CN-122020605-A - Power distribution network running state deduction method, risk assessment method, equipment and medium

CN122020605ACN 122020605 ACN122020605 ACN 122020605ACN-122020605-A

Abstract

The invention discloses a power distribution network running state deduction method, a risk assessment method, equipment and a medium, wherein the state deduction method comprises the steps of obtaining electrical parameters of each node of a power distribution network in the current moment and the past preset time period, constructing model input of the current network topology structure and the target candidate network topology structure of the power distribution network, inputting the model into a pre-trained power distribution network running state deduction model, and deducting and outputting a predicted voltage amplitude sequence of each node in a future prediction time domain. Compared with the existing power distribution network running state deduction based on a fixed topological structure, the method can realize the power distribution network running state deduction under a cross-scene and a cross-topology condition, and can provide effective data support for power distribution network running risk assessment after topology switching. The multi-step voltage track can be obtained through one-time forward deduction, and a more efficient analysis way is provided for obtaining future operation voltage safety information.

Inventors

XU LIN
CHEN LIZHI
LUO LI
WU SHIFAN
ZHANG YA
QIN BING
ZHENG JUNFENG
LIU XIAO
HE DIAN
ZENG BIN
TANG MENGXIAN
HE YAOQI

Assignees

国网湖南省电力有限公司长沙供电分公司
国网湖南省电力有限公司
国家电网有限公司

Dates

Publication Date: 20260512
Application Date: 20260115

Claims (13)

1. The power distribution network running state deduction method is characterized by comprising the following steps of: Acquiring electrical parameters of each node of the power distribution network at the current moment and in the past preset time period, and constructing model input by the current network topology structure and the target candidate network topology structure of the power distribution network; inputting the model into a pre-trained power distribution network running state deduction model, and deducting and outputting a predicted voltage amplitude sequence of each node in a future prediction time domain; The power distribution network running state deduction model comprises a feature extraction module, a feature fusion module and a non-autoregressive output module, wherein the feature extraction module comprises three parallel feature extraction branches based on a graph attention network, a residual error network and a gate control circulation unit network, the feature fusion module is used for carrying out feature fusion on output features of the three feature extraction branches, taking fusion features of a fusion feature matrix at the last moment to carry out graph attention alignment on a target candidate network topological structure, and the non-autoregressive output module carries out voltage deduction of each node in a future prediction domain based on a graph attention alignment result.
2. The method for deriving operating states of a power distribution network according to claim 1, wherein the current network topology Ã 1 and the target candidate network topology Ã 2 of the power distribution network are calculated by symmetric normalized adjacency matrices according to the following formula: ; In the formula, For the original adjacency matrix of the network topology of the distribution network, I is a unit array, D is Q is 1 or 2, and corresponds to the current network topology and the target candidate network topology of the power distribution network respectively.
3. The power distribution network operation state deduction method according to claim 1, wherein the graph attention encoding is performed on each moment of model input on the current network topology Ã 1 of the power distribution network based on the feature extraction branch of the graph attention network, and the calculation formula is represented as follows: ; ; ; Wherein h gat,(t,i) is the characteristic of the node i of the current characteristic extraction branch at the time t, x (t,i) is the input characteristic of the node i at the time t, W (k) is the linear transformation matrix of the kth attention head, a k is the attention vector of the head, and K is the total number of the attention heads; representing a vector concatenation operation; representing the score of the kth attention head to the feature, N 1 (i) is the neighborhood of node i; the characteristics of all nodes at each moment are aggregated into a graph meaning network output characteristic H gat .
4. The power distribution network operation state deduction method according to claim 1, wherein, based on the feature extraction branch of the residual network, for each node, two one-dimensional convolution layers with the convolution kernel length of 3 are stacked along the time dimension of the historical feature sequence, and linear residual connection is introduced, specifically expressed as follows: ; In the formula, h res,(t,i) is the characteristic of the node i of the current characteristic extraction branch at the time t; a sequence of features is entered for the history of node i, As an input feature of node i at time t, Representing a two-layer one-dimensional convolution stack along time, In the case of a non-linear activation, Representing to perform feature dimension mapping; the characteristics of all nodes at each moment are aggregated into a residual network output characteristic H res .
5. The power distribution network operation state deduction method according to claim 1, wherein the feature extraction is performed through a bidirectional GRU based on a feature extraction branch of a gating cycle unit network, specifically comprising the following steps: ; In the formula, h gru,(t,i) is the characteristic of the node i of the current characteristic extraction branch at the time t; Is in a forward hidden state; W o represents that the bidirectional stitching is subjected to characteristic dimension mapping; representing a vector concatenation operation; And aggregating the characteristics of all nodes at each moment into a gating cycle unit network output characteristic H gru .
6. The power distribution network running state deduction method according to claim 1, wherein the feature fusion module adopts an attention gating mechanism to score and normalize output features of three feature extraction branches to obtain self-adaptive fusion feature representations of each node at different moments, and the mathematical expression is as follows: ; ; In the formula, The fusion characteristic of the node i at the time t is obtained; Respectively representing the characteristics of a node i at a moment t, which is output by branch extraction based on three parallel characteristics of a graph attention network, a residual error network and a gating circulation unit network, wherein W f is a trainable weight matrix for mapping the spliced characteristics into three-dimensional scoring vectors, and b f is a corresponding offset; Pi (t,i) is a weight vector after attention normalization of three mapping scores, and the three components are all in (0, 1) and the sum is 1; The deduction is carried out on the target candidate network topology Ã 2 , the characteristic of the last moment H of the fusion characteristic matrix is taken to carry out drawing attention alignment on the target candidate network topology once, and the specific formula is expressed as follows: ; ; Wherein U i is the voltage of node i, and N 2 (i) is the neighborhood of node i under Ã 2 ; the fusion characteristic of the node i at the moment H is obtained, and W u 、b u is a trainable parameter.
7. The power distribution network operation state deduction method according to claim 1, wherein the non-autoregressive output module deducts the voltage of each node in the future prediction time based on the graph attention alignment result, and the voltage amplitude prediction formula of the node i in the future f-th hour is as follows: ; In the formula, U i is the voltage of the node i in the graph attention alignment result; Time position codes, and W c 、W p 、w o 、b c and c are trainable parameters.
8. The power distribution network operation state deduction method according to claim 1, wherein when training the power distribution network operation state deduction model, the following loss function is adopted: ; wherein L represents the total loss, The predicted voltage amplitude and the real voltage amplitude which are output by the power distribution network running state deduction model are respectively obtained by the U; And All weights, E </i > ] represents the average of the voltage out-of-limit penalties calculated for all samples in the current training batch, all times in the predicted time domain, and all nodes.
9. The power distribution network operation risk assessment method is characterized by comprising the following steps of: Obtaining a predicted voltage amplitude sequence of each node in a predicted time domain by adopting the power distribution network running state deduction method according to any one of claims 1 to 8; Calculating three risk indexes of voltage out-of-limit rate, system average voltage deviation and voltage fluctuation rate based on a predicted voltage amplitude sequence of each node in a prediction time domain; respectively determining subjective weights and objective weights of the three risk indexes by adopting an analytic hierarchy process and an entropy weight process, and then introducing a game theory to seek optimal balance between the subjective weights and the objective weights so as to obtain comprehensive weights of the three risk indexes; and (3) introducing a cloud model theory, and mapping the comprehensive weights of the three risk indexes into a comprehensive risk cloud containing expected entropy and super entropy, so as to realize visual grading and visual representation of the power distribution network risk.
10. The method for evaluating the running risk of a power distribution network according to claim 9, wherein when a game theory is introduced to seek the optimal balance between subjective weights and objective weights of three risk indexes, an objective function is set as follows: ; In the formula, And Subjective and objective weights for three risk indicators, respectively, lambda 1 and lambda 2 , respectively And Is used for the fusion coefficient of (a), ; And obtaining lambda 1 and lambda 2 by solving through a Lagrangian multiplier method, and finally obtaining the comprehensive weight of the three risk indexes as follows: ; In the formula, Representing the integrated weight of the three risk indicators.
11. The power distribution network operation risk assessment method according to claim 9, wherein mapping the comprehensive weights of three risk indexes into a comprehensive risk cloud containing expected-entropy-super entropy specifically comprises: The digital characteristic calculation formula of the comprehensive risk cloud is as follows: ; In the formula, 、 And Respectively representing expectations, entropy and super entropy in the comprehensive risk cloud; representing the composite weight of the first risk indicator, 、 And The expectation, entropy and super entropy of the first risk index are respectively represented.
12. An electronic device, comprising: A memory having a computer program stored thereon; a processor for loading and executing the computer program to implement the power distribution network operating state deduction method according to any one of claims 1 to 8 or the power distribution network operating risk assessment method according to any one of claims 9 to 11.
13. A computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the power distribution network operation state deduction method according to any one of claims 1 to 8 or the power distribution network operation risk assessment method according to any one of claims 9 to 11.

Description

Power distribution network running state deduction method, risk assessment method, equipment and medium Technical Field The invention relates to the field of analysis and optimization of power distribution networks of power systems, in particular to a power distribution network running state deduction method, a risk assessment method, equipment and a medium. Background In recent years, under the pushing of energy green low-carbon transformation strategy, a power distribution network is gradually transformed into an active power distribution system with high coupling of source network charge storage from a traditional radiation type power supply network. With the large-scale access of distributed photovoltaic, energy storage and electric vehicles, a power distribution system simultaneously faces the challenges of source and load double-side power fluctuation, namely, the midday photovoltaic output peak can exceed the local absorption capacity, the power flow of a power grid is induced to be reversed, the upper limit of terminal voltage is caused, and the electric vehicles are charged in a concentrated mode in a specific period to form a load peak, so that node voltage dip is caused, and even the lower limit is caused. Friendly interaction of source network charge storage is an effective regulation and control means for improving the operation flexibility of a power distribution network, but complexity and uncertainty of a system operation state are increased, so that risk early warning is necessary to be developed based on accurate space-time state deduction. Currently, a graph-time sequence depth fusion model has become a mainstream tool in the field of state deduction. In the space dimension, the graph annotation force network (graph attention network, GAT) can adaptively allocate weights according to the electrical connection relation among the nodes to effectively characterize the space coupling characteristic of the distribution network, and in the time dimension, the residual network (ResNet) and the gating cycle unit (gated recurrent units, GRU) are good at extracting the short-term fluctuation characteristic and the long-term evolution trend respectively. Most of the prior researches use the modules in series, and under the scene of severe fluctuation of source and load, the sensitive capture of power peaks and the stable maintenance of operation trends are difficult to be simultaneously considered. In addition, in the related research of network reconstruction, the existing optimization method often depends on an accurate physical model, has heavy calculation load, and is difficult to support cross-topology state deduction and online rapid risk assessment for different topology schemes. The existing power distribution network running state deduction and risk assessment mainly have the following defects: (1) The model has poor adaptability to topology change, the main stream space-time model is usually trained for a fixed topology structure, lacks a state transition mechanism crossing the topology, and is difficult to directly deduce the running state of the power distribution network after topology switching; (2) The prediction target is disjointed with the operation requirement, the existing deep learning model takes the prediction error minimization as the optimization target, and is directly related to the lack of safety indexes such as voltage out-of-limit rate, qualification rate and the like which are concerned by operators, so that the prediction accuracy is improved and the effective enhancement of risk recognition capability is not necessarily brought; (3) The risk quantification means is single, and the deterministic point prediction is difficult to describe a risk interval formed by superposition of multi-source uncertainties, and a comprehensive characterization method capable of fusing multiple types of operation indexes and realizing risk classification is lacked. Disclosure of Invention The invention provides a power distribution network running state deduction method, a risk assessment method, equipment and a medium, which are used for effectively solving the voltage deduction difficult problem oriented to risk assessment under multi-scene and multi-topology conditions. In a first aspect, a power distribution network operation state deduction method is provided, including the following steps: Acquiring electrical parameters of each node of the power distribution network at the current moment and in the past preset time period, and constructing model input by the current network topology structure and the target candidate network topology structure of the power distribution network; inputting the model into a pre-trained power distribution network running state deduction model, and deducting and outputting a predicted voltage amplitude sequence of each node in a future prediction time domain; The power distribution network running state deduction model comprises a feature extraction m