence analysis of model self-supervised training.

[0034] Record the multi-objective loss function variation curves of the model during the pre-training phase, including total loss, mask reconstruction loss, and physical residual loss. Analyze the decreasing trend of each loss as the number of iterations increases, thereby verifying the effectiveness of the physical constraint mechanism in guiding the model parameters to converge to the physical truth, and ensuring that the deep learning model does not violate the fundamental physical laws of power systems; Step E: Perform global robustness statistics under different measurement missing rates.

[0035] The statistical distribution of topology identification accuracy and state estimation error of the test model under extremely harsh observation conditions with a gradually increasing rate of missing measurement data is examined. By comparing the decay rate of various performance indicators, the efficiency of the proposed method in utilizing sparse data and its global robustness under the double-blind dilemma are evaluated.

[0036] Furthermore, the implementation process of step A is as follows: To verify the effectiveness, reliability and operability of the present invention under different regional energy network operating environments.

[0037] In this embodiment, the standard IEEE 33-node regional energy network system is first selected as the test benchmark. Considering the operation characteristics of high proportion of renewable energy access in modern active regional energy networks, the present invention connects distributed photovoltaic units with a capacity of 400 kilowatts to nodes 6, 21 and 31 of the test system, and generates continuous 24-hour operation snapshot data based on the actual collected load curve and photovoltaic output curve.

[0038] like Figure 2 As shown in the method flow, in the data construction phase, to cover various topologies that may occur in the operation of the regional energy network, this invention designed a hybrid sample library containing six different topology scenarios. Table 1, showing the switch state settings, details the network structure of these six scenarios: Scenario A is the original radial topology, i.e., no switching action is performed; Scenario B simulates a power transfer scenario after a local line fault by opening switches 6-7 and closing switches 20-7; Scenarios C to F further simulate more complex line maintenance or multiple fault situations, involving the closing of multiple sets of tie switches and the opening of sectionalizing switches, respectively. Based on the above topology scenarios, this invention generates corresponding power flow data and superimposes Gaussian white noise into the measurement data, while randomly setting measurement missing values ​​at a rate of 10% to 50%, thereby constructing a hybrid scenario simulation dataset that approximates a real double-blind environment, serving as the basis for subsequent model training and testing.

[0039] Table 1 Furthermore, the implementation process of step B is as follows: This step aims to conduct in-depth verification. Figure 3 The physical information graph masked autoencoder deep neural network architecture is shown to demonstrate the state estimation fidelity under a static single topological profile, and the ability of this model to overcome the error cascading propagation effect in traditional serial estimation architectures is evaluated.

[0040] In the specific experimental implementation, this invention loads a fully pre-trained PI-GMAE model and selects scenario B defined in step A as a typical test condition. This scenario simulates the operation of a regional energy network after a fault, where power is transferred by disconnecting switches six and seven and closing switches twenty and seven. To simulate the harsh observation environment in a real power grid, this invention sets the missing rate of node measurement data to 30%. To construct a performance benchmark with objective reference value, the experiment introduces two control groups: the first group uses the ideal weighted least squares method assuming a known real topology, which represents the theoretical upper limit of state estimation accuracy under an omniscient perspective; the second group uses the lag weighted least squares method commonly used in engineering practice, which uses the old topology parameters from the previous time step in the current time step, aiming to simulate the topology-aware lag phenomenon commonly found in actual scheduling systems.

[0041] like Figure 4 The figure shows a comparison of the spatial distribution of voltage amplitude estimation errors between the proposed method and the two benchmark methods described above across 33 nodes in the entire network. The horizontal axis represents the physical node number of the regional energy network, and the vertical axis represents the absolute deviation of the voltage amplitude estimation result from the true value. Through detailed analysis of the error curve shape in the figure, the following conclusions can be drawn: First, the error curve representing the lag-weighted least squares method exhibits significant pulse-like abrupt changes across the entire network, demonstrating that the accuracy of the topology parameters is fundamental to ensuring the convergence of state estimation. It is noteworthy that this large voltage estimation deviation does not follow a uniform distribution like Gaussian white noise, but rather displays strong spatial clustering and physical correlation. For example... Figure 4 As shown, the error peaks are mainly concentrated in the heavily loaded areas of the system and at the boundary nodes where distributed photovoltaic power sources are connected, such as near nodes six, eighteen, and thirty-three. This phenomenon profoundly reveals from a physical mechanism that in active regional energy networks with drastic fluctuations on both the source and load sides, the cognitive bias of the topology is no longer an isolated parameter error. It is propagated along the physical connection lines of the power grid and, through the nonlinear solution process of the power flow equations, the structural error is further amplified into a node-level state estimation bias, resulting in a complete distortion of the situational awareness results at key nodes, i.e., the so-called error cascade amplification effect.

[0042] In contrast, the error curve representing the method of this invention is extremely flat throughout the coordinate system and always closely follows the zero-error axis, with its numerical trajectory highly overlapping with the weighted least squares curve under ideal conditions. This result strongly demonstrates... Figure 3The end-to-end neural network architecture shown successfully decouples and coordinates topological and state features in the deep latent space by introducing a fully differentiable physical constraint mechanism. Even in the face of the "double-blind" dilemma of losing 30% of the measurement data, the model can still accurately infer the topology and electrical state that highly matches the current physical cross-section using the remaining sparse data fragments. This effectively cuts off the link of topology error propagation to the state estimation stage, achieving zero-error reconstruction of the voltage state of the entire network and demonstrating its excellent robustness in complex regional energy network environments.

[0043] Furthermore, the implementation process of step C is as follows: This step aims to comprehensively verify the model's instantaneous capture capability and dynamic adaptability in the face of unplanned switching actions or sudden failures in the regional energy network.

[0044] To simulate continuous and unpredictable network reconfiguration events that may occur in a real power grid, this invention constructs a long-term dynamic test sequence containing 120 consecutive time segments, with a time resolution of five minutes. This sequence simulates the complex dynamic evolution of a regional energy network from stable operation in scenario B, to a sudden switch to scenario C, and then back to scenario F. A 30% continuous measurement mask is applied throughout the process to test the model's robustness under the dual pressures of data loss and structural abrupt changes.

[0045] like Figure 5 The figure shows a comparison of the real-time tracking performance of the method of the present invention at the moment of topological change. The horizontal axis of the figure represents the time section, and the vertical axis represents the F1 score of topology identification. Observing the instant the switching action occurs, it can be found that the benchmark method, due to its reliance on historical topology information for state estimation, has an inherent perception lag, causing its F1 score to drop sharply at the switching point, and the system falls into a temporary topology blind zone. In contrast, the F1 score of the method of the present invention remains stable above 0.97, and its predicted branch connection probability completes an instantaneous flip at the same moment the physical topology changes, achieving zero-delay accurate identification of topology changes. This excellent transient response capability is attributed to the model learning the inherent physical mapping relationship between measurement data patterns and network topology structure in self-supervised training, enabling it to immediately infer the corresponding structural change the moment the measurement data reflects a new physical entity.

[0046] like Figure 6As shown, the curves illustrating the changes in the model's comprehensive performance indicators during a continuous dynamic switching sequence lasting 120 minutes are further presented. The figure details the evolution of the mean absolute error of state estimation, root mean square error, and topology identification accuracy over time. Analysis of the curves reveals that even after multiple complex topology reconfigurations and load fluctuations, the state estimation error curve of the proposed method remains flat and stable, without any divergence or oscillations, and the topology identification accuracy closely matches the theoretical upper limit of 100%. In stark contrast, the baseline method experiences order-of-magnitude error spikes during the transition window of each topology switch, indicating that the system completely loses observability during this period. These results strongly demonstrate that the unified differential architecture proposed in this invention can effectively overcome the perception blind spots of traditional methods, maintaining high-fidelity panoramic perception capabilities during long-term continuous operation, and providing reliable real-time data support for fault location, isolation, and power restoration in regional energy networks.

[0047] Furthermore, the implementation process of step D is as follows: This step aims to deeply analyze the training dynamics of deep neural networks and verify the convergence effectiveness of fully differentiable physical constraints in end-to-end self-supervised learning.

[0048] To address the problem that pure data-driven models are prone to getting trapped in local minima or outputting results that violate physical principles during training, this invention innovatively introduces a fully identifiable soft admittance matrix and physical power flow residuals as strong regularization terms, internalizing Kirchhoff's laws of the power system into part of the loss function, thereby reshaping the model's optimal solution space.

[0049] like Figure 7 As shown, the convergence curve of the multi-objective loss function of the model during the pre-training phase is illustrated. The figure details the evolution trends of the average total loss, edge reconstruction loss, and node feature reconstruction loss with each iteration. Through refined analysis of the training process, significant stage-specific characteristics can be observed: In the early stages of training, specifically the first twenty iterations, all three loss curves exhibited a rapid decline with extremely high slopes. This indicates that the graph attention encoder is efficiently capturing low-dimensional key features of the regional energy network topology and electrical state from the high-dimensional masked image data, and the model parameters are rapidly adjusted to adapt to the statistical distribution of the data.

[0050] As the iterations progressed to around twenty rounds, the curve gradually stabilized and converged to the minimum region without significant oscillations or divergence. This phenomenon has important physical significance: it proves that the introduced physical power flow residual loss did not cause gradient vanishing or gradient exploding problems. Instead, it served as an effective physical guidance mechanism, establishing a smooth gradient propagation channel from the physical residual to the network parameters. This mechanism forces the model's search path to proceed along a manifold direction that conforms to physical laws, thereby accelerating the convergence of model parameters to the physical truth and ensuring that the final generated topology and state results strictly adhere to the physical operating rules of the power system while meeting data fitting accuracy requirements.

[0051] Furthermore, the implementation process of step E is as follows: This step aims to quantitatively evaluate the model's global robustness and applicability boundaries under extremely sparse observation conditions. By traversing the measurement data missing rates at different gradients, the model's survivability in harsh conditions such as sensor failure, communication packet loss, or malicious network attacks is verified.

[0052] The experimental design sets the missing rate of node measurement data to gradually increase from 10% to 50%, thereby constructing a continuous test scenario from normal operation to large-scale data failure. Table 2, as a joint estimation performance index statistics table, details the key performance indicators of the model on the joint validation set, such as the mean absolute error of voltage amplitude estimation, root mean square error, and F1 score and accuracy of topology identification, under different missing rate levels.

[0053] A deep analysis of the multidimensional data in the statistical tables reveals that, with a significant increase in the data missing rate, the performance indicators of the method in this invention exhibit a resilient and gradual decline, rather than the precipitous drop seen in traditional methods. Specifically, in mild to moderate missing data scenarios with a measurement missing rate of 20% or less, the model demonstrates excellent reconstruction capabilities. Its topology identification F1 score and accuracy both reach the theoretical maximum of 1.0, indicating that the model successfully achieves error-free reconstruction of the regional energy network topology, while maintaining an extremely low state estimation error. Even under extreme conditions with a measurement missing rate of 50%, i.e., when half of the sensor data is completely unavailable, the model's performance remains highly stable. Data shows that the average absolute error of voltage estimation at this point only increases slightly compared to a 10% missing rate, proving that the state estimation function does not fail due to the large amount of data loss; simultaneously, the topology identification F1 score remains high at 0.985, and the accuracy remains above 0.973.

[0054] This experimental result powerfully demonstrates from a data perspective that the physical information graph mask autoencoder architecture employed in this invention possesses a unique "part-to-whole" inference capability. Through high-intensity mask training, this architecture successfully forces the model to learn the deep nonlinear mapping relationship between sparse measurement data and the global physical state of the system, rather than relying solely on surface statistical regularities. This allows for the effective utilization of remaining fragmented information to achieve accurate panoramic perception of the entire network's operational status even under the double-blind dilemma.

[0055] Table 2 Example 2: To achieve the above objectives, based on Example 1, the present invention discloses... Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0056] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0057] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0058] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A joint estimation method for energy-consuming networks based on physical mask autoencoders, characterized in that, The method includes the following steps: The topology and measurement data of the regional energy consumption network are obtained; the topology and measurement data of the regional energy consumption network are mapped to obtain an undirected graph model; a double random masking operation is performed on the undirected graph model to obtain masked graph data, wherein the masked graph data includes a set of masked nodes and a set of masked edges. The mask image data is input into a pre-trained physical information graph-based mask autoencoder model, and the output is the estimation results of the regional energy network topology and operating status that satisfy physical constraints. The physical information graph mask autoencoder model is trained using pre-received sample mask graph data and a multi-objective joint loss function that includes physical residual loss; the physical residual loss is generated based on the soft admittance matrix and the reconstructed node voltage state.

2. The joint estimation method for energy-consuming networks based on physical mask autoencoders according to claim 1, characterized in that, The double random masking operation on the undirected graph model includes randomly replacing node feature vectors to generate a masked node set, and randomly removing edge connections in the graph to generate a masked edge set, as follows: An undirected graph model for establishing an operational snapshot of a regional energy consumption network at a certain moment. G =( A,X,E ),in A Let be the adjacency matrix, representing the adjacency matrix that reflects the connection relationships of the edge set; X This is a node feature matrix, which contains active power, reactive power, and voltage measurement data for each node in the regional energy network. E This is the set of edges, corresponding to the line connections and switching states in the regional energy network. Randomly sample a subset from the node set. X mask ⊂ X As a set of mask nodes; the original feature vector of each node in the set of mask nodes is replaced with a unified and learnable mask identifier vector, which is automatically updated through backpropagation during training; Randomly select a subset of edges E mask ⊂ E As a set of masked edges; from the adjacency matrix A Removing the corresponding connections results in an incomplete mask diagram structure. G̃ =( X̃, Ẽ ), as mask image data.

3. The joint estimation method for energy-consuming networks based on physical mask autoencoders according to claim 1, characterized in that, The physical information graph mask autoencoder model includes a graph attention encoder, a state reconstruction decoder, and a topology identification decoder. The process of obtaining the physical residual loss is as follows: The sample mask image data is acquired and input into the graph attention encoder, which outputs a latent feature representation matrix. The sample mask image data and the latent feature representation matrix are input into the state reconstruction decoder, which outputs the reconstructed node voltage state. The reconstructed node voltage state includes the voltage amplitude and phase angle of the node. The latent feature representation matrix is ​​input into the topology recognition decoder, and the output is the probability adjacency matrix; Based on the predicted branch existence probability in the probabilistic adjacency matrix, the physical admittance parameters of candidate branches are weighted and aggregated to generate a soft admittance matrix.

4. The joint estimation method for energy-consuming networks based on physical mask autoencoders according to claim 3, characterized in that, The process of inputting sample mask image data into the graph attention encoder and outputting the latent feature representation matrix is ​​as follows: For the encoder's first l Layer, only aggregate nodes i The set of visible neighbor nodes Information. compute nodes i Its visible neighbor nodes Attention coefficient between The calculation formula is: in, For nodes i In the l Features of the layer W (l) For a learnable weight matrix, || denotes the concatenation operation; The node features are updated by weighted aggregation and non-linear activation of neighbor node features based on the attention coefficient: in, σ It is a non-linear activation function; after multi-layer aggregation, the latent feature representation is obtained. Z .

5. The joint estimation method for energy-consuming networks based on physical mask autoencoders according to claim 4, characterized in that, The process of inputting the sample mask image data and the latent feature representation matrix into the state reconstruction decoder and outputting the reconstructed node voltage state is as follows: The output latent feature representation matrix Z The incomplete adjacency matrix of the sample mask image data is simultaneously input into the state reconstruction decoder; the state reconstruction decoder includes at least one layer of auxiliary graph attention network, which uses the partial topological information provided by the incomplete adjacency matrix to perform secondary aggregation and feature refinement on the latent feature representation matrix to obtain the refined state feature matrix. Z' ; The refined state feature matrix Z' The input is fed into the linear mapping layer, which projects the feature dimensions of each node onto the power system state space and directly outputs the voltage magnitude vector and voltage phase angle vector of all nodes in the network at the current moment, as the reconstructed node voltage state. The process of inputting the latent feature representation matrix into the topological identification decoder and outputting the probability adjacency matrix is ​​as follows: For any two nodes in the regional energy consumption network i and j From the latent feature representation matrix Z Extract the corresponding feature vector z i , z j Perform feature concatenation operation to obtain the joint feature vector of node pairs; The joint feature vector of the node pairs is input into the topology identification decoder, which adopts a two-layer multilayer perceptron structure. First, the hidden layer features are calculated through a first-layer linear transformation and a nonlinear activation function. Then, the connection discrimination logic value is output through a second-layer linear transformation, and its calculation formula is as follows: in, W 1 , b 1 The first layer weight matrix and bias vector are... w 2 , b 2 For the second layer weight vector and bias scalar, σ a For activation functions; The Sigmoid activation function is used to map the connection discrimination logic values ​​to branch connection probabilities in the interval (0,1). pᵢ : Using multilayer perceptrons to process node pairs ( i , j The latent feature concatenation vector outputs the connection probability. pᵢ The calculation formula is as follows: The connection probabilities of all node pairs are summed to form a probabilistic adjacency matrix.  , where matrix elements  ij Representation Nodes i With nodes j There is confidence that there is a physical connection between them.

6. The joint estimation method for energy-consuming networks based on physical mask autoencoders according to claim 5, characterized in that, The process of weighted aggregation of the physical admittance parameters of candidate branches to generate a soft admittance matrix is ​​as follows: Branch connection probabilities from the topology decoding branch output p k Admittance parameters of candidate branches Y k Perform weighted summation to construct the soft admittance matrix. Y soft : in, Let be the set of all existing branches; based on the soft admittance matrix, make the probability of topology prediction differentiable with respect to the physical power flow residual.

7. The joint estimation method for energy-consuming networks based on physical mask autoencoders according to claim 1, characterized in that, The physical residual loss is constructed based on the soft admittance matrix and the reconstructed node voltage state, and the power flow physical residual equation is constructed according to Kirchhoff's laws. The process is as follows: Based on the node voltage magnitude vector within the reconstructed node voltage state |V̂| With voltage phase vector θ By combining the data, a complex voltage vector is constructed for all nodes in the network. V̂ , of which i The complex voltage at each node is expressed as: ; Based on the fully readable admittance matrix Y soft Using Kirchhoff's current law and power equations, the estimated complex injected power vector of all nodes in the network is calculated. Ŝ Its matrix operation formula is: in, diag ( V̂ ) represents a diagonal matrix with complex voltage vectors as its diagonal elements; Obtain the actual measured power vector from the regional energy network measurement snapshot. S ,vector S It consists of measured active power and measured reactive power; the complex injected power vector is calculated and estimated. Ŝ Compared with the actual measured complex power vector S The Euclidean distance between them defines the physical power flow residual loss. L phy : By minimizing the physical power flow residual loss, the voltage state reconstructed by the constraint model and the predicted topology satisfy the power flow physical equations of the power system.

8. The joint estimation method for energy-consuming networks based on physical mask autoencoders according to claim 1, characterized in that, The training of the pre-trained physical information graph mask autoencoder model is as follows: A multi-objective joint loss function is constructed, which includes mask node reconstruction loss, topology classification loss, and physical residual loss. The encoder and dual-branch decoder are then trained end-to-end by minimizing the joint loss function, as follows. Construct a multi-objective joint loss function that includes masked node reconstruction loss, topology classification loss, physical power flow residual loss, and non-masked node anchor point loss: in, L recon This represents the reconstruction error of the mask nodes. L topo For the binary cross-entropy loss of the topology, L phy The physical power flow residual is calculated based on the soft admittance matrix. L unmask For anchor point loss of non-masked nodes; λ T , λ Φ , β Hyperparameters are used to balance the weights of each loss term; Using physical power flow residual loss L phy As a physical consistency regularization term; due to the fully compliant microadmittance matrix Y soft Enforcing Kirchhoff's laws, the physical information graph mask autoencoder model minimizes... L phy During the process, the actual measured power vector will be automatically filtered. S Random measurement noise that does not satisfy the physical conservation law of the power system is used to guide the system state manifold based on the physical information graph mask autoencoder model to converge to the physical truth value, thereby realizing physical denoising of noisy measurement data.

9. A joint estimation system for energy-consuming networks based on a physical mask autoencoder, employing the joint estimation method for energy-consuming networks based on a physical mask autoencoder as described in any one of claims 1 to 8, characterized in that, include: The random mask module is used to obtain the topology and measurement data of the regional energy consumption network; The topology and measurement data of the regional energy consumption network are mapped to obtain an undirected graph model. The undirected graph model is then subjected to a double random masking operation to obtain masked graph data, wherein the masked graph data includes a set of masked nodes and a set of masked edges. The joint estimation module is used to input the mask map data into a pre-trained physical information graph-based mask autoencoder model and output the estimation results of the regional energy network topology and operating status that meet the physical constraints. The physical information graph mask autoencoder model is trained using pre-received sample mask graph data and a multi-objective joint loss function that includes physical residual loss; the physical residual loss is generated based on the soft admittance matrix and the reconstructed node voltage state.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs the energy network joint estimation method based on physical mask autoencoder as described in any one of claims 1 to 8.

Images