CN-121980970-A - Online power grid voltage stability margin prediction method based on graph structure

CN121980970ACN 121980970 ACN121980970 ACN 121980970ACN-121980970-A

Abstract

The invention discloses a grid voltage stability margin on-line prediction method based on a graph structure, which comprises the following steps of constructing a grid attribute graph model in a graph database, establishing an event driving mechanism to realize incremental update of topology change, dynamically generating a graph database query statement based on a parameterized graph query template, acquiring a node characteristic matrix and an adjacent matrix at a prediction moment, converting the node characteristic matrix and the adjacent matrix into a sparse tensor format, constructing a space-time graph convolution network prediction model, driving the space-time graph convolution network prediction model through incremental update of the topology change driven by the event driving mechanism, and writing a prediction result and a model intermediate layer node embedded vector back to the graph database as attributes. The invention realizes millisecond-level online prediction of the voltage stability margin of the power grid, and meets the requirements of real-time performance, accuracy and adaptability of online safety analysis of a power system.

Inventors

YANG LEIHONG
ZHOU YAN
ZHANG CHEN
WU JING

Assignees

浙江创邻科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260403

Claims (10)

1. The grid voltage stability margin on-line prediction method based on the graph structure is characterized by comprising the following steps of: storing physical equipment, connection relation and real-time measurement data of a power grid in a graph database in the form of an attribute graph, constructing a power grid attribute graph model, and establishing an event-driven mechanism to realize incremental update of topology change; Dynamically generating a graph database query statement based on a parameterized graph query template, acquiring a node characteristic matrix and an adjacent matrix at a prediction moment, and converting the node characteristic matrix and the adjacent matrix into a sparse tensor format; Constructing a space-time diagram convolutional network prediction model, wherein the space-time diagram convolutional network prediction model comprises an encoder and a decoder, and the encoder comprises a plurality of stacked space-time convolutional blocks; Driving the space-time diagram convolutional network prediction model through incremental update of topology change driven by the event driving mechanism, inputting the node characteristic matrix and the dynamic adjacency matrix into the space-time diagram convolutional network prediction model, extracting high-order space-time characteristics through the stacked space-time convolutional blocks, and layering and outputting a node-level voltage stability index and a system-level voltage stability margin through the decoder; and writing the prediction result and the model middle layer node embedded vector back to the graph database as attributes.
2. The grid voltage stability margin on-line prediction method based on graph structure as claimed in claim 1, wherein, Nodes in the power grid attribute graph model represent buses, generators and loads, and the attributes of the nodes comprise static parameters and dynamic time sequence measurement; Edges in the power grid attribute graph model represent a power transmission line and a transformer, and the attributes of the power grid attribute graph model comprise impedance, admittance and a switch state; the adjacency matrix is determined and weighted by the connection relation of the edges.
3. The grid voltage stability margin on-line prediction method based on graph structure as claimed in claim 1, wherein, The event driving mechanism is that when the power grid is subjected to switching deflection and equipment switching, the graph database completes incremental updating under the guarantee of a transaction, and synchronously triggers the data updating flow of the prediction model.
4. The grid voltage stability margin on-line prediction method based on graph structure as claimed in claim 1, wherein, And dynamically generating a query statement by the parameterized graph query template according to the prediction task range, the time window T and the physical quantity type, and acquiring the node characteristic matrix and the dynamic adjacency matrix of the node through the index-free adjacency characteristic of the graph database.
5. The grid voltage stability margin on-line prediction method based on graph structure as claimed in claim 1, wherein, The space-time convolution block comprises a space-diagram convolution layer, a time convolution layer and a gating fusion mechanism; The space diagram convolution adopts Chebyshev polynomial approximation or self-adaptive diagram convolution, and a formula can be expressed as follows: , Wherein, the Is a normalized graph Laplace matrix, the meaning of Z (l) is the first layer of spatial convolution output, k is the order of a Chebyshev polynomial, T k is the Chebyshev polynomial, θ k is a learnable parameter, σ is an activation function, and X （l-1） is the output of the last layer; The time convolution layer captures time sequence dependence along a time dimension by adopting a one-dimensional convolution or gating circulating unit; The gating fusion mechanism dynamically adjusts the fusion proportion g of the spatial feature Z s and the temporal feature Z t by calculating fusion gating, and outputs a final feature H: , , wherein [ ] represents a splice, For element-by-element multiplication.
6. The method for online prediction of grid voltage stability margin based on graph structure according to claim 5, wherein, The space diagram convolution layer adopts a self-adaptive adjacent matrix fusing physical topology and self-adaptive learning, and specifically comprises the following steps: Introducing a learnable adaptive adjacency matrix A adapt by taking a physical topology adjacency matrix A phy from a graph database as an initial value; In the model training process, the parameters of A adapt and other parameters of the space-time diagram convolution network prediction model are updated and optimized together, and an optimization formula is as follows: , Wherein alpha and beta are weight coefficients which can be learned, the matrix multiplication is represented, Z is an output characteristic matrix of a space convolution layer, X is an input node characteristic matrix, Is a matrix of weight parameters that can be learned, σ is an activation function.
7. The grid voltage stability margin on-line prediction method based on graph structure as claimed in claim 1, wherein, The decoder layered output includes: the node layer is used for mapping and outputting the node-level voltage stability index through the full-connection layer; A system layer for aggregating global features through an attention mechanism to output the system level voltage stability margin; the system layer introduces a multi-headed attention mechanism.
8. The grid voltage stability margin on-line prediction method based on graph structure as claimed in claim 1, wherein, The grid voltage stability margin online prediction method based on the graph structure further comprises the steps that when partial topology change occurs, an affected electric influence domain is calculated through an incremental graph analysis algorithm to serve as a minimum partial sub-graph, the space-time graph convolution network prediction model only carries out model reasoning and parameter reloading on the minimum partial sub-graph, and when a prediction task only aims at the minimum partial sub-graph, a scheduling system only carries out reasoning on model parameter slices of nodes related to the minimum partial sub-graph.
9. The method for online prediction of grid voltage stability margin based on graph structure of claim 8, wherein, And the calculation of the electrical influence domain adopts a method of combining quick estimation based on the power flow sensitivity or impedance matrix with graph traversal, and extends outwards by taking the change point as the center until the electrical distance exceeds a node of a threshold epsilon to form a minimum local subgraph which needs to be recalculated at the time.
10. The grid voltage stability margin on-line prediction method based on graph structure as claimed in claim 1, wherein, The on-line prediction method for the voltage stability margin of the power grid based on the graph structure further comprises the following steps: On-line fine tuning training, when a brand new topology mode is identified based on historical features in a graph database, K historical snapshots which are most similar to the current new topology are retrieved from the graph database based on graph structure similarity measurement; and in the offline training stage, model parameters are pre-trained by adopting a meta-learning strategy so as to provide a parameter initial point for online fine adjustment.

Description

Online power grid voltage stability margin prediction method based on graph structure Technical Field The invention belongs to the technical field of power system data processing based on a graph database, and particularly relates to an online power grid voltage stability margin prediction method based on a graph structure. Background Voltage stabilization is an important indicator of safe and stable operation of the power system. The voltage stability margin (Voltage Stability Margin, VSM) reflects the safety distance from voltage collapse of the power grid in the current running state, and is a key basis for dispatching operators to evaluate the safety level of the system and formulate a preventive control strategy. The traditional voltage stability margin analysis method mainly depends on numerical simulation calculation. The continuous power flow method (Continuation Power Flow, CPF) tracks the PV curve of the system by stepping up the load and solving the power flow equation, thereby determining the voltage collapse point and the stability margin. The method has the inherent defects of long calculation time consumption and difficulty in meeting the real-time requirement of on-line monitoring. For large-scale interconnected power grids, single continuous power flow calculation usually needs several minutes or even tens of minutes, the running state of the power grid is changed instantaneously, decisions based on offline calculation results often lag behind the actual system evolution, and effective support cannot be provided for real-time safety early warning. In recent years, with the development of artificial intelligence technology, models such as an artificial neural network (ARTIFICIAL NEURAL NETWORK, ANN), a support vector machine (Support Vector Machine, SVM) and the like are used to establish a mapping relationship between power grid operation characteristics and a voltage stability margin. The method utilizes the historical operation data to carry out offline training, and can rapidly obtain the prediction result only by forward reasoning when in online application, thereby remarkably improving the calculation speed. However, existing machine learning methods typically treat the grid data as independent eigenvectors in euclidean space, ignoring the critical a priori knowledge of the topological connection relationships between the grid elements. When the power grid is subjected to topological change due to faults, overhauls and the like, the dimension or the semantics of the input features of the model are changed, so that the prediction accuracy is greatly reduced, even the power grid is completely invalid, and the generalization capability is weaker. In addition, most of such models are in a 'black box' structure, the prediction results lack the interpretability, and visual decision support is difficult to provide for the dispatcher. The space-time diagram convolutional network (Spatio-Temporal Graph Convolutional Network, ST-GCN) is a combination of a diagram neural network and a time-series modeling technique, capable of processing diagram structure data and time-series data simultaneously. However, existing implementations of ST-GCN are disjoint from the underlying data storage system. The traditional architecture mostly adopts a relational database (such as Oracle and MySQL) to store power grid topology and measurement data, and when millisecond topology updating and massive real-time data are faced, delay of multi-table connection query is high, data conversion is complex, and an end-to-end low-delay flow from data acquisition to model reasoning is difficult to support. In addition, the ST-GCN model usually holds a fixed, predefined adjacency matrix, cannot sense dynamic changes of the power grid topology, and when the actual topology is inconsistent with the training topology, the model input and the structure are mismatched, so that the prediction reliability is difficult to guarantee. The Graph Database is used as a Database system for storing and managing Graph structure data in a native mode, and a node-side attribute Graph model is adopted, so that the associated query and path traversal can be executed efficiently, and the Graph Database is suitable for flexible management of power grid topology. However, in the prior art, the graph database is mainly used as a "storage and query" engine, complex graph calculation (such as spectrum analysis and feature extraction) still needs to be completed externally, the calculation and transportation amount is large, and the calculation instantaneity is not high. Disclosure of Invention The invention provides an online power grid voltage stability margin prediction method based on a graph structure, which solves the technical problems, and specifically adopts the following technical scheme: an online prediction method for a grid voltage stability margin based on a graph structure comprises the following steps: storing physical equipment, conne