A multi-source data fusion power distribution network fault accurate positioning method and system

Abstract

The application discloses a kind of multi-source data fusion's distribution network fault accurate positioning method and system, comprising: obtaining quasi-steady state measurement data and dynamic synchronous phasor data, construct dynamic graph network model, quasi-steady state measurement data and dynamic synchronous phasor data are input dynamic graph network model and are handled, obtain the causal effect after deduction, construct dynamic weight self-adapting fusion model, quasi-steady state measurement data and dynamic synchronous phasor data are input dynamic weight self-adapting fusion model and are handled, obtain space-time convolution operator, obtain distribution network fault position based on the causal effect after deduction and space-time convolution operator.The application constructs dynamic graph network model and dynamic weight self-adapting fusion model, fuses quasi-steady state measurement data and dynamic synchronous phasor data, can more accurately capture the dynamic characteristics and fault propagation path of distribution network, improve the precision of fault location.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent power grid fault diagnosis, and particularly relates to a multi-source data fusion power distribution network fault accurate positioning method and system. BACKGROUND

[0002] With the in-depth promotion of new power system construction, as an important link connecting the power supply side and the user side, the operation safety and intelligent level of the power distribution network have become the key indicators for measuring the stability and resilience of modern power systems. In recent years, the large-scale access of distributed new energy (such as photovoltaic and wind power), the dynamic changes of multiple types of loads, the rapid development of energy storage devices, and the widespread application of electric vehicles have made the power distribution network present highly complex, dynamic and multi-source heterogeneous characteristics. This complex evolution trend has posed unprecedented challenges to the fault monitoring and diagnosis capability of the power distribution network.

[0003] Traditional power distribution network fault positioning methods are mostly based on steady-state models, current sudden change detection, waveform identification or rule-based reasoning. Although they have certain practicality in scenarios with stable structure, homogeneous data and controllable operation conditions, their positioning accuracy and response efficiency are difficult to meet the high-standard operation requirements of intelligent power distribution networks in the face of current multi-source heterogeneous data frequent interaction, network structure dynamic reconstruction, nonlinear uncertainty of fault propagation path, etc.

[0004] At the same time, the popularity of intelligent sensing technology has promoted the widespread deployment of multi-source measurement terminals such as SCADA, PMU and AMI. The operation data of the power distribution network presents high-dimensional, heterogeneous and non-synchronous characteristics. How to effectively fuse multiple types of data and extract key fault features has become a hot and difficult point in current research. In addition, with the rise of new generation intelligent computing methods such as artificial intelligence, deep learning and graph neural network, a new theoretical basis and technical path are provided for fault modeling and causal inference of complex systems. However, most existing models focus on structure correlation modeling and lack deep description of the dynamic structure and fault causal mechanism of the power distribution network. In practical applications, there are still problems such as weak generalization ability and poor interpretability.

[0005] Therefore, the development of an intelligent fault positioning method with spatiotemporal modeling capability, causal reasoning capability and multi-source fusion capability not only is the core support for the digital transformation of the power distribution network, but also is a key technical breakthrough for realizing the whole-process closed-loop control of "active sensing - rapid response - accurate positioning". It has important practical significance and strategic value for improving the safe operation level of the power grid and promoting the development of intelligent power grids. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a multi-source data fusion power distribution network fault accurate positioning method and system, which aims to solve the problems in the background art.

[0007] To achieve the above object, the application provides the following technical scheme: a multi-source data fusion power distribution network fault accurate positioning method, comprising:

[0008] Step S1: obtaining quasi-steady state measurement data and dynamic synchronous phasor data;

[0009] Step S2: constructing a dynamic graph network model, inputting the quasi-steady state measurement data and dynamic synchronous phasor data into the dynamic graph network model for processing to obtain deduced causal effects;

[0010] Step S3: constructing a dynamic weight adaptive fusion model, inputting the quasi-steady state measurement data and dynamic synchronous phasor data into the dynamic weight adaptive fusion model for processing to obtain a space-time convolution operator;

[0011] Step S4: obtaining the power distribution network fault location based on the deduced causal effects and the space-time convolution operator.

[0012] Further, the processing process of the dynamic graph network model is:

[0013] Based on the quasi-steady state measurement data and the dynamic synchronous phasor data, the time-varying power distribution network is defined as a four-tuple ; wherein, represents the differential manifold space at the time, the differential manifold space at the time is constructed based on the quasi-steady state measurement data; represents the node feature field at the time, which is constructed based on the quasi-steady state measurement data and the dynamic synchronous phasor data; represents the edge connection degree tensor field at the time; represents the causal propagation form at the time;

[0014] The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, which is represented as:

[0015] ;

[0016] In the formula, represents the rate of change of the node feature field at the t time; represents the covariant derivative; is a manifold curvature adjustment factor; ^ represents an outer product operator; is a dual operator;

[0017] The node neighborhood is determined by the geodesic distance function ; and respectively represent the node i and the node j at the Time-differentiated manifold space The geometric position in the middle; Adaptive neighborhood radius;

[0018] Adaptive neighborhood radius, indicating:

[0019] ;

[0020] In the formula, Indicates the minimum value that meets the conditions; Indicates the p i Ball with center and radius r; Indicates the volume element; Indicates the threshold value;

[0021] The first Time edge connection degree tensor field weight is calculated by curvature coupling mechanism, indicating:

[0022] ;

[0023] In the formula, Indicates the edge connection degree tensor field weight of the attention score between node And node At time t; Indicates the geodesic path of node And node ; Is the action of the Riemann curvature tensor in the direction of the geodesic path tangent vector, reflecting the bending degree of the geodesic path in the geometric sense; Indicates the geodesic path; Indicates the geodesic path element length;

[0024] Based on the joint analysis of dynamic graph attention mechanism and causal intervention reasoning to construct information propagation chain, the joint analysis of causal intervention reasoning is composed of characteristic transformation equation, node neighborhood , adaptive neighborhood radius and edge connection degree tensor field weight; The information propagation chain is constrained and screened to form the fault propagation path F ， .

[0025] Further, in the four-tuple , the index theorem of Atiyah-Singer is used to process And , calculate the index of differential operator , establish the corresponding relationship between the topological invariant of the index of differential operator And the fault propagation path F ， , indicating:

[0026] ;

[0027] ;

[0028] ;

[0029] In the formula, Describing differential operators Indicators; Indicates the first Nodal feature field at time step Chen's characteristics; This represents the outer product operator; Indicates the first Differential manifold space at time of Topology class; For differential operators The indicators and fault propagation path F ， The correspondence; This represents the i-th path fault node on the fault propagation path; For the fault propagation path and Edge; For the fault propagation path Edge; This is the nth path failure node on the fault propagation path; For the first in the fault propagation path One path failure node;

[0030] based on and F ， The established correspondence will link the fault propagation path F ， Formalized as a dynamic causal network model, it can be represented as:

[0031] ;

[0032] In the formula, Indicates the first Time-of-flight dynamic causal network model; Indicates the first The set of nodes in a time-varying dynamic causal network model; Indicates the first The set of edges in a time-varying dynamic causal network model; Indicates the first Time-node feature matrix; Indicates the first The set of constraints for a time-varying dynamic causal network model;

[0033] A dynamic causal network model is used to capture the evolution of the distribution network topology. The topology includes voltage, current, and power, and is represented as follows:

[0034] ;

[0035] where, is the spatio-temporal attention weight of node and node at time t; denotes the learnable transpose weight vector; is the learnable weight matrix; and are the state features of voltage and current of node i and node j at time t, respectively; denotes the power state feature of node at time t; is the activation function; is the set of neighboring nodes of node i at time t;

[0036] the spatio-temporal attention weight of node and node at time t, can be further decomposed as

[0037] ;

[0038] ;

[0039] ;

[0040] where, denotes the state feature concatenation vector of voltage and current of node i and node j at time t; denotes the attention score between node i and node j at time t; is the attention score of node i and node at time t;

[0041] based on the spatio-temporal attention weight of node and node at time t, a causal graph structure is constructed, which represents the causal relationship between node and node in the power distribution network, the causal relationship between node and node is learned through a dynamic graph attention mechanism, and a path of the causal relationship is obtained; based on the path of the causal relationship and the spatio-temporal attention weight of node and node at time t, an intervention on node j is simulated; denotes the intervention operator; denotes the node Intervention variables;

[0042] Based on the simulation of intervention at node j Define causal effect ,express:

[0043] ;

[0044] In the formula, Represents a node The causal effect; This represents the expected value operator; Indicates the state of node i;

[0045] Introducing the parent node set and exogenous variables from the structural equation model, we can represent:

[0046] ;

[0047] In the formula, Represents a node The structural equation function; For nodes The set of parent nodes; Let i be an exogenous variable of node i;

[0048] By introducing the parent node set and exogenous variables from the structural equation model, we can further derive the causal effect, which is expressed as:

[0049] ;

[0050] In the formula, This represents the derived causal effect.

[0051] Furthermore, the processing procedure of the dynamic weight adaptive fusion model is as follows:

[0052] Quasi-steady-state measurement data and dynamic synchronous phasor data are divided into the k-th type of data, and the k-th type of data is in the... The feature matrix of time nodes is defined as follows ; For the real number field; Let be the number of nodes at time t; Let k be the feature dimension of the k-th class of data;

[0053] The dynamic fusion weights are derived from the network state entropy and data quality metrics, and are expressed as follows:

[0054] ;

[0055] In the formula, This indicates that the weights of the k-th data source are dynamically fused at time t. Indicates the first the data quality indicator of the kth data source at the tth time; denotes the network state entropy at the tth time; is a weight parameter of the network state entropy at the tth time; is a weight parameter of the data quality indicator of the kth data source at the tth time; the data quality indicator of the kth data source at the tth time; is the total number of classes of the kth data;

[0056] the network state entropy at the tth time, denoted by:

[0057] ;

[0058] wherein, denotes the probability of node v in the failure state; is a logarithmic function; wherein, ; denotes the degree of node v; is the degree of node u;

[0059] the data quality indicator of the kth data source at the tth time, denoted by:

[0060] ;

[0061] wherein, denotes the real node feature matrix corresponding to the kth data source at the tth time; denotes the Frobenius norm of the real node feature matrix;

[0062] the fused spatio-temporal feature tensor is generated by and , denoted by:

[0063] ;

[0064] wherein, is the fused spatio-temporal feature tensor; denotes the concatenation operation of the feature dimension; denotes a nonlinear activation function; denotes the projection matrix of the kth data source node feature matrix.

[0065] Further, the causal enhanced spatio-temporal propagation equation of the fused spatio-temporal feature tensor is defined, denoted by:

[0066] ;

[0067] wherein, denotes the fused spatio-temporal feature tensor of the l+1th layer; denotes a time convolution operation; is an activation function; denotes the dynamic attention coefficient of the fused spatio-temporal feature tensor of node j and node i at the t-th moment; denotes the fused spatio-temporal feature tensor of node j at the t-th moment in the l-th layer; denotes the intervention operation on ; denotes a gradient operator; denotes an estimated value of a causal effect; is a causal enhancement coefficient; is a set of neighbor nodes of node i;

[0068] The dynamic attention coefficient of the fused spatio-temporal feature tensor of node j and node i at the t-th moment is calculated, denoted as:

[0069] ;

[0070] In the formula, denotes the fused spatio-temporal feature tensor of node i at the t-th moment in the l-th layer; denotes a difference vector of the fused spatio-temporal feature tensor of node i and node j at the t-th moment; is a set of neighbor nodes of node i and node the fused spatio-temporal feature tensor at the t-th moment; is a concatenation operation;

[0071] The estimated value of the causal effect is calculated, denoted as:

[0072] ;

[0073] In the formula, denotes an expected value of a target variable of node x obtained after intervention on node x at the t-th moment; denotes an expected value of a target variable of node obtained after no intervention on node at the t-th moment; denotes a loss function with respect to ; denotes the fused spatio-temporal feature tensor of node v at the t-th moment in the l-th layer; denotes a node feature matrix of node v at the t-th moment; denotes a partial derivative with respect to ; is a partial derivative;

[0074] The spatio-temporal convolution operator fuses the bidirectional gated recurrent unit and the spatio-temporal propagation equation of causal enhancement through a TConv unit, denoted as:

[0075] ;

[0076] In the formula, The spatio-temporal convolution operation is performed on the input to obtain a spatio-temporal convolution operator; Bidirectional gated recurrent unit is represented; is the inverse square root of the t-th time degree matrix dynamically generated based on ; is the t-th time adjacency matrix dynamically generated based on ; Element-wise multiplication is represented; Long short-term memory network is represented; is a weight matrix of ; is a spatio-temporal feature tensor.

[0077] Further, the specific process of step S4 is as follows:

[0078] Obtain meteorological data, use a dynamic time warping algorithm to perform spatio-temporal alignment on quasi-steady state measurement data, dynamic synchronous phasor data and meteorological data, and construct a multi-source tensor ; is the time dimension; N is the number of nodes; and D is the feature dimension;

[0079] In view of the asynchronous sampling characteristics of the meteorological data in the multi-source tensor, the following interpolation function is designed, which is represented as:

[0080] ;

[0081] In the formula, represents the meteorological data after interpolation processing; represents a normalization factor of an interpolation time window; represents the length of a time window ending at the t-th time; represents a meteorological influence decay kernel function; represents the value of the initial meteorological data at time S; is an integral variable;

[0082] The meteorological data after interpolation processing and the spatio-temporal convolution operator construct a dynamic graph structure , the dynamic graph structure and are fused to define the following:

[0083] ;

[0084] In the formula, represents a fusion function; represents a fault propagation path inference function; represents a composite operation of functions; represents a fusion operation; represents the dynamic fusion weight of the kth data source at the tth moment; represents a graph convolution network-long short-term memory network;

[0085] the fusion function is taken as the final comprehensive spatio-temporal feature tensor, and based on the final comprehensive spatio-temporal feature tensor and a fault propagation causal graph is constructed, and an average causal effect ACE is calculated through the fault propagation causal graph, the fault propagation causal graph including fault nodes , and representing:

[0086] ;

[0087] in the formula, represents a result variable; is a result variable; represents an intervention operation on the node , so that the node is in a normal state ; represents an intervention operation on the fault node , so that the fault node is in a fault state ;

[0088] when |ACE|>delta, delta is a set threshold value; it is considered that the fault node has a causal effect on the result variable ; the average causal effect ACE values of all fault nodes having a causal effect are sorted, and the fault node having a causal effect with the largest ACE value after sorting is selected as the fault location of the power distribution network.

[0089] A multi-source data fusion power distribution network fault accurate positioning system for implementing a multi-source data fusion power distribution network fault accurate positioning method, comprising:

[0090] an acquisition module for acquiring quasi-steady state measurement data and dynamic synchronous phasor data;

[0091] a causal processing module for constructing a dynamic graph network model, inputting the quasi-steady state measurement data and the dynamic synchronous phasor data into the dynamic graph network model for processing, and obtaining a deduced causal effect;

[0092] A fusion module is configured to construct a dynamic weight adaptive fusion model, input the quasi-steady state measurement data and dynamic synchronous phasor data into the dynamic weight adaptive fusion model for processing, and obtain a space-time convolution operator;

[0093] A fault positioning module is configured to obtain a fault position of the power distribution network based on the derived causal effect and the space-time convolution operator.

[0094] Compared with the prior art, the present application has the following beneficial effects: the present application can more accurately capture the dynamic characteristics and fault propagation path of the power distribution network by constructing a dynamic graph network model and a dynamic weight adaptive fusion model, fusing quasi-steady state measurement data and dynamic synchronous phasor data, and introducing a causal reasoning mechanism. Compared with the traditional method, the accuracy and response speed of fault positioning are significantly improved, especially in a complex and variable power distribution network environment, the fault position can be quickly locked, and the fault troubleshooting time is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0095] Figure 1 The method flowchart of the present application.

[0096] Figure 2 The accuracy comparison chart of different fault positioning methods of the present application. DETAILED DESCRIPTION

[0097] As shown in the drawings, Figure 1 the present application provides a technical solution: a multi-source data fusion power distribution network fault accurate positioning method, comprising:

[0098] Step S1: obtaining quasi-steady state measurement data and dynamic synchronous phasor data;

[0099] Step S2: constructing a dynamic graph network model, inputting the quasi-steady state measurement data and dynamic synchronous phasor data into the dynamic graph network model for processing, and obtaining a derived causal effect;

[0100] Step S3: constructing a dynamic weight adaptive fusion model, inputting the quasi-steady state measurement data and dynamic synchronous phasor data into the dynamic weight adaptive fusion model for processing, and obtaining a space-time convolution operator;

[0101] Step S4: obtaining a fault position of the power distribution network based on the derived causal effect and the space-time convolution operator.

[0102] The specific process of step S2 is as follows:

[0103] Based on the quasi-steady state measurement data and dynamic synchronous phasor data, the time-varying power distribution network is defined as a four-tuple ; wherein, represents the differential manifold space at the moment, and the The time differential manifold space is constructed based on quasi-steady state measurement data; The time differential manifold space is constructed based on quasi-steady state measurement data; The time differential manifold space is constructed based on quasi-steady state measurement data; The time differential manifold space is constructed based on quasi-steady state measurement data; The time differential manifold space is constructed based on quasi-steady state measurement data; The time differential manifold space is constructed based on quasi-steady state measurement data; The time differential manifold space is constructed based on quasi-steady state measurement data; The time differential manifold space is constructed based on quasi-steady state measurement data;

[0104] The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as:

[0105] The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as:

[0106] The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as:

[0107] The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as:

[0108] The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as:

[0109] The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as:

[0110] The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as:

[0111] The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: i The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as: The evolution of the dynamic graph network model follows the characteristic transformation equation on the four-tuple manifold, and is represented as:

[0112] No. The edge connectivity tensor field weights at each time step are calculated using a curvature coupling mechanism, and are expressed as follows:

[0113] ;

[0114] In the formula, Represents the node at time t. and nodes The edge connectivity tensor field weights between attention scores; Represents a node and nodes Geodetic path; This represents the action of the Riemann curvature tensor in the direction of the tangent vector of the geodesic path, reflecting the degree of curvature of the geodesic path in a geometric sense. Indicates the geodetic path; Indicates the length of a geodesic path element;

[0115] An information propagation chain is constructed based on the dynamic graph attention mechanism and the joint analysis of causal intervention reasoning. The joint analysis of causal intervention reasoning consists of feature transformation equations and node neighborhoods. The system consists of an adaptive neighborhood radius and edge connectivity tensor field weights; it constrains and filters the information propagation chain to form a fault propagation path F. ， .

[0116] Among them, to enhance the fault propagation path F ， The explanatory power of the quadruple In the middle, the Atiyah-Singer index theorem is used to... and Process and calculate the differential operator. The index is used to establish the differential operator. Among the indicators, topological invariants and fault propagation paths F ， The correspondence is represented as follows:

[0117] ;

[0118] ;

[0119] ;

[0120] In the formula, Describing differential operators Indicators; Indicates the first Nodal feature field at time step Chen's characteristics; This represents the wedge product operator; Indicates the first Differential manifold space of time instant of topological class (Atiyah-Singer); index of differential operator and failure propagation path F ， correspondence; is the ith path failure node on failure propagation path F is the ith path failure node on failure propagation path F and edge on failure propagation path F is the ith path failure node on failure propagation path F is the ith path failure node on failure propagation path F is the ith path failure node on failure propagation path F is the ith path failure node on failure propagation path F is the ith path failure node on failure propagation path F

[0121] Based on the correspondence established by and F ， , failure propagation path F ， is formalized as a dynamic causal network model, represented as:

[0122] ;

[0123] wherein, represents the dynamic causal network model at the th time instant; represents the node set in the dynamic causal network model at the th time instant; represents the edge set in the dynamic causal network model at the th time instant; represents the node feature matrix at the th time instant; represents the constraint condition set of the dynamic causal network model at the th time instant;

[0124] The dynamic causal network model is adopted to capture the evolution law of the topology structure of the power distribution network, including voltage, current and power, represented as:

[0125] ;

[0126] wherein, is the spatio-temporal attention weight of node and node at the th time instant; represents the learnable transpose weight vector; is the learnable weight matrix; and The voltage and current characteristics of nodes i and j at time t are respectively. Represents the node at time t. Power state characteristics; Activation function for Leaky Rectified LinearUnit; Let be the set of neighboring nodes of node i at time t;

[0127] No. Time Node and nodes The spatiotemporal attention weights can be further decomposed into, representing:

[0128] ;

[0129] ;

[0130] ;

[0131] In the formula, This represents the concatenated vector of voltage and current characteristics of nodes i and j at time t. This represents the attention score between node i and node j at time t; Let node i and node t be the nodes at time t. Attention score;

[0132] Based on the Time Node and nodes The spatiotemporal attention weights are used to construct a causal graph structure, which represents the nodes in the distribution network. and nodes The causal relationships between nodes are learned through a dynamic graph attention mechanism. and nodes The causal relationship between them, and the path to obtain the causal relationship; the path based on the causal relationship and the first Time Node and nodes The spatiotemporal attention weights are used to simulate intervention on node j. ; Indicates the budget; Represents a node Intervention variables;

[0133] Based on the simulation of intervention at node j Define causal effect ,express:

[0134] ;

[0135] In the formula, Represents a node The causal effect; This represents the expected value operator; Indicates the state of node i;

[0136] Introducing the parent node set and exogenous variables from the structural equation model, we can represent:

[0137] ;

[0138] In the formula, Represents a node The structural equation function; For nodes The set of parent nodes; Let i be an exogenous variable of node i;

[0139] By introducing the parent node set and exogenous variables from the structural equation model, we can further derive the causal effect, which is expressed as:

[0140] ;

[0141] In the formula, This represents the derived causal effect.

[0142] The processing procedure for the dynamic weight adaptive fusion model is as follows:

[0143] Quasi-steady-state measurement data and dynamic synchronous phasor data are divided into the k-th type of data, and the k-th type of data is in the... The feature matrix of time nodes is defined as follows ; For the real number field; Let be the number of nodes at time t; Let k be the feature dimension of the k-th class of data;

[0144] The dynamic fusion weights are derived from the network state entropy and data quality metrics, and are expressed as follows:

[0145] ;

[0146] In the formula, This indicates that the weights of the k-th data source are dynamically fused at time t. Indicates the first Data quality metrics for a data source at time t; This represents the network state entropy at time t; The weight parameters are the network state entropy at time t; For the k-th type of data source, the weight parameters of the data quality index at time t are: Let k be the data quality index of the k-th data source at time t; a total number of categories of the kth data;

[0147] a network state entropy at the tth moment, denoted as:

[0148] ;

[0149] wherein, denotes a probability of the node v in a failure state; is a logarithmic function; wherein, ; denotes a degree of the node v; is a degree of the node u;

[0150] a data quality index of the kth data source at the tth moment, denoted as:

[0151] ;

[0152] wherein, denotes a real node feature matrix corresponding to the kth data source at the tth moment; denotes a Frobenius norm of the real node feature matrix;

[0153] the fused spatio-temporal feature tensor is generated by and , denoted as:

[0154] ;

[0155] wherein, is the fused spatio-temporal feature tensor; denotes a concatenation operation of feature dimensions; denotes a nonlinear activation function; denotes a projection matrix of the kth data source node feature matrix.

[0156] wherein, a causal enhanced spatio-temporal propagation equation of the fused spatio-temporal feature tensor is defined, denoted as:

[0157] ;

[0158] wherein, denotes the fused spatio-temporal feature tensor of the l+1th layer; denotes a time convolution operation; is an activation function; denotes a dynamic attention coefficient of the fused spatio-temporal feature tensor of the node j and the node i at the tth moment; denotes the fused spatio-temporal feature tensor of the node j at the tth moment of the lth layer; denotes an intervention operation on ; denotes a gradient operator; represents an estimated value of a causal effect; is a causal enhancement coefficient; is a set of neighborhood nodes of node i;

[0159] The fused spatiotemporal feature tensor dynamic attention coefficient calculation of node j and node i at the t-th moment is represented as:

[0160] ;

[0161] In the formula, represents the fused spatiotemporal feature tensor of node i at the t-th moment of the l-th layer; represents the fused spatiotemporal feature tensor difference vector of node i and node j at the t-th moment; is the fused spatiotemporal feature tensor difference vector of node i and node at the t-th moment; is a concatenation operation;

[0162] The estimated value calculation of a causal effect is represented as:

[0163] ;

[0164] In the formula, represents the expected value of the target variable of node x obtained after intervention on node x at the t-th moment; represents the expected value of the target variable of node obtained after no intervention on node at the t-th moment; represents the partial derivative of the loss function with respect to ; represents the fused spatiotemporal feature tensor of node v at the t-th moment of the l-th layer; represents the node feature matrix of node v at the t-th moment; represents the partial derivative of with respect to ; is the partial derivative;

[0165] The spatiotemporal convolution operator fuses the bidirectional gated recurrent unit and the spatiotemporal propagation equation of causal enhancement through a TConv unit, and is represented as:

[0166] ;

[0167] In the formula, represents the spatiotemporal convolution operation on the input to obtain the spatiotemporal convolution operator; represents the bidirectional gated recurrent unit; is based on The dynamically generated t-th time scale matrix inverse square root; is based on The dynamically generated t-th time adjacency matrix; Indicates element-by-element multiplication; Indicates a long short-term memory network; For The weight matrix of The space-time feature tensor.

[0168] The specific process of step S4 is:

[0169] The present application obtains meteorological data (such as temperature, humidity, wind speed, etc.) for revealing the influence of extreme weather on power distribution network equipment thermal instability and fault propagation; however, due to the differences in sampling frequency and time stamp between meteorological data and quasi-steady state measurement data and dynamic synchronous phasor data, dynamic time warping algorithm is needed to realize time sequence alignment;

[0170] Obtain meteorological data: obtain real-time observation data from meteorological stations installed in the coverage area of the power distribution network, and use the real-time observation data as meteorological data, which includes temperature, humidity, wind speed, air pressure, etc.

[0171] Adopt dynamic time warping algorithm to perform space-time alignment on quasi-steady state measurement data, dynamic synchronous phasor data and meteorological data, and construct multi-source tensor ; The time dimension is N, the node number is D, and the feature dimension is D.

[0172] In view of the asynchronous sampling characteristics of meteorological data in the multi-source tensor, the following interpolation function is designed, which is represented as:

[0173] ;

[0174] In the formula, Indicates the meteorological data after interpolation processing; Indicates the normalization factor of the interpolation time window; Indicates the length of the time window with t as the end point; Indicates the meteorological influence decay kernel function; Indicates the value of the initial meteorological data at time S; Is the integral variable;

[0175] The meteorological data after interpolation processing And the space-time convolution operator constructs a dynamic graph structure , the dynamic graph structure And Perform feature fusion, and are defined as follows:

[0176] ;

[0177] In the formula, Indicates the fusion function; This represents the inference function for the fault propagation path. Represents the composition of functions; Indicates a fusion operation; This represents the dynamic fusion weights of the k-th data source at time t; Represents a graph convolutional network - a long short-term memory network;

[0178] Calculate the time rate of change of the dynamic fusion weights of the k-th data source at time t, and express it as:

[0179] ;

[0180] In the formula, This represents the rate of change of the dynamic fusion weights of the k-th data source at time t; Indicates the learning rate; Represents the task loss function right The gradient;

[0181] fusion function As the final integrated spatiotemporal feature tensor, based on the final integrated spatiotemporal feature tensor and Construct a fault propagation causal graph and calculate the average causal effect (ACE) using the graph. The fault propagation causal graph includes fault nodes. ,express:

[0182] ;

[0183] In the formula, express Outcome variable; for Outcome variable; Indicates a node Perform intervention operations to make the node In normal condition ; Indicates the faulty node Perform intervention operations to make the faulty node In fault state ;

[0184] When |ACE|>δ, where δ is a set threshold, the faulty node is considered to be... For outcome variables Having causal influence; all fault nodes with causal influence The average causal effect ACE values ​​were sorted, and the fault node with the highest ACE value after sorting was selected as having causal influence. as a power distribution network fault location.

[0185] A multi-source data fusion power distribution network fault accurate positioning system for implementing a multi-source data fusion power distribution network fault accurate positioning method, comprising:

[0186] An acquisition module for acquiring quasi-steady state measurement data and dynamic synchronous phasor data;

[0187] A causal processing module for constructing a dynamic graph network model, inputting the quasi-steady state measurement data and dynamic synchronous phasor data into the dynamic graph network model for processing to obtain a derived causal effect;

[0188] A fusion module for constructing a dynamic weight adaptive fusion model, inputting the quasi-steady state measurement data and dynamic synchronous phasor data into the dynamic weight adaptive fusion model for processing to obtain a spatio-temporal convolution operator;

[0189] A fault location module for obtaining a power distribution network fault location based on the derived causal effect and the spatio-temporal convolution operator.

[0190] As shown in Figure 2 , wherein the present application uses MATLAB / Simulink platform to simulate test the method of the present application, and verify its effect in power distribution network fault accurate positioning;

[0191] From Figure 2 , it can be seen that under single-phase grounding, two-phase short circuit and three-phase short circuit fault types, the accuracy of the present application method is best, the accuracy of the present application method is distributed between 99% and 98.2%; the causal enhanced spatio-temporal graph neural network algorithm is second, the accuracy of the causal enhanced spatio-temporal graph neural network algorithm is distributed between 94.5% and 93.2%; the accuracy of the dynamic weighted multi-source data fusion algorithm is distributed between 90.8% and 92%, which is better than the traditional graph neural network algorithm (accuracy is distributed between 87.2% and 88.5%) and the long short-term memory network algorithm; the accuracy of the long short-term memory network algorithm is the lowest, the accuracy is distributed between 84.8% and 85.8%;

[0192] According to Figure 2 data analysis shows that the present application method performs best in single-phase grounding, the accuracy is about 4.5% higher than that of the causal enhanced spatio-temporal graph neural network algorithm; in two-phase short circuit and three-phase short circuit faults, the present application method is about 4.5% and 5.0% higher than the causal enhanced spatio-temporal graph neural network algorithm respectively; this shows that the present application method has more obvious advantages in dealing with complex faults; it is worth noting that in Figure 2The accuracy of the method, the cause and effect enhanced spatiotemporal graph neural network algorithm, the dynamic weighted multi-source data fusion algorithm, the traditional graph neural network algorithm and the long short-term memory network algorithm all decrease when processing three-phase short-circuit faults, but the decrease amplitude of the method is the smallest, only about 0.3%, which embodies the stability of the method.

[0193] While the embodiments of the application have been shown and described, it is to be understood that the embodiments proposed are only by way of example and are not to be construed in a limiting sense, but the scope of the present application should be gauged from the appended claims and their equivalents.

Claims

1. A method for accurate fault location in a distribution network through multi-source data fusion, characterized in that, include: Step S1: Acquire quasi-steady-state measurement data and dynamic synchronization phasor data; Step S2: Construct a dynamic graph network model, input quasi-steady-state measurement data and dynamic synchronous phasor data into the dynamic graph network model for processing, and obtain the derived causal effect; Step S3: Construct a dynamic weight adaptive fusion model. Input the quasi-steady-state measurement data and dynamic synchronization phasor data into the dynamic weight adaptive fusion model for processing to obtain the spatiotemporal convolution operator. Step S4: Obtain the location of the distribution network fault based on the derived causal effect and spatiotemporal convolution operator; The specific process of step S4 is as follows: Quasi-steady-state measurement data and dynamic synchronous phasor data are divided into the k-th type of data, and the k-th type of data is in the... The feature matrix of time nodes is defined as follows ; For the real number field; Let be the number of nodes at time t; Let k be the feature dimension of the k-th class of data; Meteorological data is acquired, and a dynamic time warping algorithm is used to spatiotemporally align quasi-steady-state measurement data with dynamically synchronized phasor data and meteorological data to construct a multi-source tensor. ; The time dimension is represented by N; the number of nodes is represented by D; and the feature dimension is represented by D. To address the asynchronous sampling characteristics of meteorological data in multi-source tensors, an interpolation function is designed, representing: ； In the formula, This represents the meteorological data after interpolation. The normalization factor represents the interpolation time window; This represents the length of the time window that ends at time t. The kernel function represents the attenuation of meteorological effects; This represents the initial meteorological data value at time S; For integration variables; Meteorological data after interpolation Constructing dynamic graph structures with spatiotemporal convolution operators , dynamic graph structure and Feature fusion is performed, defined as follows: ； In the formula, Indicates the fusion function; This represents the inference function for the fault propagation path. This represents the composition of functions; Indicates a fusion operation; This represents the dynamic fusion weight of the k-th data source at time t; Represents a graph convolutional network - a long short-term memory network; The total number of categories in the k-th category; Let be the adjacency matrix at time t; fusion function As the final integrated spatiotemporal feature tensor, a fault propagation causality graph is constructed based on the final integrated spatiotemporal feature tensor and the derived causal effects. The average causal effect ACE is calculated through the fault propagation causality graph, which includes fault nodes. ,express: ； In the formula, express Outcome variable; for Outcome variable; Indicates a node Perform intervention operations to make the node In normal condition ; Indicates the faulty node Perform intervention operations to make the faulty node In fault state ; This represents the expected value operator; When |ACE|>δ, δ is the set threshold; The faulty node is considered For outcome variables Having causal influence; all fault nodes with causal influence The average causal effect ACE values ​​were sorted, and the fault node with the highest ACE value after sorting was selected as having causal influence. As the location of a fault in the distribution network.

2. The method for accurate fault location in a distribution network based on multi-source data fusion according to claim 1, characterized in that: The processing procedure for dynamic graph network models is as follows: Based on quasi-steady-state measurement data and dynamic synchronous phasor data, time-varying distribution networks are defined as quaternions. ;in, Indicates the first The time-differentiable manifold space, the first The time-differential manifold space is constructed based on quasi-steady-state measurement data; Indicates the first The time-node feature field is constructed based on quasi-steady-state measurement data and dynamic synchronous phasor data; Indicates the first The edge connectivity tensor field at any given time; Indicates the first The form of causal propagation at any given moment; The evolution of dynamic graph network models follows the characteristic transformation equation on the quadrupole manifold, which is expressed as: ； In the formula, This represents the rate of change of the nodal characteristic field at time t; Denotes the covariant derivative; This represents the manifold curvature adjustment factor; ^ denotes the outer product operator; It is a dual operator; Node Neighborhood From the geodesic distance function Sure; and These represent the positions of nodes i and j respectively. Time-differentiable manifold space Geometric position in the middle; For adaptive neighborhood radius; Adaptive neighborhood radius, representing: ； In the formula, This represents the minimum value that satisfies the condition. It means that it is p i A sphere centered at r with radius r; Represents a volume element; Indicates the threshold; No. The edge connectivity tensor field weights at each time step are calculated using a curvature coupling mechanism, and are expressed as follows: ； In the formula, Represents the node at time t. and nodes The edge connectivity tensor field weights between attention scores; Represents a node and nodes Geodetic path; It represents the action of the Riemann curvature tensor in the direction of the tangent vector of the geodesic path, reflecting the degree of curvature of the geodesic path in a geometric sense; Indicates the geodetic path; Indicates the length of a geodesic path element; An information propagation chain is constructed based on the dynamic graph attention mechanism and the joint analysis of causal intervention reasoning. The joint analysis of causal intervention reasoning consists of feature transformation equations and node neighborhoods. The system consists of an adaptive neighborhood radius and edge connectivity tensor field weights; it constrains and filters the information propagation chain to form a fault propagation path F. ， .

3. The method for accurate fault location in a distribution network based on multi-source data fusion according to claim 2, characterized in that: In the quadruple In the middle, the Atiyah-Singer index theorem is used to... and Process and calculate the differential operator. The index is used to establish the differential operator. Among the indicators, topological invariants and fault propagation paths F ， The correspondence is represented as follows: ； ； In the formula, Describing differential operators Indicators; Indicates the first Nodal feature field at time step Chen's characteristics; This represents the outer product operator; Indicates the first Differential manifold space at time of Topology class; For differential operators Indicators and Fault Propagation Path F ， The correspondence; This represents the i-th path fault node on the fault propagation path; For the fault propagation path and Edge; For the fault propagation path Edge; This is the nth path failure node on the fault propagation path; For the first in the fault propagation path One path failure node; based on and F ， The established correspondence will link the fault propagation path F ， Formalized as a dynamic causal network model, it can be represented as: ； In the formula, Indicates the first Time-of-flight dynamic causal network model; Indicates the first The set of nodes in a time-varying dynamic causal network model; Indicates the first The set of edges in a time-varying dynamic causal network model; Indicates the first Time-node feature matrix; Indicates the first The set of constraints for a time-varying dynamic causal network model; A dynamic causal network model is used to capture the evolution of the distribution network topology. The topology includes voltage, current, and power, and is represented as follows: ； In the formula, For the first Time Node and nodes Spatiotemporal attention weights; Represents a learnable transpose weight vector; The weight matrix is ​​a learnable matrix; and The voltage and current characteristics of nodes i and j at time t are respectively. Represents the node at time t. Power state characteristics; For activation functions; Let be the set of neighboring nodes of node i at time t; No. Time Node and nodes The spatiotemporal attention weights are decomposed as follows: ； ； ； In the formula, This represents the concatenated vector of voltage and current characteristics of nodes i and j at time t. This represents the attention score between node i and node j at time t; Let node i and node t be the nodes at time t. Attention score; Based on the Time Node and nodes The spatiotemporal attention weights are used to construct a causal graph structure, which represents the nodes in the distribution network. and nodes The causal relationships between nodes are learned through a dynamic graph attention mechanism. and nodes The causal relationship between them, and the path to obtain the causal relationship; the path based on the causal relationship and the first Time Node and nodes The spatiotemporal attention weights are used to simulate intervention on node j. ; Indicates the budget; Represents a node Intervention variables; Based on the simulation of intervention at node j Define causal effect ,express: ； In the formula, Represents a node The causal effect; Indicates the state of node i; Introducing the parent node set and exogenous variables from the structural equation model, we can represent: ； In the formula, Represents a node The structural equation function; For nodes The set of parent nodes; Let i be an exogenous variable of node i; By introducing the parent node set and exogenous variables from the structural equation model, the causal effect is derived, which is represented as: ； In the formula, This represents the derived causal effect.

4. The method for accurate fault location in a distribution network based on multi-source data fusion according to claim 3, characterized in that: The processing procedure for the dynamic weight adaptive fusion model is as follows: The dynamic fusion weights are derived from the network state entropy and data quality metrics, and are expressed as follows: ； In the formula, This indicates that the weights of the k-th data source are dynamically fused at time t. Indicates the first Data quality metrics for a data source at time t; This represents the network state entropy at time t; The weight parameters are the network state entropy at time t; For the k-th type of data source, the weight parameters of the data quality index at time t are: Let k be the data quality index of the k-th data source at time t; The network state entropy at time t is represented by: ； In the formula, This represents the probability that node v is in a fault state. Let be a logarithmic function; where, ; Indicates the degree of node v; Let u be the degree of node u; The data quality metric for the k-th data source at time t is represented as: ； In the formula, This represents the feature matrix of the real nodes corresponding to the k-th type of data source at time t; The Frobenius norm represents the feature matrix of the real nodes; The fused spatiotemporal feature tensor is composed of and Generate, meaning: ； In the formula, For the fused spatiotemporal feature tensor; This indicates a concatenation operation of feature dimensions; Represents a non-linear activation function; This represents the projection matrix of the feature matrix of the k-th data source node.

5. The method for accurate fault location in a distribution network based on multi-source data fusion according to claim 4, characterized in that: Define the causal enhancement spatiotemporal propagation equation for the fused spatiotemporal feature tensor, which is expressed as: ； In the formula, This represents the spatiotemporal feature tensor after fusion at layer (l+1). Indicates a temporal convolution operation; For activation functions; The dynamic attention coefficient of the spatiotemporal feature tensor after the fusion of node j and node i at time t is represented. This represents the spatiotemporal feature tensor of node j after fusion at time t in layer l; Indicates to Intervention procedures; Represents the gradient operator; Indicates an estimate of the causal effect; The causal enhancement coefficient; Let i be the set of neighboring nodes of node i; The dynamic attention coefficients of the spatiotemporal feature tensor after the fusion of nodes j and i at time t are calculated, and are represented as follows: ； In the formula, This represents the spatiotemporal feature tensor of node i after fusion at time t in layer l; This represents the spatiotemporal feature tensor difference vector of nodes i and j after fusion at time t; For node i and node The spatiotemporal feature tensor difference vector after fusion at time t; For splicing operations; The estimated value of the causal effect is calculated and expressed as follows: ； In the formula, This indicates the expected value of the target variable of node x after intervention at time t; Represents the relationship between nodes at time t. After non-intervention, the node is obtained. The expected value of the target variable; Represents the loss function right The partial derivatives; This represents the spatiotemporal feature tensor of node v after fusion at time t in layer l; Let v represent the node feature matrix at time t; express right The partial derivatives; These are partial derivatives; The spatiotemporal convolution operator, through the fusion of bidirectional gated recurrent units and the causal-enhanced spatiotemporal propagation equations using TConv units, is expressed as: ； In the formula, Indicates input Perform spatiotemporal convolution operations to obtain the spatiotemporal convolution operator; Indicates a bidirectional gated loop unit; Based on The inverse square root of the dynamically generated degree matrix at time t; This indicates element-wise multiplication; Represents the Long Short-Term Memory network; for The weight matrix; It is a spatiotemporal feature tensor.

6. A multi-source data fusion-based distribution network fault accurate location system, used to implement the multi-source data fusion-based distribution network fault accurate location method according to any one of claims 1-5, characterized in that, include: The acquisition module is used to acquire quasi-steady-state measurement data and dynamic synchronization phasor data; The causal processing module is used to construct a dynamic graph network model. Quasi-steady-state measurement data and dynamic synchronous phasor data are input into the dynamic graph network model for processing to obtain the derived causal effects. The fusion module is used to construct a dynamic weight adaptive fusion model. Quasi-steady-state measurement data and dynamic synchronization phasor data are input into the dynamic weight adaptive fusion model for processing to obtain the spatiotemporal convolution operator. The fault location module is used to obtain the location of the fault in the distribution network based on the derived causal effect and the spatiotemporal convolution operator.

Images