CN-121983946-A - Power system optimization method, device, medium and equipment based on physical constraint heterograph

CN121983946ACN 121983946 ACN121983946 ACN 121983946ACN-121983946-A

Abstract

The application discloses an electric power system optimization method, device, medium and equipment based on physical constraint heterograms. The method comprises the steps of constructing a grid abnormal graph based on a static topological connection relation of a power grid and power grid operation data in a power system, performing differential feature coding on different types of nodes and edges in the grid abnormal graph to determine node feature vectors and edge feature vectors, inputting the node feature vectors and the edge feature vectors into a multi-layer heterogeneous convolution model, performing neighborhood information aggregation on the node feature vectors and the edge feature vectors through the multi-layer heterogeneous convolution model to generate a target feature vector of a central node, performing global pooling on the target feature vector, outputting power grid state variables through a residual multi-layer perceptron, solving an inverse problem of a power grid power balance equation based on the power grid state variables, calculating control parameters of the power system, and adjusting the power system based on the control parameters. The generalization capability of the model to the power grid topology change can be improved, and the prediction result is ensured to strictly meet the physical safety constraint of the power grid.

Inventors

ZHOU LIANGCAI
XU FENG
XU WANXIN
Jiao Xingchao
LIU LINLIN
XU HAO
LUAN WEIJIE
ZHOU YI
CHEN XIN
GAO JIANING
CHEN RUI
WANG SHU

Assignees

国家电网有限公司华东分部
西安交通大学

Dates

Publication Date: 20260505
Application Date: 20251201

Claims (10)

1. A method for optimizing an electrical power system based on a physical constraint differential pattern, the method comprising: constructing a power grid heterogram based on the power grid static topological connection relation and the power grid operation data in the power system; Performing differential feature coding on nodes and edges of different types in the power grid heterograms, and determining node feature vectors and edge feature vectors, wherein the nodes comprise generator nodes and load nodes, and the edges comprise power transmission lines and load interconnection lines; Inputting the node feature vector and the edge feature vector into a multi-layer heterogeneous convolution model, and carrying out neighborhood information aggregation on the node feature vector and the edge feature vector through the multi-layer heterogeneous convolution model to generate a target feature vector of a central node, wherein the multi-layer heterogeneous convolution model is obtained by optimizing a loss function of physical constraint mixing; carrying out global pooling on the target feature vector, and outputting a power grid state variable through a residual multi-layer perceptron, wherein the power grid state variable at least comprises voltage amplitude values and voltage phase angles of all nodes; And solving an inverse problem of a power grid power balance equation based on the power grid state variables, calculating control parameters of the power system, and adjusting the power system based on the control parameters.
2. The method for optimizing a power system based on a physical constraint differential graph according to claim 1, wherein the differential feature coding of nodes and edges in the power grid differential graph comprises: And respectively carrying out characteristic projection on the node and the edge through linear transformation based on the linear coding layer corresponding to the node type and the edge type to obtain the node characteristic and the edge characteristic.
3. The power system optimization method based on physical constraint heterograms according to claim 1, wherein the multi-layer heterogeneous convolution model includes a plurality of heterogram attention layers, the neighborhood information aggregation is performed on the node feature vector and the edge feature vector, and a target feature vector of a central node is generated, and the method comprises: Combining a node feature vector of a central node output by a previous heterogram attention layer with a neighbor feature vector of a neighbor node to generate a node feature vector of the central node of the current heterogram attention layer, wherein the neighbor node is a first-order node connected with the central node; And if the current heterographing attention layer is the last heterographing attention layer of the multi-layer heterographing convolution model, taking a node feature vector of a central node of the current heterographing attention layer as the target feature vector.
4. A power system optimization method based on physical constraint heterograms according to claim 3, wherein the step of combining the node feature vector of the central node output by the previous heterogram attention layer with the neighbor feature vector of the neighbor node to generate the node feature vector of the central node of the current heterogram attention layer comprises: Splicing a node characteristic vector of a central node, a node characteristic vector of a neighbor node and an edge characteristic vector of a connecting edge under the attention layer of the previous iso-graph to form a target vector, wherein the connecting edge is an edge connecting the central node and the neighbor node; performing dot product on the target vector and the weight vector to generate attention weight of the target vector; And generating the node feature vector of the central node under the current heterograph attention layer based on the node feature vectors of all the neighbor nodes connected with the central node and the node feature vector of the central node by superposition of the attention weights.
5. The method for optimizing a power system based on a physical constraint heterogram according to claim 1, wherein the global pooling of the target feature vector and outputting a power grid state variable through a residual multi-layer perceptron comprises: global pooling is carried out on the target feature vector, and a global context vector is generated; And fusing the global context vector with the target feature vector through a residual multi-layer perceptron to obtain the power grid state variable.
6. The method for optimizing a power system based on physically constrained iso-patterning according to any one of claims 1 to 5, further comprising: Normalizing the voltage amplitude to a preset safety range by a Sigmoid function and/or, And constraining the voltage phase angle to a preset phase angle range through a Tanh function.
7. The method for optimizing a power system based on physically constrained iso-patterning of any one of claims 1 to 5, The loss function comprises a mean square error term between a predicted value and a reference solution and a penalty term violating inequality constraints, wherein the inequality constraints comprise a generator output limit value, a voltage amplitude safety boundary and a line transmission capacity limit.
8. An electrical power system optimization apparatus based on physical constraint heterograms, the apparatus comprising: The construction module is used for constructing a power grid abnormal pattern based on the power grid static topological connection relation and the power grid operation data in the power system; the processing module is used for carrying out differential feature coding on different types of nodes and edges in the power grid abnormal pattern, determining node feature vectors and edge feature vectors, wherein the nodes comprise generator nodes and load nodes, the edges comprise power transmission lines and load interconnection lines, and Inputting the node feature vector and the edge feature vector into a multi-layer heterogeneous convolution model, carrying out neighborhood information aggregation on the node feature vector and the edge feature vector through the multi-layer heterogeneous convolution model to generate a target feature vector of a central node, wherein the multi-layer heterogeneous convolution model is obtained by optimizing a loss function of physical constraint mixing, and Carrying out global pooling on the target feature vector, and outputting a power grid state variable through a residual multi-layer perceptron, wherein the power grid state variable at least comprises voltage amplitude values and voltage phase angles of all nodes; and the optimization module is used for solving the inverse problem of the power balance equation of the power grid based on the power grid state variable, calculating the control parameter of the power system and adjusting the power system based on the control parameter.
9. A readable storage medium having stored thereon a program or instructions, which when executed by a processor, implements the power system optimization method based on physical constraint iso-patterning of any one of claims 1 to 7.
10. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the power system optimization method based on physically constrained iso-patterning as claimed in any one of claims 1 to 7 when executing the program.

Description

Power system optimization method, device, medium and equipment based on physical constraint heterograph Technical Field The application relates to the technical field of power systems, in particular to a power system optimization method, device, medium and equipment based on physical constraint heterograms. Background Currently, the Optimal Power Flow (OPF) problem of a power system is mainly solved by a numerical optimization algorithm, and typically includes interior point method (Interior Point Methods, IPMs) and sequential quadratic programming (Sequential Quadratic Programming, SQP). Such methods require processing highly nonlinear physical constraints through multiple rounds of iterations. In a medium-scale power grid (e.g., an IEEE 118 node system), a single OPF calculation by a conventional MATPOWER solver takes about 124 milliseconds. In the face of a high-proportion renewable energy source access scene, the system needs to solve thousands of working conditions in real time under + -10% load fluctuation, the total time consumption exceeds 10 minutes, and the second-level scheduling response requirement cannot be met. In the prior art, although the calculation speed is improved through a machine learning scheme, the method has fundamental limitations that firstly, a power grid adjacent matrix is flattened into a feature vector, connection characteristics between nodes are ignored, and secondly, a pure data driving model is not embedded into a power balance equation, and a prediction result is easy to violate power grid safety constraint. Therefore, the real-time performance, the safety and the topology generalization capability cannot be considered, and the intelligent optimization process of the high-volatility power grid is restricted. Disclosure of Invention In view of the above, the present application provides a method, apparatus, medium and device for optimizing an electric power system based on a physical constraint iso-graph, so as to solve the problems set forth in the related art. According to one aspect of the present application, there is provided a power system optimization method based on physical constraint heterograms, comprising: constructing a power grid heterogram based on the power grid static topological connection relation and the power grid operation data in the power system; Performing differential feature coding on nodes and edges of different types in the power grid heterograms, and determining node feature vectors and edge feature vectors, wherein the nodes comprise generator nodes and load nodes, and the edges comprise power transmission lines and load interconnection lines; Inputting the node feature vector and the edge feature vector into a multi-layer heterogeneous convolution model, and carrying out neighborhood information aggregation on the node feature vector and the edge feature vector through the multi-layer heterogeneous convolution model to generate a target feature vector of a central node, wherein the multi-layer heterogeneous convolution model is obtained by optimizing a loss function of physical constraint mixing; carrying out global pooling on the target feature vector, and outputting a power grid state variable through a residual multi-layer perceptron, wherein the power grid state variable at least comprises voltage amplitude values and voltage phase angles of all nodes; And solving an inverse problem of a power grid power balance equation based on the power grid state variables, calculating control parameters of the power system, and adjusting the power system based on the control parameters. Optionally, the differential feature coding of the nodes and edges in the grid heterograph includes: And respectively carrying out characteristic projection on the node and the edge through linear transformation based on the linear coding layer corresponding to the node type and the edge type to obtain the node characteristic and the edge characteristic. Optionally, the multi-layer heterogeneous convolution model includes a plurality of heterographic attention layers, and the neighborhood information aggregation is performed on the node feature vector and the edge feature vector to generate a target feature vector of a central node, including: Combining a node feature vector of a central node output by a previous heterogram attention layer with a neighbor feature vector of a neighbor node to generate a node feature vector of the central node of the current heterogram attention layer, wherein the neighbor node is a first-order node connected with the central node; And if the current heterographing attention layer is the last heterographing attention layer of the multi-layer heterographing convolution model, taking a node feature vector of a central node of the current heterographing attention layer as the target feature vector. Optionally, the combining the node feature vector of the central node output by the previous iso-composition attention layer with the neighb