A power state estimation method and system based on an improved graph isomorphism network

Abstract

The application relates to the field of data analysis, in particular to a power state estimation method and system based on an improved graph isomorphism network, wherein the method is characterized in that: original measurement data and topological information of a target power system are acquired first, then a graph structure is constructed according to the topological information, the original measurement data are converted into node features and edge features of the graph structure, graph structure data are obtained, and finally, the graph structure data are input into a preset deep GINE model to perform feature extraction, message transmission and residual aggregation, and a target state estimation value is output. Compared with the prior art, the method has higher estimation accuracy and stronger robustness.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology. More specifically, this invention relates to a power state estimation method and system based on an improved graph isomorphic network. Background Technology

[0002] With the advancement of power technology, the complexity of power distribution networks has increased significantly, rendering traditional manual estimation and analysis methods inadequate for modern applications. To more quickly estimate power system data, existing technologies employ weighted least squares or deep neural network-based algorithms. However, both methods have the following drawbacks: (1) Weighted least squares method usually requires an accurate system topology matrix (Jacobi matrix) of the power system and linearization of the nonlinear power flow equations. This process is computationally complex, and its estimation accuracy and robustness will decrease significantly when there are small changes in the system topology or bad data. In addition, when a high proportion of distributed energy (such as wind power or photovoltaic) is connected, the volatility of grid operation data will increase. The above-mentioned weighted least squares method often has difficulty in identifying / distinguishing between the fluctuating data of new energy sources and the real bad data, resulting in lower estimation accuracy and slower estimation speed.

[0003] (2) Neural network-based algorithms struggle to effectively utilize the topological connections between nodes and the physical quantities on edges when processing graph-structured data, resulting in poor estimation accuracy. Furthermore, deep DNN networks also suffer from problems such as high training difficulty, vanishing gradients, and overfitting, which exacerbate the decrease in computational accuracy.

[0004] Therefore, the main problem with existing technologies for power system estimation is low estimation accuracy. Summary of the Invention

[0005] To address the aforementioned technical problem of low estimation accuracy, this invention discloses a power state estimation method and system based on an improved graph isomorphic network.

[0006] In a first aspect, the present invention discloses a power state estimation method based on an improved graph isomorphic network, comprising: Acquire raw measurement data and topology information of the target power system; A graph structure is constructed based on the topological information, and the original measurement data is converted into the node features and edge features of the graph structure to obtain the graph structure data. The graph-structured data is input into a pre-defined deep GINE (Graph Isomorphism Network with Edgefeatures) model for feature extraction, message passing, and residual aggregation, and the target state estimate is output.

[0007] Beneficial effects: The method of this invention inputs the power grid topology and edge features into a deep GINE model for in-depth processing, which can better fit the nonlinear physical relationships of the power system. Among them, residual aggregation makes it possible to train a deeper and more stable network. The deep network has stronger expressive power and noise resistance, and can provide better fault tolerance or robustness to partial measurement loss or noise, which can improve the estimation accuracy at the final estimation level.

[0008] Preferably, the deep GINE model includes an input layer, multiple stacked residual blocks, and an output layer; in, The input layer is used to extract features from the graph structure data to obtain the first feature; The residual block consists of a GINE convolutional layer and residual connection paths; the GINE convolutional layer is used to aggregate node features and edge features to obtain convolutional features; the residual connection paths are used to add the convolutional features to obtain the second feature. The output layer is used to receive the second feature of the last residual block output and map it to the target state estimate of the corresponding node.

[0009] Preferably, the residual block further includes a normalization layer and an activation function layer; The expression for the processing procedure of the residual block is:

[0010] In the formula, Indicates the first Input of each residual block; Represent edge features; This represents the GINE convolution operation; Presentation layer normalization operation; This represents the activation function. Indicates the first The output of each residual block.

[0011] Preferably, the topology information includes at least the buses and branches in the target power system; the raw measurement data includes the voltage amplitude measurement, active load and reactive load of each bus, and the active power measurement and reactive power measurement of each branch.

[0012] Preferably, a graph structure is constructed based on the topological information, and the original measurement data is converted into the node features and edge features of the graph structure, specifically as follows: The busbars in the target power system are defined as nodes of the graph structure, and the branches in the target power system are defined as edges of the graph structure. Extract the voltage amplitude measurement, active load and reactive load of each bus as the node characteristics of the corresponding node; Extract the active power and reactive power measurements from each branch and use them as edge features for the corresponding edges.

[0013] Preferably, the deep GINE model is trained by minimizing the loss function.

[0014] The preferred loss function to be minimized is as follows:

[0015] In the formula, Indicates the first The true value of each measurement item Indicates the first The predicted value of each measurement item, Indicates the first The weight of each measurement item, Represents the physical constraint loss term. Represents the balance coefficient. This indicates the total number of measurement items. This represents the loss value.

[0016] Preferably, before obtaining the graph structure data, the method of the present invention further includes: The physical quantities corresponding to node features and edge features are standardized.

[0017] Preferably, the target state estimate includes node load and inter-node transmission power.

[0018] In a second aspect, the present invention discloses a power state estimation system based on an improved graph isomorphic network, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the power state estimation method based on the improved graph isomorphic network described in the first aspect is implemented.

[0019] The beneficial effects of this invention are as follows: (1) Compared with the prior art, the method of the present invention has higher estimation accuracy and stronger robustness.

[0020] (2) Compared with the prior art, the method of the present invention has better generalization ability and topology adaptability.

[0021] (3) Compared with the prior art, the method of the present invention combines data-driven and model-driven approaches, guides the model to learn results that conform to physical laws, and improves the rationality and reliability of the estimation results. Attached Figure Description

[0022] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This is a flowchart of the power state estimation method based on an improved graph isomorphic network in Embodiment 1 of the present invention; Figure 2 This is an architecture diagram of the deep GINE model in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the power state estimation system based on an improved graph isomorphic network in Embodiment 2 of the present invention. Detailed Implementation

[0023] 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, not all, of the embodiments of the present invention. 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.

[0024] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] Example 1 like Figure 1 As shown, this invention discloses a power state estimation method based on an improved graph isomorphic network, comprising: S10: Obtain the raw measurement data and topology information of the target power system.

[0026] In this embodiment, the aforementioned raw measurement data mainly refers to the voltage amplitude of nodes, active / reactive loads, and active / reactive power flow on branches (edges). Topology information is mainly used to clarify how various components (such as buses, transmission lines, and transformers) in the target power system are connected, and it includes bus information and node information.

[0027] S20: Construct a graph structure based on the topology information, and convert the original measurement data into the node features and edge features of the graph structure to obtain the graph structure data.

[0028] Specifically, step S20, concerning the construction of the graph structure, involves: S21: Define the busbars in the target power system as nodes of the graph structure, and define the branches in the target power system as edges of the graph structure.

[0029] At this point, the graph structure obtained based on step S21 above can be regarded as a static network topology used to describe the structural layout of the target power system.

[0030] Based on step S21, dynamic data representation also needs to be incorporated into the graph structure, specifically: S22: Extract the voltage amplitude measurement, active load and reactive load of each bus as the node characteristics of the corresponding node.

[0031] Specifically, each bus in the target power system is defined as a graph node. The initial feature vector of each node includes the measured voltage amplitude, active load, and reactive load of that node. These features directly reflect the local operating state of the node.

[0032] S23: Extract the active power and reactive power measurements from each branch and use them as edge features for the corresponding edge.

[0033] Specifically, the edges of the graph are constructed based on the physical connectivity of the target power system (defined by a topology file similar to edge.csv, containing source and target nodes). The feature vector associated with each edge includes the active and reactive power measurements on that branch. Step S23 directly transforms the physical connectivity and power flow information of the target power system into the topology and edge features of the graph, providing rich contextual information for subsequent deep GINE models.

[0034] Through the above steps S21-S23, a bridge connecting the physical power grid and the deep GINE model is constructed. Its core purpose is to convert the original data into graph structure data that the deep GINE model can process.

[0035] Preferably, before obtaining the graph structure data, the method in this embodiment also needs to perform the following steps: S200: Standardize the physical quantities corresponding to node features and edge features.

[0036] It should be explained that step S200 above belongs to the data preprocessing process. Among them, per-unit value is a relatively important data representation method in power system analysis. It unifies physical quantities with different dimensions and magnitudes into dimensionless relative values ​​through a simple ratio, which can greatly simplify the calculation and analysis process, eliminate the numerical differences caused by different physical dimensions, thereby accelerating model training and improving stability.

[0037] S30: Input the graph structure data into the preset deep GINE model for feature extraction, message passing and residual aggregation, and output the target state estimate.

[0038] like Figure 2As shown, the method in this embodiment provides a deep GINE model, which includes an input layer, multiple stacked residual blocks, and an output layer.

[0039] The input layer extracts features from the graph-structured data to obtain the first feature. The residual block consists of a GINE convolutional layer and residual connection paths; the GINE convolutional layer aggregates node and edge features to obtain convolutional features; the residual connection paths sum the convolutional features to obtain the second feature. The output layer receives the second feature output from the last residual block and maps it to the target state estimate of the corresponding node.

[0040] Regarding the GINE convolutional layer, it's important to note that as the core feature extractor, it's a variant of the Graph Isomorphic Network (GIN), its core advantage lying in its ability to explicitly handle edge features. During message passing, it not only aggregates the state information of neighboring nodes but also incorporates the power features (active or reactive power) along the connecting edges into the node's feature update. This message passing mechanism allows the model to more accurately model the mutual influence between nodes established through line power. More specifically, in power systems, the "edges" of a graph represent not only logical connections but also actual transmission lines, carrying active and reactive power measurements that contain key physical constraints reflecting the grid's operating state. Ordinary graph algorithms often simplify these edges to weights or use them merely as paths, leading to the loss of significant high-dimensional physical information. The GINE model, through the aforementioned message passing mechanism, can explicitly and non-linearly fuse edge features (line power data) with the feature vectors of neighboring nodes when aggregating neighbor node information. This approach has profound physical implications: it simulates the strong coupling between "node voltage state" and "line power flow distribution" in a power system—that is, the state of a node depends not only on its neighboring nodes but also directly on the magnitude and direction of power transmitted on the connecting lines. Through this mechanism, the model effectively embeds the physical laws of the power grid (such as the constraints implicit in Kirchhoff's laws) into the feature extraction process, thereby more sensitively capturing minute state disturbances caused by fluctuations in renewable energy sources or changes in line load, significantly improving the physical consistency and accuracy of state estimation. Therefore, this mechanism is extremely important for accurately estimating the power grid state.

[0041] Regarding residual connection paths, it's worth noting that to build effective deep networks and address gradient vanishing or network degradation issues, residual connections can be introduced between GINE convolutional layers. Specifically, a standard residual block design is used, with several GINE convolutional layers in each residual block. The input to the residual block is directly added to the output through a "short-circuit path." This transforms the entire mapping relationship into H(x,w_h)+x, where H(x,w_h) is the residual mapping learned by the GINE convolutional layers. This structure allows for direct backpropagation of gradients, greatly alleviating the training challenges of deep networks and making it possible to build deeper networks to improve model expressive power and receptive field.

[0042] Based on the above technical solutions, the deep GINE model provided in this embodiment also implements multi-layer stacking and feature abstraction. By stacking multiple "GINE + residual" modules, the network can abstract features layer by layer. Among them, the shallow network mainly captures local neighbor information (such as directly connected lines and nodes), while the deep network can aggregate information from a wider range of neighbors, thereby capturing the complex long-range dependencies and nonlinear power flow effects in the power system.

[0043] Furthermore, the aforementioned residual block also includes a normalization layer and an activation function layer; Combining the above normalization layer and activation function layer, the processing expression for the residual block is as follows:

[0044] In the formula, Indicates the first Input of each residual block; Represent edge features; This represents the GINE convolution operation; Presentation layer normalization operation; This represents the activation function. Indicates the first The output of each residual block.

[0045] It should be explained that the normalization layer mentioned above is used to stabilize the training process, accelerate convergence, and mitigate gradient vanishing / exploding. The activation function layer, on the other hand, introduces nonlinear analysis, enabling the network to fit complex functions and thus enhance its expressive power.

[0046] Preferably, the activation function mentioned above can be ELU or SELU to improve the stability of the model training process.

[0047] Furthermore, the aforementioned deep GINE model undergoes a pre-training process, which primarily involves training by minimizing the loss function.

[0048] The aforementioned loss function to be minimized is specifically:

[0049] In the formula, Indicates the first The true value of each measurement item Indicates the first The predicted value of each measurement item, Indicates the first The weight of each measurement item, Represents the physical constraint loss term. Represents the balance coefficient. This indicates the total number of measurement items. This represents the loss value.

[0050] It should be explained that the aforementioned loss function is used to guide the training direction of the model. The loss function, which incorporates physical information constraints (such as the weighted mean square error loss function) and combines it with physical information loss based on power flow equation constraints, minimizes the weighted residual between the state estimate and the measurement value, thereby encouraging the estimation results to conform to the physical laws of the power system. In this embodiment, the physical constraint loss term is used to constrain the target state estimate to satisfy Kirchhoff's laws or nonlinear power flow equations.

[0051] For example, the formula for calculating the above physical constraint loss term can be:

[0052] In the formula, This represents the total number of nodes. Indicates the first The active power of each node (corresponding to the measurement item). Indicates the first The injected reactive power of each node, Indicates the first The flow from the node to the first The active power of the line at each node Indicates the first The flow from the node to the first The reactive power of the line at each node, Indicates and The set of all neighboring nodes connected to a given node. The physical constraint loss term calculated the difference between the sum of injected and outflowing power at each node; ideally, this difference should be zero. By minimizing the sum of squares of this difference, the loss function forces the learned state variables (voltage magnitude and phase angle) to satisfy the active and reactive power balance of the entire network.

[0053] For the output layer, it is essentially a fully connected layer that maps the node embeddings output by the last GINE convolutional layer to two target state estimates for each node, namely the node load (including active and reactive loads) and the inter-node transmission power (including active and reactive power).

[0054] It should be further explained that after training, the deep GINE model, when inputting graph structure data, can rely on the graph structure data to perform data simulation of each node and edge. Its output target state estimate can be equivalent to a prediction value or a real-time value of a newly connected node / edge.

[0055] Through the above steps S10-S30, the method of the present invention has at least the following technical advantages: Firstly, it offers higher estimation accuracy and stronger robustness. The method in this embodiment utilizes the power grid topology and edge features (power measurements) through a deep GINE model, enabling a better fit to the nonlinear physical relationships of the power system. Its residual connections make it possible to train deeper, more stable networks, while deep networks have stronger expressive power and noise resistance, exhibiting good fault tolerance for missing measurements or noise.

[0056] Secondly, it has better generalization ability and topology adaptability. The above model learns the message passing mechanism on the graph structure rather than the mapping of fixed nodes. Therefore, it can better generalize to different power grid topologies (such as the network after line switching), providing an effective solution for power grid state estimation with a high proportion of new energy sources.

[0057] Third, by incorporating physical constraints such as power flow equations into the loss function, and combining data-driven and model-driven approaches, the model is guided to learn results that conform to physical laws, thereby improving the rationality and reliability of the estimation results.

[0058] For example, based on the above technical solution, this embodiment selects the IEEE 118-node standard test system to simulate a regional power grid with a high proportion of distributed energy resources.

[0059] 1. Graph Structure Construction Phase For topology construction: 118 buses are defined as graph nodes, and 186 transmission lines are defined as graph edges.

[0060] Node characteristics: Select the measured values ​​of bus voltage amplitude, active load and reactive load collected by the SCADA system.

[0061] Side characteristics: Select the active and reactive power flow measurements at both ends of the line.

[0062] Preprocessing: Standardize all physical quantities to unify their dimensions.

[0063] Sample generation: Gaussian white noise (simulating measurement error) and random distributed photovoltaic power output fluctuations (simulating new energy uncertainty) are superimposed on the baseline power flow data to generate 10,000 sets of "measurement-state" pairs as the training set.

[0064] 2. Model Configuration and Training Construct a deep GINE model containing 8 stacked residual blocks: Residual block structure: Each residual block contains a GINE convolutional layer, a LayerNorm normalization layer, and a ReLU activation function, and sets residual connections (Shortcuts) to prevent gradient vanishing.

[0065] Loss function: The output result is forced to satisfy the nonlinear power flow equation (i.e. Kirchhoff's laws) by minimizing the loss function.

[0066] Training process: End-to-end training is performed using the Adam optimizer until the mixed loss converges.

[0067] 3. Effect Description and Verification To verify the superiority of this solution in dealing with significant fluctuations in new energy sources and the lack of measurement data, the following experiment was conducted: Test scenario: Assume that the output of the wind farm connected to node 45 in the system suddenly drops by 50%, and the line measurement data between nodes 45 and 49 is lost.

[0068] Based on the operation process of Examples 1-3 above, the following verification results can be obtained: Thanks to GINE's explicit utilization of power flow information from surrounding lines, it can infer values ​​from the edge features of neighboring nodes even in cases of local measurement loss. The final output strictly satisfies physical constraints, with a voltage amplitude MAE of only 0.4%, achieving a 77% improvement in accuracy compared to ordinary graph convolutional networks. Furthermore, its computation time is only one-tenth that of the traditional weighted least squares method, enabling millisecond-level real-time response.

[0069] Example 2 like Figure 3 As shown, this embodiment discloses a power state estimation system based on an improved graph isomorphic network, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the power state estimation method based on the improved graph isomorphic network described in Embodiment 1 is implemented.

[0070] The system in this embodiment also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art, and therefore will not be described in detail here.

[0071] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions that can be stored or otherwise maintained by such a computer-readable medium.

[0072] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0073] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A power state estimation method based on an improved graph isomorphic network, characterized in that, include: Acquire raw measurement data and topology information of the target power system; A graph structure is constructed based on the topology information, and the original measurement data is converted into node features and edge features of the graph structure to obtain graph structure data. The graph structure data is input into a preset deep GINE model for feature extraction, message passing, and residual aggregation, and the target state estimate is output.

2. The power state estimation method based on improved graph isomorphic networks according to claim 1, characterized in that, The deep GINE model includes an input layer, multiple stacked residual blocks, and an output layer; in, The input layer is used to extract features from the graph structure data to obtain a first feature; The residual block includes a GINE convolutional layer and a residual connection path; the GINE convolutional layer is used to aggregate the node features and the edge features to obtain convolutional features; the residual connection path is used to add the convolutional features to obtain a second feature; The output layer is used to receive the second feature output by the last residual block and map it to the target state estimate of the corresponding node.

3. The power state estimation method based on improved graph isomorphic networks according to claim 2, characterized in that, The residual block also includes a normalization layer and an activation function layer; The processing expression for the residual block is as follows: In the formula, Indicates the first Input of each residual block; Represent edge features; This represents the GINE convolution operation; Presentation layer normalization operation; This represents the activation function. Indicates the first The output of each residual block.

4. The power state estimation method based on improved graph isomorphic networks according to claim 1, characterized in that, The topology information includes at least the buses and branches in the target power system; the raw measurement data includes the voltage amplitude measurement, active load and reactive load of each bus, and also includes the active power measurement and reactive power measurement of each branch.

5. The power state estimation method based on improved graph isomorphic networks according to claim 4, characterized in that, A graph structure is constructed based on the topological information, and the original measurement data is converted into node features and edge features of the graph structure, specifically as follows: The busbars in the target power system are defined as nodes of the graph structure, and the branches in the target power system are defined as edges of the graph structure. Extract the voltage amplitude measurement, active load and reactive load of each bus as the node characteristics of the corresponding node; Extract the active power and reactive power measurements from each branch and use them as edge features for the corresponding edges.

6. The power state estimation method based on improved graph isomorphic networks according to claim 1, characterized in that, The deep GINE model is trained by minimizing the loss function.

7. The power state estimation method based on improved graph isomorphic networks according to claim 6, characterized in that, The loss function to be minimized is specifically: In the formula, Indicates the first The true value of each measurement item Indicates the first The predicted value of each measurement item, Indicates the first The weight of each measurement item, Represents the physical constraint loss term. Represents the balance coefficient. This indicates the total number of measurement items. This represents the loss value.

8. The power state estimation method based on improved graph isomorphic networks according to claim 1, characterized in that, Before obtaining the graph structure data, the method further includes: The physical quantities corresponding to the node features and edge features are normalized.

9. The power state estimation method based on improved graph isomorphic networks according to claim 1, characterized in that, The target state estimate includes node load and inter-node transmission power.

10. A power state estimation system based on an improved graph isomorphic network, characterized in that, It includes a processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the power state estimation method based on the improved graph isomorphic network as described in any one of claims 1-9.

Images