CN-122017474-A - PINN-transducer model-based power distribution network single-ended traveling wave fault positioning method and system

CN122017474ACN 122017474 ACN122017474 ACN 122017474ACN-122017474-A

Abstract

The invention discloses a method and a system for locating single-ended traveling wave faults of a power distribution network based on PINN-transducer model, comprising the following steps: acquiring a fault voltage transient time-frequency diagram in real time, inputting the fault voltage transient time-frequency diagram to a trained PINN-transducer model, and synchronously outputting a fault feeder result and a precise positioning distance; the invention has the advantages of fault time-frequency characteristic extraction of deep learning and the mechanism constraint characteristic of a physical information neural network, and ensures that the positioning result improves the ranging precision under complex working conditions to a certain extent on the premise of meeting the physical consistency by constructing a tiny physical residual error item.

Inventors

DENG FENG
LIU ZHENGYANG
TANG TIAN
WANG ZHENTONG
Liu Xianghan
REN XINGYU
YANG MINGYANG
TANG CHANG

Assignees

长沙理工大学

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (11)

1. A power distribution network single-ended traveling wave fault positioning method based on PINN-transducer model is characterized by comprising the following steps: The method comprises the steps of collecting a fault voltage transient state time-frequency diagram in real time, inputting the fault voltage transient state time-frequency diagram to a trained PINN-transducer model, and synchronously outputting a fault feeder line result and an accurate positioning distance, wherein the training process of the PINN-transducer model comprises the following steps: S1, collecting fault traveling wave samples in a set time window after faults, and obtaining input features corresponding to corresponding sample fault voltage transient state time-frequency diagrams ; S2, will N encoder modules are respectively input in parallel and used for capturing global time-frequency characteristics of input signals, each encoder module comprises a multi-head attention layer and a feedforward neural network, and output characteristic vectors of the N encoder modules are output to a fault line selection branch and a fault distance measurement branch in parallel to respectively obtain probability distribution and predicted fault distances; S3, converting the joint loss function into a gradient flow through a back propagation algorithm, and driving the network parameters to approach to a physical consistency area in a solution space until a PINN-fransformer model converges, wherein the joint loss function comprises Data loss term and transient physical consistency loss term And boundary constraint terms 。
2. The method for locating single-ended traveling wave faults of a power distribution network based on PINN-transducer model as claimed in claim 1, wherein the specific content of S1 comprises: S11, constructing an electromagnetic transient model of the power distribution network, setting a plurality of feeder lines and traversing fault conditions formed based on key fault parameters to simulate, wherein the key fault parameters comprise single-phase grounding fault types, transition resistances and fault initial angles; S12, recording bus transient voltage and outlet transient current after faults occur, and intercepting a transient short window voltage sequence and a current sequence after a fault triggering moment; s13, establishing a sample index table, and associating fault positions with branch number labels; S14, performing continuous wavelet transformation on the voltage sequence, and then performing normalization processing to obtain a fault voltage transient state time-frequency diagram ; S15, will Mapping to input matrix X and overlapping position codes to serve as input features of deep learning model 。
3. The method for locating the single-ended traveling wave fault of the power distribution network based on PINN-transducer model according to claim 2, wherein the specific contents of S15 include: Transient time-frequency diagram of fault voltage Mapping to an input matrix R represents X as a real matrix, T as a time step, Is a feature dimension; generating a position code by adopting sine and cosine functions: ; ; In the formula, A position index representing a time step is presented, Representing a feature dimension index; Then the input features as a deep learning model after superposition position coding Expressed as: ; Wherein PE represents position encoding.
4. The method for locating single-ended traveling wave faults of a power distribution network based on PINN-transducer model as claimed in claim 1, wherein the specific content of S2 comprises: s21 input features Respectively entering N encoder modules in parallel, and respectively carrying out the following contents in each encoder module: ① Input features Respectively entering h attention heads; ② In each attention head, a feature is entered Through a weight matrix 、 And Mapping to obtain a query matrix Key matrix Sum matrix Based on 、 And Obtain corresponding single-head attention score ; ③ Multiple-head attention is executed in parallel in step ② i times, the outputs of all attention heads are spliced and then are subjected to linear transformation Obtain the final output of the multi-head attention layer Output after residual connection and layer normalization ; ④ Input to a feedforward neural network FFN, and FFN output characteristics are obtained through two full-connection layers After residual connection and layer normalization, final output feature vector is output ; S22. Parallel input fault line selection branch and fault ranging branch, respectively outputting fault probability distribution of each feeder line branch And predicting the fault distance 。
5. The method for locating a single-ended traveling wave fault in a power distribution network based on PINN-transducer model as claimed in claim 4, wherein the specific contents of step ② include: Input features Through a weight matrix 、 And Mapping to obtain a query matrix Key matrix Sum matrix : ; Normalized by Softmax function to obtain weight and act on Obtaining a corresponding single-head attention score : ; In the formula, Is that And (3) with Is used for the number of columns of (a), Is a scaling factor.
6. The method for single-ended traveling wave fault location in a power distribution network based on PINN-transducer model as claimed in claim 4, wherein the final output of multiple attention layers in step ③ The method comprises the following steps: ; ; In the formula, Is the first A self-attention head output is provided, 、、 Is the first A weight matrix of the individual attention headers.
7. The method for locating a single-ended traveling wave fault in a power distribution network based on PINN-transducer model as claimed in claim 4, wherein the specific contents of step ④ include: Input to a feed-forward neural network FFN, and pass through two fully connected layers, wherein a ReLU activation function is adopted in the middle layer to obtain FFN output characteristics : ; In the formula, As a matrix of weights, the weight matrix, Is a bias term; After residual connection and layer normalization, final output feature vector is output : ; ; In the formula, And The mean and standard deviation of the inputs are respectively, In order to prevent a small amount of zero removal, And Is a learnable affine transformation parameter.
8. The method for locating single-ended traveling wave faults of a power distribution network based on PINN-transducer model according to claim 1, wherein the specific content of S3 includes: Within each training round Epoch, training set samples are shuffled and divided into batches, and the following is performed for each batch: Performing forward propagation according to the time-frequency diagram corresponding to the fault transient state of the batch And (3) with Calculating a joint loss value based on the joint loss function; performing back propagation and gradient calculations to calculate joint loss values for all learnable parameters of the network based on the chain law Gradient of (2) ; By calculating gradients First moment estimation of (a) Sum and second moment estimation The estimated value and the current learning rate are further combined to carry out parameter adjustment on the current time parameter; Evaluating the generalization performance and the physical residual error level of the model by using a verification set after each training round of Epoch is finished: when the joint loss of the verification set is not reduced any more, the physical residual error converges to a preset physical tolerance threshold or the maximum iteration round is reached, judging that the model converges, triggering the early shutdown system to finish training, and solidifying the current optimal parameters to obtain a final PINN-transducer fault positioning model; Otherwise, continuing the next training round until convergence.
9. The method for locating single-ended traveling wave faults in power distribution network based on PINN-transducer model as claimed in claim 8, wherein the gradient is The method comprises the following steps: ; Updating first moment estimates Sum and second moment estimation : ; ; In the formula, And All are exponential decay rates; estimation of first moment Sum and second moment estimation Performing deviation correction to eliminate zero deviation influence caused by initialization: ; ; In the formula, And Respectively is And Is a correction value of (2); the current time parameter is adjusted by combining the correction value and the current learning rate as follows: ; In the formula, For the parameters updated at the next moment in time, As a parameter of the current moment of time, Is the learning rate.
10. The method for locating single-ended traveling wave faults of a power distribution network based on PINN-transducer model as claimed in claim 8, wherein the joint loss function is: ; In the formula, Is a data loss term; Is a transient physical consistency loss term; Is a boundary constraint term; And (3) with Regularized weight coefficients for the physical constraint terms; Data loss term : ; ; ; In the formula, And As the weight coefficient of the light-emitting diode, Is regression loss; And (3) with Respectively the first The actual fault distance and model prediction distance of each sample, n is the number of samples of the current training batch, For cross entropy loss, K is the total number of feeder branches, For the value indicated by the real tag, The prediction probability of the ith sample belonging to the kth branch is calculated for the model; Transient physical consistency loss term : Selecting a short time window near a fault wave head And taking the baseline value before the short window to construct an incremental signal: ; ; In the formula, As a function of the transient voltage increment, For the transient bus voltage, the voltage of the bus, In order to increase the amount of transient current, In order to correspond to the transient line out current, And Respectively is And According to the traveling wave theory, the short-time traveling wave component satisfies the approximate consistency relation: ; ; In the formula, For the characteristic impedance of the line, For the line inductance to be a function of the line inductance, Is the line capacitance; Defining transient voltage-current consistency residuals as: ; In order to eliminate amplitude scale differences caused by different voltage levels and load fluctuation, a normalized factor den is introduced to construct a dimensionless relative physical residual error, and a mean absolute error MAE form is adopted to define physical loss: ; ; Then: ; Wherein, the The dimensionless index after the amplitude normalization of the physical consistency residual error is used for training constraint and evaluating the confidence coefficient of the result; (3) Boundary constraint term : ; In the formula, For the maximum distance of the corresponding line from the measuring end to the tail end The term is 0 when And when the boundary is exceeded, the punishment is rapidly increased, and the prediction result is constrained in a physical feasible domain through a gradient descent forcing model.
11. A single-ended travelling wave fault location system for a power distribution network based on PINN-transducer model, comprising a processor, a memory and a computer program stored on the memory and executable by the processor, wherein the processor implements a single-ended travelling wave fault location method for a power distribution network based on PINN-transducer model as defined in any one of claims 1 to 10 when the computer program is executed by the processor.

Description

PINN-transducer model-based power distribution network single-ended traveling wave fault positioning method and system Technical Field The invention relates to the technical field of power distribution network fault location, in particular to a power distribution network single-ended traveling wave fault location method and system based on PINN-transducer model. Background The existing power distribution network fault positioning method mainly comprises an impedance method, a traveling wave method, a signal injection method, a power distribution automation method, an artificial intelligent algorithm and the like. The method is characterized in that after a single-phase grounding fault occurs in the system, a signal of a specific frequency is actively injected into the power grid through a neutral point of a transformer or special injection equipment, the method has high positioning precision and is not influenced by the grounding mode of the neutral point, but has high hardware investment cost, the positioning effect is easily influenced by parameters such as a wire-to-ground distributed capacitor, a grounding resistor and the like, a traditional artificial intelligent algorithm utilizes a deep learning construction feature and a position mapping model, has the advantage of nonlinear characterization, is taken as a pure data driving model, belongs to a black box model in nature, is easy to excessively rely on statistical correlation of samples in time limit when the transient state signal is processed, is easy to cause the statistics of the samples or easy to change in the fault or has poor fitting performance of the sample, and cannot guarantee the positioning performance of the system is difficult to be interpreted as a basic result. In a line wave fault positioning scene, a Transformer, CNN-LSTM and other pure data driving models are easy to output non-physical feasible solutions when working conditions such as high-resistance grounding, strong noise interference and the like are processed, and the problem that a positioning result exceeds the whole length of a line, and the voltage and current traveling wave increment does not meet the physical inconsistency of a transient mechanism and the like occurs. Disclosure of Invention In view of the above, the invention provides a single-ended traveling wave fault positioning method and system for a power distribution network based on PINN-transducer model, which are used for at least solving the problems of strong parameter dependence of traveling wave physical model and limited generalization performance of the traditional deep learning method under high-resistance and strong-noise scenes in the prior art. In order to achieve the above purpose, the present invention adopts the following technical scheme: A power distribution network single-ended traveling wave fault positioning method based on PINN-transducer model comprises the following steps: The method comprises the steps of collecting a fault voltage transient state time-frequency diagram in real time, inputting the fault voltage transient state time-frequency diagram to a trained PINN-transducer model, and synchronously outputting a fault feeder line result and an accurate positioning distance, wherein the training process of the PINN-transducer model comprises the following steps: S1, collecting fault traveling wave samples in a set time window after faults, and obtaining input features corresponding to corresponding sample fault voltage transient state time-frequency diagrams ; S2, willN encoder modules are respectively input in parallel and used for capturing global time-frequency characteristics of input signals, each encoder module comprises a multi-head attention layer and a feedforward neural network, and output characteristic vectors of the N encoder modules are output to a fault line selection branch and a fault distance measurement branch in parallel to respectively obtain probability distribution and predicted fault distances; S3, converting the joint loss function into a gradient flow through a back propagation algorithm, and driving the network parameters to approach to a physical consistency area in a solution space until a PINN-fransformer model converges, wherein the joint loss function comprises Data loss term and transient physical consistency loss termAnd boundary constraint terms。 Preferably, the specific content of S1 includes: S11, constructing an electromagnetic transient model of the power distribution network, setting a plurality of feeder lines and traversing fault conditions formed based on key fault parameters to simulate, wherein the key fault parameters comprise single-phase grounding fault types, transition resistances and fault initial angles; S12, recording bus transient voltage and outlet transient current after faults occur, and intercepting a transient short window voltage sequence and a current sequence after a fault triggering moment; s13, establishing a sample inde