CN-122020059-A - Power system fault positioning method based on space-time causal graph network

CN122020059ACN 122020059 ACN122020059 ACN 122020059ACN-122020059-A

Abstract

The invention discloses a power system fault positioning method based on a space-time causal graph network, which relates to the field of power, and comprises the following steps of step 1, preprocessing data, cleaning, time synchronization, interpolation and standardization of multi-node time sequence data from a PMU (Power management unit) so as to ensure the integrity and consistency of input data; compared with the prior art, the intelligent power grid diagnosis method has the beneficial effects that nonlinear statistical detection, deep learning space-time feature extraction, graph structure causal learning and spectrogram theoretical quantization sequencing are creatively combined through the progressive analysis flow of data preprocessing, fault detection and variable isolation, space-time causal graph network modeling and causal balance root cause positioning, so that core pain points of the existing method in the aspects of sensitivity, accuracy, interpretability and adaptability to complex systems are systematically solved, and a brand-new and reliable solution is provided for the efficient and safe fault diagnosis of the intelligent power grid.

Inventors

JIN ZIYAN
He Zhenzheng
SHI WENJUAN
HU QIJIN
WANG SHUAI
LI SIYUAN
YU SHI
BAO JIN
YIN KE
SUN BOWEN

Assignees

国网江西省电力有限公司信息通信分公司

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (8)

1. The power system fault positioning method based on the space-time causal graph network is characterized by comprising the following steps of: step 1, data preprocessing, namely cleaning, time synchronization, interpolation and standardization are carried out on multi-node time sequence data from a PMU (power management unit) so as to ensure the integrity and consistency of input data; step 2, fault detection and variable isolation, performing early abnormality detection on the preprocessed data based on a kernel principal component analysis model, and screening out a candidate abnormal variable set; Step 3, network modeling of a space-time causal graph, namely extracting multi-layer time sequence features in a time dimension by utilizing time sequence data of a candidate abnormal variable set, and learning and constructing a causal graph structure among variables in the space dimension, so that multi-stage propagation modeling of faults among multiple nodes is realized, and a multi-stage prediction contribution matrix P of a quantized causal relation is output; and 4, causal balance root positioning, namely based on the multi-stage prediction contribution matrix P, iteratively solving disturbance source nodes with the most influence in the system by quantifying causal contribution relation among nodes, and finally outputting root positioning results, propagation paths and variable importance sequencing.
2. The method for locating faults in an electrical power system based on a spatio-temporal causal graph network according to claim 1, wherein the preprocessing of time series data from PMUs in step 1 is as follows: Step 11, time synchronization, namely aligning all node data according to a uniform sampling period according to time stamps of PMU equipment of different measuring points, and ensuring time sequence continuity; step 12, feature derivation and standardization, namely calculating derived features through voltage and current phasors, and carrying out normalization and standardization treatment on all the features; and step 13, segmenting the continuous data stream by adopting a sliding window mechanism so as to keep the time evolution characteristic and adapt to the sequence input requirement of a subsequent model.
3. The method for positioning faults of an electric power system based on a space-time causal graph network according to claim 1, wherein the step 2 specifically comprises the steps of detecting faults, calculating T2 and SPE statistics in real time based on a nuclear principal component analysis model, and comparing the statistics with a preset control limit; step 22, variable isolation, namely calculating the contribution value of each original variable to the statistic overrun through contribution analysis, screening out the variable with the highest contribution degree, and forming a candidate abnormal variable set Wherein For a preset contribution threshold value, Is the first in the time series variable The number of variables that can be used, Actual first Actual values of the individual predicted contributions; Wherein the kernel principal component analysis is performed by a kernel function Implicit mapping of data to high-dimensional space to efficiently capture nonlinear anomaly characteristics, where 、 Respectively is 、 Is used for the non-linear mapping of (a).
4. The method for locating a fault in an electrical power system based on a spatio-temporal causal graph network according to claim 1, wherein step 3 comprises: Step 31, extracting time features, namely extracting multi-scale time features from time sequence data of each variable in a candidate abnormal variable set by adopting a one-dimensional convolution gating circulating unit structure; step 32, constructing a causal graph and modeling multi-stage propagation, namely constructing a learnable sparse adjacency matrix to represent direct causal relation among variables, and realizing cross-level transmission of information among the variables by stacking multi-layer graph convolution layers so as to simulate the multi-stage propagation process of faults; And 33, causal path screening and contribution quantification, namely introducing a hierarchical sparse pruning mechanism to sparse the causal graph, automatically screening a key causal path, and generating a multi-stage prediction contribution matrix P for quantifying the total causal influence among variables by accumulating multi-layer propagation effects.
5. The method for locating faults of an electric power system based on a space-time causal graph network according to claim 4, wherein in step 31, a one-dimensional convolution gating circulation unit model is adopted to extract multi-level time sequence features from time sequences of variables in a candidate abnormal variable set, the model firstly carries out local feature extraction through a plurality of parallel one-dimensional convolution neural networks, a first layer uses 32 convolution kernels with the width of 3 and adopts a ReLU activation function to extract a basic time sequence mode, a second layer uses 64 convolution kernels with the width of 3 to further excavate deep time features, and convolution operations all adopt the same filling to keep the sequence length unchanged; Then, building a GRU network to capture long-term time dependence and converting convolution characteristics into characteristic vectors rich in time sequence information, wherein the core calculation process of the GRU is as follows: And (3) gating signal calculation: ; ; State update mechanism: ; ; Wherein, the To update the gate, controlling the degree of retention of the historical state; Determining a forgetting proportion of a historical state for resetting the gate; as candidate state, fusing the current input and the history information after screening, The weight matrices for the update gate, reset gate and candidate states respectively, Bias vectors for the update gate, reset gate and candidate states respectively, As a function of the non-linear activation, For the hidden state output of the current time step, For inputting a time series.
6. The method for locating faults in an electrical power system based on a space-time causal graph network of claim 4, wherein in step 32, causal graph construction and multistage propagation modeling are performed by learning causal relationships between variables and simulating hierarchical propagation of faults, firstly constructing a learnable sparse adjacency matrix A for characterizing direct predictive relationships between variables, then stacking L graph information fusion layers to realize cross-layer transfer and aggregation of node information in the causal graph, and setting a first layer The layer input features are The output characteristics are calculated as follows: ; Wherein, the Represent the first The input characteristics of the layer(s), For the corresponding weight matrix to be used, Is the adjacency matrix after normalization processing, Is that Is used for the degree matrix of the (c), , Is a nonlinear activation function; To capture indirect causal relationships across multiple nodes, modules stack L information fusion layers, a The calculated general formula of the layer is: ; through multi-layer propagation, the information of the root cause variable is transmitted to the downstream variable through the intermediate node to form a complete multi-stage causal chain; on the basis of determining a network architecture, a joint optimization target for constructing a time sequence capable of accurately predicting and simultaneously learning a causal graph structure with physical significance is provided, and multi-element time sequence prediction and graph structure learning are unified under a joint optimization framework: ; wherein the parameter settings satisfy Wherein And The L1 and L2 penalty factors of the sparse canonical selection constraint method respectively, Adjacency matrices of layers 1 to L are used for representing the causal relationship intensity between different layers, Is the norm of the L1 matrix, The method is a Frobenius matrix norm, and in order to avoid that node relations in a graph structure are too isolated and information loss caused by L1 regularization, minimum spanning tree constraint is introduced: ; In the formula, Representing edges in a minimum spanning tree Selected among the minimum spanning trees.
7. The method for locating faults in an electrical power system based on a space-time causal graph network as claimed in claim 4, wherein in step 33, causal path screening and contribution quantization are performed by introducing a hierarchical sparse pruning mechanism for improving the interpretability of the causal graph and focusing on critical propagation paths, and the hierarchical sparse pruning mechanism performs sparsification processing on the adjacency matrix A by introducing elastic network constraints, and simultaneously, adopting one of the following methods Pruning function, selectively strengthening important connection and weakening unimportant connection in training process of deep network to obtain matrix after processing Through L layers of information propagation, a multi-stage prediction contribution matrix P is constructed to quantify the total causal effect among variables: ; ; In the matrix Is the final prediction contribution matrix and, Is that Is the first of (2) Row of lines Column elements.
8. The method for locating faults in an electrical system based on a spatio-temporal causal graph network according to any of claims 1 to 7, characterized in that in step 4, for each variable, based on a multi-level prediction contribution matrix P Calculating a root cause score , wherein, Score of As non-negative real numbers, for measuring the amount of predictive information provided by a variable in a system, the higher the score, the more likely the variable is the root cause; Based on the nature of the root cause variables, the score calculation follows three key principles: Direct contribution principle, variables The greater the predicted contribution to other variables, i.e The greater the value, the score The higher should be, this reflects the direct impact of the variables; Principle of indirect influence if the variables For variable With predicted contribution, then Should be matched with Positive correlation; part of the prediction information of (a) originates from High scoring Indicating that the important information is derived from ; Normalization principle-root cause scores of all variables should constitute a probability distribution, i.e Ensuring the interpretability and comparability of the score; Establishing a balance equation for predicting information transfer according to the principle: ; Wherein the product term Has definite physical meaning and represents a variable Directional variable The amount of predictive information transferred, which takes into account the variables at the same time The multi-layer information flow propagation and quantification mechanism completely reflects two core principles of root cause assessment, namely, not only considering the direct prediction contribution of the variable, but also including indirect contribution generated by influencing other important variables; The balance equation is expressed as a matrix form, and can be obtained: ; Wherein the method comprises the steps of Is a column vector consisting of all variable root cause scores; Based on the calculation result, the algorithm automatically selects the variable with the highest score as the system root cause: 。

Description

Power system fault positioning method based on space-time causal graph network Technical Field The invention relates to the field of electric power, in particular to an electric power system fault positioning method based on a space-time causal graph network. Background With the rapid development of smart grids, the power system structure is increasingly complex, the operation uncertainty is obviously increased, and the safe and stable operation of the smart grid faces serious challenges. Disturbance and fault propagation in a power system have obvious space-time coupling and multistage conduction characteristics, and if a fault source cannot be positioned timely and accurately, large-range voltage drop, frequency instability and even cascading power failure accidents can be caused. Therefore, the rapid and accurate fault positioning is realized, and the method is important to prevent the expansion of faults, shorten the power failure time and improve the power supply reliability. The current fault diagnosis method mainly has the following limitations: 1. Although the methods based on protection actions and knowledge reasoning (such as expert systems, petri networks and the like) are logically intuitive, the methods are highly dependent on the correctness and completeness of protection and circuit breaker action signals. In an actual system, protection against action, malfunction or communication interruption occurs, so that the diagnosis accuracy of the method is drastically reduced, and it is difficult to describe a dynamic propagation path and a multistage influence process of a fault on electric quantity. 2. Anomaly detection is performed by measuring statistical features of data based on methods of multivariate statistics and contribution analysis (e.g., principal Component Analysis (PCA), independent Component Analysis (ICA), etc.). However, most of the methods are static or quasi-static models, the dynamic evolution of faults along with time is difficult to describe, and downstream monitoring variables which are obviously affected by the faults are easily misjudged as sources, so that the positioning accuracy is limited. At the same time, it lacks the ability to quantify multi-level, indirect causal chains. 3. Methods based on traditional causal analysis (e.g., gland cause and effect, transfer entropy, bayesian networks, etc.) attempt to construct inter-variable causal graphs to reveal propagation paths. However, feedback control loops such as Automatic Voltage Regulators (AVRs) and Power System Stabilizers (PSS) commonly existing in a power system can introduce cyclic causality, interfere with the judgment of such methods, and cause blurring or even error of root cause identification. In addition, the traditional method has the defects of modeling capability, quantization precision and result interpretability when processing the causal relationship of indirect and dynamic coupling in a complex system. In summary, when the existing method is used for coping with a fault of a complex power system, the existing method generally has the problems of insufficient early detection sensitivity, easy downstream interference in positioning, incapability of quantitatively describing a multi-stage propagation path, poor adaptability to a feedback link, poor interpretability and the like, and needs to be improved. Disclosure of Invention The invention aims to provide a power system fault positioning method based on a space-time causal graph network, which aims to solve the problems in the background technology. In order to achieve the above purpose, the present invention provides the following technical solutions: a power system fault locating method based on a space-time causal graph network comprises the following steps: Step 1, data preprocessing, namely cleaning, time synchronization, interpolation and standardization are carried out on multi-node time sequence data (voltage, current, power, frequency and the like) from a PMU (phasor measurement unit) so as to ensure the integrity and consistency of input data; step 2, fault detection and variable isolation, performing early abnormality detection on the preprocessed data based on a kernel principal component analysis model, and screening out a candidate abnormal variable set; Step 3, network modeling of a space-time causal graph, namely extracting multi-layer time sequence features in a time dimension by utilizing time sequence data of a candidate abnormal variable set, and learning and constructing a causal graph structure among variables in the space dimension, so that multi-stage propagation modeling of faults among multiple nodes is realized, and a multi-stage prediction contribution matrix P of a quantized causal relation is output; and 4, causal balance root positioning, namely based on the multi-stage prediction contribution matrix P, iteratively solving disturbance source nodes with the most influence in the system by quantifying causal contribution r