A power transmission line fault diagnosis method and system based on a QRNN-Bi-GRU model

Abstract

The application discloses a power transmission line fault diagnosis method and system based on an improved QRNN-Bi-GRU model, and the method first constructs a QRNN-Bi-GRU fault recognition classification model comprising a QRNN feature extraction module, a Bi-GRU time sequence modeling module and a full-connection classification module, then obtains power transmission line fault data, extracts multi-dimensional quantum enhanced features based on the QRNN feature extraction module, then performs bidirectional time sequence modeling based on the Bi-GRU time sequence modeling module, obtains bidirectional hidden states, and then performs linear transformation through the full-connection classification module to output probability distribution of various faults, and the power transmission line fault determination is completed; the application fuses the fault recognition classification model of the QRNN quantum recurrent neural network and the Bi-GRU bidirectional recurrent gate unit, strengthens the extraction of key features, effectively alleviates the gradient vanishing problem, and improves the accuracy of recognizing various power transmission line faults in power transmission and transformation engineering.

Description

Technical Field

[0001] This invention relates to fault diagnosis methods, belonging to the field of power system fault detection, and particularly to a method and system for diagnosing transmission line faults based on the QRNN-Bi-GRU model. Background Technology

[0002] In modern power systems, transmission lines are the core carriers for transmitting electrical energy, and their safe and stable operation is an important foundation for ensuring a continuous and reliable power supply. Because transmission lines are in a complex natural environment for a long time and also have to withstand various electrical stresses within the power system, they are prone to various faults such as lightning strikes, flashovers, and broken conductor strands. These faults not only cause power outages, disrupting normal social production and residents' lives, but also cause huge economic losses and even induce various safety accidents. Currently, when faced with multiple faults, high-resistance grounding faults, and complex operating conditions such as harmonic and noise interference, the accuracy of traditional fault diagnosis methods is significantly reduced, and they can no longer meet the requirements of actual engineering applications.

[0003] Currently, various fault diagnosis technologies for transmission lines have emerged. Among them, fault diagnosis models based on convolutional neural networks (CNNs) can build corresponding network structures to complete fault diagnosis. However, these models struggle to deeply explore the long-term dependencies of fault signals over time, resulting in significant shortcomings in overall recognition accuracy when faced with complex time-series fault patterns. Some diagnostic schemes integrate CNNs, bidirectional gated recurrent units, and optimization algorithms to extract fault features through signal decomposition and fine-tune network parameters using optimization algorithms. However, the optimization algorithms employed perform poorly in multi-objective parameter optimization scenarios, easily leading to repeated model training and impacting overall operational efficiency. Other fault diagnosis methods based on ensemble learning rely on traditional methods to extract fault features and build classification models. However, traditional feature extraction methods cannot fully exploit the nonlinear and non-stationary complex characteristics of transmission line fault data, thus limiting the final fault diagnosis accuracy of ensemble learning models. Therefore, a more accurate fault detection method is urgently needed to address the aforementioned shortcomings of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned defects and problems in the prior art and to provide a transmission line fault diagnosis method and system based on the QRNN-Bi-GRU model with high accuracy.

[0005] To achieve the above objectives, the technical solution of this invention is: a method for fault diagnosis of transmission lines based on a QRNN-Bi-GRU model, comprising:

[0006] A QRNN-Bi-GRU fault identification and classification model is constructed based on QRNN quantum recurrent neural network and Bi-GRU bidirectional cyclic gate unit; the QRNN-Bi-GRU fault identification and classification model includes a QRNN feature extraction module, a Bi-GRU time series modeling module and a fully connected classification module.

[0007] Data on power transmission line faults are acquired, and quantum entanglement and feature extraction are performed based on the QRNN feature extraction module to obtain multidimensional quantum enhanced features.

[0008] Based on the multidimensional quantum enhancement feature, bidirectional temporal modeling is performed through the Bi-GRU temporal modeling module to obtain bidirectional hidden states;

[0009] Based on the bidirectional hidden state, a linear transformation is performed through a fully connected classification module to output the probability distribution of various faults, thus completing the fault determination of transmission lines.

[0010] Optionally, the QRNN feature extraction module adopts an sQRNN interleaved quantum recurrent neural network, which is composed of multiple sets of QRB quantum recurrent blocks stacked alternately;

[0011] A single QRB quantum loop block integrates a data encoding unit, an Ansatz variable quantum circuit unit, and a portion of the quantum measurement unit;

[0012] The Bi-GRU temporal modeling module includes a forward GRU subnetwork, a backward GRU subnetwork, and a hidden state splicing layer;

[0013] An attention mechanism layer is embedded after the output layer of the Bi-GRU time series modeling module.

[0014] Optionally, the acquisition of transmission line fault data, based on the QRNN feature extraction module for quantum entanglement and feature extraction, to obtain multidimensional quantum enhanced features, specifically includes:

[0015] Obtain the fault feature vector of the fault condition, and normalize the fault feature vector to... After grading, angle encoding is performed, and its expression is as follows:

[0016] ;

[0017] ;

[0018] in: Let be the unitary matrix of the revolving door; , These are the original and normalized values ​​of the fault feature vector, respectively. This represents the initial state of a qubit; The number of qubits; It is the quantum state tensor product;

[0019] The angle-encoded input data is fed into an Ansatz variable quantum circuit. Quantum entanglement between the data register Reg.D and the historical information register Reg.H is achieved through a three-layer structure of single-quantum gates and double-quantum gates within the Ansatz variable quantum circuit, thus extracting the correlation features between the input data and historical information. Each layer contains a single quantum gate. and two quantum gates Its expression is as follows:

[0020] ;

[0021] ;

[0022] in: , , , All of these are rotation angles corresponding to the rotation gates of a single quantum bit; , They act on the first The and the first Pauli-Z operator for 1 qubit; To bypass Pauli revolving door with an axle; To bypass Pauli revolving door with an axle; for Pauli revolving door;

[0023] After completing quantum entanglement and feature extraction, a projection measurement with the Pauli-Z operator as the observable is performed on the first qubit of the data register Reg.D, and a single classical eigenvalue is output. The single classical eigenvalue is used as the one-dimensional feature corresponding to a single QRB quantum loop block. All QRB quantum loop blocks are traversed in turn, and the one-dimensional features are summarized to obtain a multi-dimensional quantum enhancement eigenvector.

[0024] Optionally, the step of obtaining bidirectional hidden states by performing bidirectional temporal modeling through the Bi-GRU temporal modeling module based on multidimensional quantum enhancement features specifically includes:

[0025] The first in the multidimensional quantum enhancement eigenvector Input features at each time step The hidden state of the previous time step Calculate the reset door separately and the update gate The expression for controlling the retention and updating of historical information is as follows:

[0026] ;

[0027] ;

[0028] in: It is the sigmoid activation function; , , , Both are weight matrices; , All are bias vectors;

[0029] The first in the multidimensional quantum enhancement eigenvector Input features at each time step Concealed state after resetting the door The hidden states are generated through fusion, and their expression is as follows:

[0030] ;

[0031] in: This is a candidate hidden state; It is the hyperbolic tangent activation function; , Both are weight matrices; This is element-wise multiplication; It is the bias vector;

[0032] Based on the generated candidate hidden states, the multidimensional quantum enhancement feature vectors are traversed in forward GRU order and backward GRU order to obtain bidirectional hidden states. .

[0033] Optionally, the step of using a bidirectional hidden state to perform a linear transformation through a fully connected classification module to output the probability distribution of various faults and complete the transmission line fault determination specifically includes:

[0034] Based on the attention mechanism, feature mapping is performed on the bidirectional hidden state to generate query vector, key vector, and value vector. Then, based on the query vector, key vector, and value vector, the attention weights are obtained, and their expressions are as follows:

[0035] ;

[0036] ;

[0037] in: , , These are the query vector, key vector, and value vector, respectively. , , Both are weight matrices; Attention weights; The number of time steps; The dimension of the key vector;

[0038] We perform a weighted summation based on the attention weights to obtain the input feature representation after processing by the attention mechanism, as shown in the following expression:

[0039] ;

[0040] in: For input feature representation;

[0041] Represent the input features The input fully connected layer undergoes a linear transformation, and the probability distribution of various fault types is calculated using the cross-entropy loss function, as shown in the following expression:

[0042] ;

[0043] ;

[0044] ;

[0045] in: For each The predicted probabilities of various faults obtained by performing the Softmax operation; The original score for the fault characteristics; For the first The original score for the type of fault; For the first Fault category tags; This is the weight matrix; The total number of samples; For sample index; This is the cross-entropy loss function.

[0046] Optionally, obtaining the fault feature vector of the fault condition specifically includes:

[0047] A fault simulation model of a transmission line is constructed, historical fault signals of the transmission line are collected, and based on variational mode decomposition, the historical fault signals are decomposed into... The expression for each IMF modal component is as follows:

[0048] ;

[0049] ;

[0050] in: For the first One IMF modal component; , These are the instantaneous amplitude and phase, respectively. The center frequency; The unit impulse function; This is a convolution operation; This is a fault signal; The preset number of decomposition modes; The modal component number;

[0051] Key features are extracted from the IMF modal components, including energy percentage, kurtosis factor, and permutation entropy, and their expressions are as follows:

[0052] ; ;

[0053] ;

[0054] in: , , These are energy percentage, kurtosis factor, and permutation entropy, respectively. Signal duration; For mathematical expectation; To reconstruct the probability of the arrangement pattern in the phase space; Total number of patterns; For the first One IMF modal component;

[0055] The key features are concatenated to obtain the fault feature vector, which is expressed as follows:

[0056] ;

[0057] in: This is the fault feature vector.

[0058] Optionally, the construction of the fault simulation model for the transmission line specifically includes:

[0059] The available distributed parameters for the fault simulation model are set as follows:

[0060] ;

[0061] in: For any position on the line; , These are voltage and current vectors, respectively; For resistance; It is an inductor; Electrical conductivity; For capacitors; For frequency; The propagation constant; Characteristic impedance;

[0062] For any point on the line, the voltage and current vectors at the beginning and end of the line are as follows:

[0063] ;

[0064] in: It is a hyperbolic sine function; It is a hyperbolic cosine function;

[0065] To simulate the transient state of a transmission line fault, the transient process control equations for the transmission line are established, as follows:

[0066] ;

[0067] ;

[0068] in: For the location of the line place, time The instantaneous voltage; For the location of the line place, time The instantaneous current;

[0069] Solve the transient process control equations of the transmission line to obtain the traveling wave propagating along the line, the expression of which is as follows:

[0070] ;

[0071] ;

[0072] in: A traveling wave propagating along the path; , These are forward and reverse traveling waves, respectively. The speed of travel wave propagation;

[0073] Considering the environmental factors affecting actual transmission lines, an environmental correction is added, the expression of which is as follows:

[0074] ;

[0075] in: , , These are the corrected resistors, inductors, and capacitors for transmission lines covered by ice and snow. This represents the original cross-sectional area of ​​the transmission line; The cross-sectional area of ​​the power transmission line after it has been covered with ice and snow; This refers to the altitude above the ground. This refers to the amount of downward sag.

[0076] The expressions for setting the grounding resistance and ground modulus attenuation coefficient of transmission lines are as follows:

[0077] ; ;

[0078] in: This is the grounding resistance attenuation coefficient; This is the ground model attenuation coefficient; This represents the actual soil resistivity. Reference resistivity;

[0079] A lightning strike signal is constructed using a double exponential wave to simulate a lightning fault. The double exponential function is as follows:

[0080] ;

[0081] in: This represents the instantaneous value of the lightning strike current. This refers to the amplitude of the lightning strike current. The decay time constant; The step time constant; It is a constant coefficient.

[0082] Optionally, after constructing the QRNN-Bi-GRU fault identification and classification model, the method further includes:

[0083] The hyperparameters of the QRNN-Bi-GRU fault identification and classification model are iteratively optimized based on the IKOA-improved Kepler optimization algorithm, and the optimal hyperparameters are output.

[0084] The IKOA improved Kepler optimization algorithm specifically includes:

[0085] To improve the global search phase of the KOA algorithm, a perturbation mechanism is introduced, as shown in the following expression:

[0086] ;

[0087] ;

[0088] in: For the first The updated global optimal position of each planet; , The first The positions of the planets before and after the update; This is the fitness value; The position of the sun is randomly selected from an external archive library; The disturbance coefficient; A random number within the range [0, 1]; This represents the number of iterations. This represents the maximum number of iterations.

[0089] For the exploration and development transition phase of the KOA algorithm, a dynamic decay factor is introduced to improve the adjustment parameters, and its expression is as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] in: To adjust the parameters; It is a dynamic decay factor; These are random values ​​based on a normal distribution. A random value between [0, 1]; These are the parameters for cyclic control. To control the initial value of the loop; This is the modulo operation;

[0094] An external archive mechanism is introduced to retain multiple non-dominated solutions and provide alternative locations for randomized optimal solutions, as detailed below:

[0095] If the solution No inferior solution on all objective functions And solve It performs strictly better than the solution on at least one objective function. The term "explanation" is sometimes used to describe the process of solving problems. Dominate , recorded as The basis for judgment is:

[0096] ;

[0097] in: To solve In the The values ​​that can be obtained on the objective function; To solve In the The values ​​that can be obtained on the objective function;

[0098] The solution filtering mechanism is optimized by removing redundant solutions based on the density of the region where the solution resides. Its expression is as follows:

[0099] ;

[0100] in: Density; The number of objective functions; For the first The normalized value of the fitness value of each objective function.

[0101] Optionally, the iterative optimization of the hyperparameters of the QRNN-Bi-GRU fault identification and classification model based on the IKOA-based improved Kepler optimization algorithm to output the optimal hyperparameters specifically includes:

[0102] The parameters to be optimized are combined and encoded into planetary position vectors using the IKOA improved Kepler optimization algorithm. ,in The total number of parameters; the combination of parameters to be optimized includes the angle encoding rotation angle of the QRNN quantum recurrent neural network, the circuit parameters and measurement parameters of the Ansatz variable quantum circuit; and the gating parameters, hidden layer parameters, and training parameters of the Bi-GRU.

[0103] Set up an optimization multi-objective function, the expression of which is as follows:

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] in: , , All are optimization objectives; , These represent the number of correctly classified samples and the total number of samples, respectively. , These represent the circuit depth and required quantum coherence time for the Ansatz variable quantum circuit, respectively. , All are coefficient weights; , These represent the running time of the QRNN quantum recurrent neural network and the backpropagation time of the Bi-GRU bidirectional recurrent gate unit, respectively.

[0109] Set initialization parameters and randomly generate initial position vectors for each planet. ; Calculate the multi-objective function values ​​for each planet. Based on the positional relationship between the planet and the Sun, the gravitational force between the planet and the Sun and the planet's orbital velocity are calculated, and the expressions are as follows:

[0110] ;

[0111] ;

[0112] ;

[0113] in: The gravitational pull between the planet and the sun; , These are the masses of the Sun and the planets, respectively. It is the gravitational coefficient; The Euclidean distance between the planet and the Sun; It is a very small positive number; , These are the current planetary orbital speed and the updated planetary orbital speed, respectively. For random numbers The controllable planetary orbit direction coefficient;

[0114] The global search mechanism based on the IKOA-improved Kepler optimization algorithm performs planetary position updates, and the non-dominated solutions generated during the iteration process are stored in the external archive based on the external archive mechanism and solution filtering mechanism.

[0115] When the algorithm iterates to the preset maximum number of iterations, it outputs the Pareto front solution set, calculates the distance of each solution in the solution set to the ideal point, selects the solution with the smallest distance as the optimal solution, and obtains the optimal hyperparameters.

[0116] A transmission line fault diagnosis system based on a QRNN-Bi-GRU model, which applies the above-mentioned method, includes:

[0117] The fault model construction module is used to construct a QRNN-Bi-GRU fault identification and classification model based on QRNN quantum recurrent neural network and Bi-GRU bidirectional cyclic gate unit; the QRNN-Bi-GRU fault identification and classification model includes a QRNN feature extraction module, a Bi-GRU time series modeling module and a fully connected classification module.

[0118] The fault determination module is used to acquire transmission line fault data. Based on the QRNN feature extraction module, quantum entanglement and feature extraction are performed to obtain multi-dimensional quantum enhanced features.

[0119] Based on the multidimensional quantum enhancement feature, bidirectional temporal modeling is performed through the Bi-GRU temporal modeling module to obtain bidirectional hidden states;

[0120] Based on the bidirectional hidden state, a linear transformation is performed through a fully connected classification module to output the probability distribution of various faults, thus completing the fault determination of transmission lines.

[0121] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0122] 1. This invention discloses a method and system for diagnosing transmission line faults based on a QRNN-Bi-GRU model. The method first constructs a QRNN-Bi-GRU fault identification and classification model comprising a QRNN feature extraction module, a Bi-GRU time-series modeling module, and a fully connected classification module. Then, it acquires transmission line fault data, extracts multi-dimensional quantum enhancement features based on the QRNN feature extraction module, performs bidirectional time-series modeling based on the Bi-GRU time-series modeling module to obtain bidirectional hidden states, and finally performs linear transformation through the fully connected classification module to output the probability distribution of various fault types, thus completing the transmission line fault determination. In application, this design, by integrating the QRNN quantum recurrent neural network and the Bi-GRU bidirectional cyclic gate unit fault identification and classification model, strengthens the extraction of key features, effectively alleviates the gradient vanishing problem, and improves the accuracy of identifying various transmission line faults in power transmission and transformation projects.

[0123] 2. The present invention discloses a transmission line fault diagnosis method and system based on the QRNN-Bi-GRU model. The QRNN-Bi-GRU fault identification and classification model includes a QRNN feature extraction module, a Bi-GRU time series modeling module, and a fully connected classification module. In application, the QRNN feature extraction module can quickly extract local key features of fault data, reduce computational load, and improve operating efficiency, while Bi-GRU can bidirectionally mine the temporal dependencies of data, fully capturing the dynamic evolution process of the fault. The combination of the two achieves deep fusion of local features and time series information, preserving fine-grained fault features while comprehensively representing the temporal change pattern, effectively overcoming the problems of insufficient feature extraction and one-sided time series information mining by a single model.

[0124] 3. This invention relates to a transmission line fault diagnosis method and system based on a QRNN-Bi-GRU model. It optimizes the Kepler optimization algorithm by introducing a perturbation mechanism, dynamic attenuation factor, solution filtering, and external archive mechanism. Furthermore, it iteratively optimizes the hyperparameters of the QRNN-Bi-GRU fault identification and classification model using an improved Kepler optimization algorithm based on IKOA. In application, this design effectively improves the problems of traditional Kepler optimization algorithms, such as being prone to local optima, slow convergence speed, and insufficient search diversity, through multiple optimization strategies. Simultaneously, it utilizes the IKOA algorithm to achieve adaptive optimization of model hyperparameters, quickly selecting the optimal parameter combination, fully releasing the feature extraction and temporal modeling performance of the QRNN-Bi-GRU model, and further enhancing the model's ability to identify multiple types of faults.

[0125] 4. In the transmission line fault diagnosis method and system based on the QRNN-Bi-GRU model of this invention, when performing hyperparameter optimization based on the IKOA improved Kepler optimization algorithm, a multi-objective function is set. In application, this design, by setting a multi-objective optimization function, simultaneously optimizes the classification accuracy of fault identification, quantum circuit overhead and stability, and model running efficiency, effectively avoiding the performance imbalance problem caused by single-objective optimization. While ensuring the accuracy of fault identification, it also takes into account the engineering feasibility and real-time running efficiency of the solution, making the model more robust and practically deployable. Attached Figure Description

[0126] Figure 1 This is a flowchart of the method of the present invention.

[0127] Figure 2 This is a network structure diagram of the QRNN-Bi-GRU fault identification and classification model in Embodiment 1 of the present invention.

[0128] Figure 3 This is a flowchart of the fault identification process in Embodiment 2 of the present invention.

[0129] Figure 4 This is a schematic diagram of the confusion matrix of the test set in Embodiment 2 of the present invention.

[0130] Figure 5 This is an accuracy iteration curve diagram in Embodiment 2 of the present invention.

[0131] Figure 6 This is a system structure diagram of the present invention.

[0132] Figure 7 This is a structural diagram of the device of the present invention.

[0133] In the diagram: Fault model construction module 1, fault determination module 2, processor 3, memory 4, computer program code 41. Detailed Implementation

[0134] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0135] Example 1:

[0136] See Figure 1 A method for fault diagnosis of transmission lines based on a QRNN-Bi-GRU model, comprising:

[0137] A QRNN-Bi-GRU fault identification and classification model is constructed based on QRNN quantum recurrent neural network and Bi-GRU bidirectional cyclic gate unit; the QRNN-Bi-GRU fault identification and classification model includes a QRNN feature extraction module, a Bi-GRU time series modeling module and a fully connected classification module.

[0138] Furthermore, the QRNN feature extraction module adopts an sQRNN interleaved quantum recurrent neural network, which is composed of multiple sets of QRB quantum recurrent blocks stacked alternately;

[0139] A single QRB quantum loop block integrates a data encoding unit, an Ansatz variable quantum circuit unit, and a portion of the quantum measurement unit;

[0140] The Bi-GRU time series modeling module has a two-layer stacked structure, with 64 hidden units in each layer, including a forward GRU sub-network, a backward GRU sub-network, and a hidden state splicing layer; it has a two-layer stacked structure, with 64 hidden units in each layer.

[0141] An attention mechanism layer is embedded after the output layer of the Bi-GRU time series modeling module.

[0142] In this embodiment, QRNN (Quantum Recurrent Neural Network) is a novel model that integrates the characteristics of quantum computing and recurrent neural networks. Its core relies on parameterized quantum circuits (PQC) to achieve time-series data learning, breaking through the computational bottleneck of classical RNNs. QRNN can encode fault feature vectors into quantum states and realize the dynamic evolution of quantum states through quantum gate operations. At the same time, the historical quantum state information stored in the quantum register (Reg.H) is combined with the current input quantum register (Reg.D) to transform quantum measurement into prediction results.

[0143] Bi-GRU, by introducing a reverse recurrent loop, consists of two layers of gated recurrent units that process sequences from both the forward and reverse directions. This allows it to fully capture temporal information and concatenate the hidden vectors of the hidden neurons as the final output. With its gating mechanism, Bi-GRU can selectively transmit information during the learning process, including forget gates, update gates, and output gates. This effectively alleviates the gradient vanishing problem and improves the ability to understand and model complex relationships in the entire sequence.

[0144] See Figure 2 , Figure 2 The diagram shows the network structure of the QRNN-Bi-GRU fault identification and classification model. The QRNN structure consists of two parts: quantum recurrent blocks (QRBs) and network stacking. The QRB is the core component of the QRNN, which includes three major modules: data encoding, Ansatz variable quantum circuits, and partial quantum measurement. The functional division is achieved through dual quantum registers (Reg.D and Reg.H).

[0145] The QRB employs an "interleaved stacking" method, where the qubits of Reg.H operate alternately in time steps; the Bi-GRU layer performs bidirectional temporal modeling of the quantum features output by the QRNN, capturing the correlation features before and after the fault. , They are respectively Forward and backward output at any time; , They are respectively Forward and backward output at any time; , They are respectively Forward and backward output at any time; for Output at all times; for Input at any time.

[0146] Furthermore, addressing the gradient vanishing problem in the backpropagation process of classical recurrent neural networks (RNNs), where the gradient decays exponentially after multiplying the weights over multiple time steps, this scheme utilizes QRNNs based on unitary quantum circuits to achieve state evolution. The unitary transformation preserves the vector inner product and magnitude, resulting in significantly higher gradient stability during quantum state evolution compared to classical recurrent structures. Therefore, QRNN quantum recurrent neural networks can mitigate the gradient decay problem to some extent. The proposed solution constructs a hybrid architecture of QRNN quantum recurrent neural networks and Bi-GRU bidirectional recurrent units, forming complementary capabilities and mutually addressing the core shortcomings of each when operating independently, as detailed below:

[0147] QRNNs are inherently unidirectional recursive structures, capable of transmitting historical quantum state information only in forward chronological order, which presents significant limitations. On one hand, pure QRNNs can only capture temporal dependencies unidirectionally, failing to fully characterize the dynamic evolution of faults; while Bi-GRU bidirectional recurrent units, with forward and backward recurrent branches processing the sequence synchronously, effectively compensate for this deficiency. On the other hand, QRNNs output multidimensional quantum enhancement features that are independent at each time step, lacking temporal dimension correlation modeling between features; Bi-GRUs, however, can perform sequence-level encoding and fusion of these features, transforming discrete quantum features into temporal representations with contextual semantics.

[0148] In the hybrid architecture composed of the two, the modules have a clear division of labor: QRNN is responsible for feature extraction in a single time step, while the mining of long-range time-series dependencies is handled by Bi-GRU bidirectional recurrent units. This design significantly reduces the quantum circuit depth and coherence time requirements, effectively alleviating the problem of continuous accumulation of quantum decoherence errors as the time step increases.

[0149] In contrast, using Bi-GRU alone has significant drawbacks: directly inputting the original signal or manually extracted traditional features into Bi-GRU makes it difficult to fully exploit the nonlinear characteristics of transmission line fault signals. QRNN, on the other hand, can output multi-dimensional enhanced features through quantum state measurements, completely preserving the fine-grained differences in fault signals. Furthermore, although Bi-GRU alleviates the gradient vanishing problem through its gating structure, it still suffers from long-term information loss when dealing with extremely long sequences. Thanks to the quantum enhanced features extracted by QRNN, which condense the core fault information, Bi-GRU no longer needs to rely on long-range original historical data, indirectly improving the gradient decay problem under long sequences.

[0150] Therefore, this scheme combines QRNN with Bi-GRU to achieve the organic integration of local fine-grained feature extraction and global temporal dependency modeling, and completely overcomes the problems of insufficient feature extraction and one-sided temporal information mining caused by a single model.

[0151] Furthermore, the acquisition of transmission line fault data, based on the QRNN feature extraction module, involves quantum entanglement and feature extraction to obtain multidimensional quantum enhanced features, specifically including:

[0152] Obtain the fault feature vector of the fault condition, and normalize the fault feature vector to... After grading, angle encoding is performed, and its expression is as follows:

[0153] ;

[0154] ;

[0155] in: Let be the unitary matrix of the revolving door; , These are the original and normalized values ​​of the fault feature vector, respectively. This represents the initial state of a qubit; The number of qubits; It is the quantum state tensor product;

[0156] The angle-encoded input data is fed into an Ansatz variable quantum circuit. Quantum entanglement between the data register Reg.D and the historical information register Reg.H is achieved through a three-layer structure of single-quantum gates and double-quantum gates within the Ansatz variable quantum circuit, thus extracting the correlation features between the input data and historical information. Each layer contains a single quantum gate. and two quantum gates Its expression is as follows:

[0157] ;

[0158] ;

[0159] in: , , , All of these are rotation angles corresponding to the rotation gates of a single quantum bit; , They act on the first The and the first Pauli-Z operator for 1 qubit; To bypass Pauli revolving door with an axle; To bypass Pauli revolving door with an axle; for Pauli revolving door;

[0160] After completing quantum entanglement and feature extraction, a projection measurement with the Pauli-Z operator as the observable is performed on the first qubit of the data register Reg.D, outputting a single classical eigenvalue. This single classical eigenvalue is used as a one-dimensional feature corresponding to a single QRB quantum loop block. All QRB quantum loop blocks are traversed sequentially, and the one-dimensional features of each block are summarized to obtain a multi-dimensional quantum enhancement feature vector. If each QRB outputs a one-dimensional feature, then 30 time steps will output a 30-dimensional quantum enhancement feature vector.

[0161] Furthermore, the method based on multidimensional quantum enhancement features, using a Bi-GRU time-series modeling module to perform bidirectional time-series modeling to capture the correlation features before and after the fault and obtain bidirectional hidden states, specifically includes:

[0162] The first in the multidimensional quantum enhancement eigenvector Input features at each time step The hidden state of the previous time step Calculate the reset door separately and the update gate The expression for controlling the retention and updating of historical information is as follows:

[0163] ;

[0164] ;

[0165] in: It is the sigmoid activation function; , , , Both are weight matrices; , All are bias vectors;

[0166] The first in the multidimensional quantum enhancement eigenvector Input features at each time step Concealed state after resetting the door The hidden states are generated through fusion, and their expression is as follows:

[0167] ;

[0168] in: This is a candidate hidden state; It is the hyperbolic tangent activation function; , Both are weight matrices; This is element-wise multiplication; It is the bias vector;

[0169] Based on the generated candidate hidden states, the forward GRU in ascending order (i.e., forward GRU by time step) is used. (Calculation), backward GRU reverse order (i.e., backward GRU in time step) (Calculation) to obtain the forward hidden state sequence and the backward hidden state sequence respectively, concatenate them and pass them to the next time step, and obtain the bidirectional hidden state by traversing the multidimensional quantum enhancement eigenvector. .

[0170] Furthermore, the method of using a bidirectional hidden state and a fully connected classification module to perform linear transformation to output the probability distribution of various fault types, thereby completing the transmission line fault determination, specifically includes:

[0171] To enable the model to focus more on the parts of the input sequence that are more relevant to the current task and improve its performance in fault identification, an attention mechanism is embedded after the output layer of the Bi-GRU. Based on the attention mechanism, feature mapping is performed on the bidirectional hidden state to generate query vectors, key vectors, and value vectors. The attention weights are then obtained based on the query vectors, key vectors, and value vectors, as expressed below:

[0172] ;

[0173] ;

[0174] in: , , These are the query vector, key vector, and value vector, respectively. , , Both are weight matrices; Attention weights, with dimensions of ; The number of time steps; The dimension of the key vector;

[0175] We perform a weighted summation based on the attention weights to obtain the input feature representation after processing by the attention mechanism, as shown in the following expression:

[0176] ;

[0177] in: For input feature representation;

[0178] Represent the input features The input fully connected layer undergoes a linear transformation, and the probability distribution of various fault types is calculated using the cross-entropy loss function, as shown in the following expression:

[0179] ;

[0180] ;

[0181] ;

[0182] in: For each The predicted probabilities of various faults obtained by performing the Softmax operation; The original score for the fault characteristics; For the first The original score for the type of fault; For the first Fault category tags; This is the weight matrix; The total number of samples; For sample index; This is the cross-entropy loss function.

[0183] Example 2:

[0184] See Figure 3 In this embodiment, an overhead transmission line model is first built on the PSCAD / EMTDC platform. Seven types of fault modules are identified and dynamic time-domain simulation is performed. Fault samples are collected by combining historical fault data, and test and training sets are divided. Then, the hyperparameters of the fault model are used as the initial population of the algorithm through the IKOA algorithm. After operations such as gravity calculation, velocity update, position update, and solution selection, the optimal parameter optimization model is output. Next, the QRNN-Bi-GRU fault identification and classification model is trained using the training set samples. Then, the accuracy of the model is tested using the test set. If the accuracy is not met, the optimization and training process is repeated until the accuracy is met to obtain the improved QRNN-Bi-GRU fault identification and classification model. Online fault identification is then performed based on the improved QRNN-Bi-GRU fault identification and classification model combined with online fault acquisition.

[0185] In this embodiment, in order to obtain the fault feature vector of the fault condition, it is necessary to construct a fault simulation model of the transmission line. Through dynamic time-domain simulation and fault history data collection, fault feature vectors containing multiple fault conditions are obtained and divided into training set and test set.

[0186] Specifically, models are built and fault simulations are performed using the PSCAD / EMTDC platform, while fault samples are collected and processed.

[0187] PSCAD / EMTDC (Power System Computer Aided Design / Electromagnetic Transient in DC systems) is a professional software platform widely used for electromagnetic transient simulation of power systems. It boasts the following significant features and functions: 1. Powerful modeling capabilities: It provides a rich component library covering power system components such as transmission lines, transformers, circuit breakers, and various power sources, enabling easy construction of complex power system topologies; 2. Accurate electromagnetic transient simulation: Employing numerical calculation methods, it can accurately simulate rapidly changing electromagnetic transient processes in power systems, such as the changes in current and voltage waveforms at the moment of a fault. The time scale of the transient process can be accurate to the microsecond level; 3. Flexible data output and analysis: It can output time-domain waveform data of various electrical quantities, such as node voltages and branch currents, facilitating subsequent analysis and processing by users.

[0188] Furthermore, the construction of the fault simulation model for the transmission line specifically includes:

[0189] Transmission lines are mainly overhead transmission lines, and their main components include conductors, towers, and grounding devices. Depending on the voltage level and transmission scenario, transmission line structures are mainly single-circuit overhead lines and multi-circuit overhead transmission lines. To be more realistic, this embodiment selects a multi-circuit transmission line with integrated protection devices as the fault simulation model.

[0190] For overhead lines of 220kV and above with a length greater than 100km, the available distributed parameters of the fault simulation model are as follows:

[0191] ;

[0192] in: For any position on the line; , These are voltage and current vectors, respectively; For resistance; It is an inductor; Electrical conductivity; For capacitors; For frequency; The propagation constant; Characteristic impedance;

[0193] For any position on the line, the voltage and current vectors at the beginning (x=0) and end (x=1) of the line are as follows:

[0194] ;

[0195] in: It is a hyperbolic sine function; It is a hyperbolic cosine function;

[0196] To simulate the transient state of transmission line faults, the transient process of overhead lines can be described by the traveling wave in the initial stage and the traveling wave in the later stage of the transient state. Therefore, the governing equations for the transient process of transmission lines are established as follows:

[0197] ;

[0198] ;

[0199] in: For the location of the line place, time The instantaneous voltage; For the location of the line place, time The instantaneous current;

[0200] Solve the transient process control equations of the transmission line to obtain the traveling wave propagating along the line, the expression of which is as follows:

[0201] ;

[0202] ;

[0203] in: A traveling wave propagating along the path; , These are forward and reverse traveling waves, respectively. The speed of travel wave propagation;

[0204] Considering the environmental factors affecting actual transmission lines, an environmental correction is added, the expression of which is as follows:

[0205] ;

[0206] in: , , These are the corrected resistors, inductors, and capacitors for transmission lines covered by ice and snow. This represents the original cross-sectional area of ​​the transmission line; The cross-sectional area of ​​the power transmission line after it has been covered with ice and snow; This refers to the altitude above the ground. This refers to the amount of downward sag.

[0207] The expressions for setting the grounding resistance and ground modulus attenuation coefficient of transmission lines are as follows:

[0208] ; ;

[0209] in: This is the grounding resistance attenuation coefficient; This is the ground model attenuation coefficient; This represents the actual soil resistivity. Reference resistivity;

[0210] The types of transmission line faults involved in this embodiment include any one or more of the following: single-pole grounding, double-pole short circuit, line breakage, lightning strike fault, lightning interference, insulator flashover fault, and foreign object intrusion fault.

[0211] Transmission lines are usually erected in the field and are very susceptible to lightning strikes. Both lightning faults and lightning interference are caused by lightning discharges directly or indirectly striking the transmission lines. If a lightning strike breaks down the line insulators, causing a flashover, it is called a lightning fault. Due to protective devices such as surge arresters, most lightning strikes do not break down the line insulation, and this is called lightning interference.

[0212] In this embodiment, a 2.6 / 50µs double exponential wave is used to construct the lightning strike signal. When simulating a lightning strike fault, the fault is delayed by 0.1ms compared to the lightning signal. The double exponential function is as follows:

[0213] ;

[0214] in: This represents the instantaneous value of the lightning strike current. This refers to the amplitude of the lightning strike current. The decay time constant; The step time constant; It is a constant coefficient.

[0215] Furthermore, the construction of the transmission line fault simulation model involves obtaining fault feature vectors containing multiple fault conditions through dynamic time-domain simulation and historical fault data collection, and dividing them into training and testing sets, specifically including:

[0216] Historical fault signals of transmission lines, such as fault current and traveling wave signals, are collected, undergo phase-mode transformation, and line values ​​are extracted. Then, based on variational mode decomposition (VMD), the historical fault signals are decomposed into... The expression for each IMF (Intrinsic Mode Functions) modal component is as follows:

[0217] ;

[0218] ;

[0219] in: For the first One IMF modal component; , These are the instantaneous amplitude and phase, respectively. The center frequency; The unit impulse function; This is a convolution operation; This is a fault signal; The preset number of decomposition modes; The modal component number;

[0220] Key features are extracted from the IMF modal components, including energy percentage, kurtosis factor, and permutation entropy, and their expressions are as follows:

[0221] ; ;

[0222] ;

[0223] in: , , These are energy percentage, kurtosis factor, and permutation entropy, respectively. Signal duration; For mathematical expectation; To reconstruct the probability of the arrangement pattern in the phase space; Total number of patterns; For the first One IMF modal component;

[0224] The key features are concatenated to obtain the fault feature vector, which is expressed as follows:

[0225] ;

[0226] in: This represents the fault feature vector;

[0227] The fault feature vectors are divided into two sets: 70% is used as the training set and 30% is used as the test set.

[0228] In this embodiment, after constructing the QRNN-Bi-GRU fault identification and classification model, the following is also included:

[0229] The hyperparameters of the QRNN-Bi-GRU fault identification and classification model are iteratively optimized based on the IKOA-improved Kepler optimization algorithm, and the optimal hyperparameters are output.

[0230] To address the weakness of traditional recurrent neural networks, such as CNNs, in extracting nonlinear feature coupling relationships, a QRNN feature extraction module is designed. QRNN maps input features to quantum state space through quantum entanglement and unitary transformation, which can capture high-order nonlinear correlations that classical methods cannot characterize. This significantly improves the distinguishability between similar working conditions such as lightning strike faults and lightning interference or grounding faults with different resistance values, thereby improving feature quality from the source.

[0231] To address the shortcomings of traditional time series modeling schemes, such as unidirectional RNN, GRU, and LSTM, in terms of bidirectional time series traversal and long sequence modeling, a Bi-GRU time series modeling module is designed. Bi-GRU performs forward and backward time series traversal simultaneously, which can not only capture fault precursor features using historical information, but also trace back the weak signal at the fault initiation time through waveform features after the fault occurs, fully covering the entire time series correlation before, during, and after the fault occurs, thereby improving the accuracy of fault type identification and fault location.

[0232] To address the limited ability of traditional fault classification schemes, such as Support Vector Machines (SVM) and Random Forests, to represent high-dimensional temporal features, a fully connected classification module is designed. This fully connected classification module is jointly trained with QRNN and Bi-GRU modules under a unified cross-entropy loss function. Gradients can be backpropagated to all front-end modules, allowing feature extraction and temporal modeling to adaptively adjust parameters around the ultimate goal of fault classification. This achieves collaborative optimization across the entire chain, resulting in overall performance superior to the segmented combination mode of feature extraction and classifier.

[0233] Furthermore, the IKOA improved Kepler optimization algorithm specifically includes:

[0234] The Kepler optimization algorithm (KOA) is a multi-objective metaheuristic algorithm based on Kepler's laws of planetary motion. Its core inspiration comes from the physical laws governing planetary motion around the sun: each combination of parameters to be optimized is considered a "planet," and the parameter optimization objective is considered the "solar gravitational field." By simulating the orbital position updates and gravitational interactions of the planets, it searches for the Pareto optimal solution to the multi-objective optimization problem. In this embodiment, it is improved and optimized as follows:

[0235] (1) Improve the KOA search process and solution selection mechanism:

[0236] Traditional KOA algorithms use the "sun position (global optimum)" as a fixed guiding direction, which easily leads to the population getting trapped in local optima in the later stages of iteration, making it impossible to explore a better parameter space. This is especially true in multi-objective optimization, where it is difficult to generate a uniformly distributed Pareto front. Therefore, for the global search phase of the KOA algorithm, a perturbation mechanism is introduced to improve the search process, and its expression is as follows:

[0237] ;

[0238] ;

[0239] in: For the first The updated global optimal position of each planet; , The first The positions of the planets before and after the update; This is the fitness value; The position of the sun is randomly selected from an external archive library; The disturbance coefficient; A random number within the range [0, 1]; This represents the number of iterations. This represents the maximum number of iterations.

[0240] In this embodiment, most existing conventional perturbation strategies are post-position random noise superposition, that is, after the position update is completed, Gaussian noise, Cauchy noise or random perturbation of fixed amplitude is superimposed on the solution. In essence, it is a random trial without direction, with a lot of invalid search and weak connection with the information solved by the algorithm.

[0241] To address the above issues, this scheme employs a unique perturbation mechanism: First, unlike random perturbations, this mechanism is based on randomly selected non-dominated solutions from an external archive, effectively utilizing the algorithm's solution information for targeted perturbation exploration. Second, the perturbation intensity dynamically decreases with the iteration process; the perturbation amplitude is large in the early stages of iteration, enhancing global exploration capabilities, while the perturbation amplitude gradually decreases in the later stages, ensuring the accuracy of local development. Finally, the perturbation directly participates in the planetary position update calculation, integrating with the original gravity-velocity update mechanism rather than being an independent post-operation, ensuring the consistency of the algorithm's convergence logic.

[0242] The perturbation mechanism and dynamic decay factor are collaborative improvement strategies designed for the IKOA algorithm from two different dimensions: First, they jointly balance the algorithm's global exploration and local exploitation. The perturbation mechanism addresses the problem of narrow search range and rapid loss of population diversity caused by a single gravitational source from the perspective of search direction diversity, thereby enhancing the breadth of global exploration. The dynamic decay factor controls the switching between global search and local exploitation, optimizing the adjustment parameters. The dynamic change pattern of KOA solves the problems in traditional KOA. The two algorithms address several issues: first, the chaotic switching and slow convergence caused by random fluctuations; second, both employ an iterative-driven adaptive evolution mechanism to avoid logical conflicts while collaboratively controlling the breadth, speed, and fluctuation range of the algorithm's solution phase; and finally, their complementary support relationship synergistically improves algorithm performance, with a dynamic decay factor adjusting the algorithm's global exploration phase. By increasing the exploration weights, the perturbation mechanism, guided by stronger, multi-directional non-dominated solutions, expands the search space. During the phase where the algorithm tends towards local exploitation, the dynamic decay factor is adjusted... By increasing the development weight, the strength of the perturbation mechanism decreases synchronously, and only small-amplitude fine perturbations are performed to help escape tiny local optima.

[0243] When a planet is close to the sun, the algorithm focuses on optimizing the exploration operator; conversely, when a planet is far from the sun, the algorithm prioritizes optimizing the exploration operator. This principle largely relies on adjusting parameters. The value of the parameter is adjusted to improve the convergence speed of the algorithm. To mitigate the impact of random fluctuations on optimization results, a dynamic decay factor is introduced to improve the adjustment parameters during the exploration and development transition phase of the KOA algorithm. Its expression is as follows:

[0244] ;

[0245] ;

[0246] ;

[0247] in: To adjust the parameters; It is a dynamic decay factor; These are random values ​​based on a normal distribution. A random value between [0, 1]; These are the parameters for cyclic control. To control the initial value of the loop; This is the modulo operation;

[0248] In this embodiment, adjusting parameters This is the core threshold in the Kepler Optimization Algorithm (KOA) that controls the switching between global exploration and local exploitation modes. The algorithm compares the Euclidean distance between the planet (the combination of parameters to be optimized) and the sun (the current optimal solution). The size of the parameter determines the search strategy for the current iteration; while the fixed parameter Traditional methods suffer from insufficient exploration, susceptibility to local optima, and slow convergence. To address these issues, the dynamic decay factor in this scheme is resolved in the following ways:

[0249] (1) Escape local optima through a dynamic loop-assisted algorithm;

[0250] To control the initial values ​​of the loop, the total iterations are divided into multiple sub-cycles, which determine... Frequency of reset; The smaller the value, the shorter the period, and the higher the frequency of escaping local optima; The larger the value, the stronger the convergence stability; by setting this parameter and combining it with the influence of the maximum number of iterations, the algorithm can be controlled when it stalls in the later stages of iteration. The "lift-decrease" cycle causes the adjustment parameters to fluctuate dynamically, helping the algorithm to escape local optima; To be preferred, positive integers are selected, with values ​​ranging from 3 to 10. In engineering applications, values ​​of 3 and 10 are commonly used. =5.

[0251] (2) Dynamic decay adjustment algorithm search rhythm;

[0252] This scheme introduces a dynamic decay factor into the adjustment parameter to form a nonlinear exponent, causing the adjustment parameter to decrease nonlinearly and monotonically with the iteration process. In the early stages of iteration, the adjustment parameter is relatively large, enabling the algorithm to perform a fast global search; in the middle stages of iteration, the adjustment parameter decays nonlinearly, while also incorporating random values. , The effect of this is to enhance the randomness of the search and improve the efficiency of the algorithm; in the later stage of the iteration, the parameter is adjusted to decay to a smaller value, which is conducive to the algorithm to perform accurate optimization in local areas.

[0253] In summary, compared to traditional fixed-parameter and conventional linear decay strategies, the following approach is adopted... Compared to the simple linear descent method, the dynamic decay factor of this scheme performs better. First, the exponential nonlinear decay of the dynamic decay factor exhibits the characteristics of slow descent in the early stage, rapid descent in the middle stage, and steady descent in the later stage, which matches the convergence rhythm of the algorithm. The optimization efficiency and accuracy are better than linear decay. At the same time, the dynamic adjustment mechanism can maintain an appropriate population diversity throughout the iteration process, and the output Pareto front has a wider coverage and more uniform distribution. It can provide more balanced alternatives for the hyperparameter selection of the QRNN-Bi-GRU model, which is more in line with the multi-objective decision-making needs of engineering practice.

[0254] Traditional KOA is a single-objective optimization algorithm that lacks a mechanism for storing multi-objective non-dominated solutions. It cannot handle multi-objective conflict scenarios and can only output a single optimal solution, failing to provide multi-dimensional decision-making options for practical engineering. Therefore, an external archive mechanism is introduced to simultaneously retain multiple non-dominated solutions, forming a complete Pareto front, and providing alternative locations for stochastic optimal solutions, as detailed below:

[0255] If the solution No inferior solution on all objective functions And solve It performs strictly better than the solution on at least one objective function. The term "explanation" is sometimes used to describe the process of solving problems. Dominate , recorded as The basis for judgment is:

[0256] ;

[0257] in: To solve In the The values ​​that can be obtained on the objective function; To solve In the The values ​​that can be obtained on the objective function;

[0258] As the iteration process progresses, the number of non-dominated solutions will accumulate and may exceed the pre-set storage capacity. Therefore, it is necessary to optimize the solution filtering mechanism to remove redundant solutions based on the density of the region where the solution is located. The expression is as follows:

[0259] ;

[0260] in: Density; The number of objective functions; For the first The normalized value of the fitness value of each objective function.

[0261] (2) Optimize the hyperparameters of the fault identification and classification model:

[0262] Furthermore, the IKOA-based improved Kepler optimization algorithm iteratively optimizes the hyperparameters of the QRNN-Bi-GRU fault identification and classification model, outputting the optimal hyperparameters, specifically including:

[0263] The parameters to be optimized are combined and encoded into planetary position vectors using the IKOA improved Kepler optimization algorithm. ,in The total number of parameters; the combination of parameters to be optimized includes the angle encoding rotation angle of the QRNN quantum recurrent neural network, the circuit parameters and measurement parameters of the Ansatz variable quantum circuit; and the gating parameters, hidden layer parameters, and training parameters of the Bi-GRU.

[0264] Set up an optimization multi-objective function, the expression of which is as follows:

[0265] ;

[0266] ;

[0267] ;

[0268] ;

[0269] in: , , All represent optimization objectives, namely, fault classification error rate, quantum circuit complexity, and total model execution time; , These represent the number of correctly classified samples and the total number of samples, respectively. , These represent the circuit depth and required quantum coherence time for the Ansatz variable quantum circuit, respectively. , All are coefficient weights; , These represent the running time of the QRNN quantum recurrent neural network and the backpropagation time of the Bi-GRU bidirectional recurrent gate unit, respectively.

[0270] In this embodiment, the three optimization objectives and the model architecture form a one-to-one correspondence: The final output performance corresponding to the fully connected classification module, Corresponding to the core constraints of the QRNN quantum feature extraction module, This corresponds to the computational efficiency of the Bi-GRU time series modeling module. The target system is perfectly aligned with the three-layer hybrid architecture of this scheme, rather than a generalized combination of multiple targets. The optimization process can accurately target the core parameters of each module, and the optimization efficiency and final results are superior to general target combinations.

[0271] The three optimization objectives correspond to three core requirements of power operation and maintenance: diagnostic accuracy to ensure grid reliability, operational efficiency to ensure rapid online fault handling, and quantum overhead to ensure the feasibility of the solution under existing hardware conditions. The final Pareto solution set can also be flexibly adapted to different application scenarios: high-precision offline analysis scenarios can be selected. The optimal solution for online real-time diagnostic scenarios can be selected. The optimal solution, which can be selected in scenarios where quantum hardware is constrained. The optimal solution. Compared to single-objective optimization, which can only output a fixed set of parameters, this system has significantly improved scenario adaptability.

[0272] Meanwhile, all three optimization objectives in this scheme are incorporated into the mandatory constraints of the optimization process, and there is no selective design. Together, they constitute a complete multi-objective optimization system, the core logic of which is as follows:

[0273] The three objectives cannot be globally optimal at the same time: improving classification accuracy usually requires deeper quantum circuits and more hidden layer nodes, which increases the complexity of quantum circuits and running time; compressing circuit depth and simplifying network structure can reduce the time consumption, but will sacrifice diagnostic accuracy.

[0274] Achieving cooperative balance through Pareto optimality mechanism: The role of the IKOA optimization algorithm is to search for a non-dominated solution set (Pareto optimal solution set) among three conflicting objectives: a solution will dominate and eliminate another solution only if it is not inferior to another solution in all objectives and is strictly better in at least one objective, thus obtaining multiple feasible parameter schemes.

[0275] Set initialization parameters, including population size, maximum number of iterations, external archive capacity, and cycle period; randomly generate the initial position vectors of each planet. ; Calculate the multi-objective function values ​​for each planet. Initialize the external archive library, and then calculate the gravitational force between the planet and the sun and the planet's orbital velocity based on the positional relationship between the planet and the sun. The expressions are as follows:

[0276] ;

[0277] ;

[0278] ;

[0279] in: The gravitational pull between the planet and the sun; , The masses of the Sun and the planets are respectively determined based on the objective function value; It is the gravitational coefficient; The Euclidean distance between the planet and the Sun; It is a very small positive number; , These are the current planetary orbital speed and the updated planetary orbital speed, respectively. For random numbers The controllable planetary orbit direction coefficient;

[0280] The global search mechanism based on the IKOA-improved Kepler optimization algorithm performs planetary position updates, and the non-dominated solutions generated during the iteration process are stored in the external archive based on the external archive mechanism and solution filtering mechanism.

[0281] (3) Optimal solution decision:

[0282] When the algorithm iterates to the preset maximum number of iterations, it outputs the Pareto front solution set, calculates the distance of each solution in the solution set to the ideal point, selects the solution with the smallest distance as the optimal solution, and obtains the optimal hyperparameters.

[0283] See Figure 4 , Figure 4 The confusion matrix for the test set shows the number of samples that were actually classified as a certain type of fault, represented by values ​​such as "28", "27", and "20". The "100%" on the far right of each row indicates that the model's accuracy in identifying the corresponding real fault category is 100%, meaning all samples of that type of real fault were correctly identified. The "100%" in the bottom row indicates that the model's accuracy in identifying each predicted fault category is also 100%. The results in the figure show that all real fault samples were correctly identified, and the overall accuracy across all categories reached 100%, with almost no category confusion. This demonstrates the high effectiveness of the improved QRNN-Bi-GRU fault identification and classification model, making it reliably applicable to relevant fault identification scenarios.

[0284] See Figure 5 , Figure 5 The accuracy iteration curve represents how the accuracy changes with the number of iterations during model training. The polar angle represents the number of iterations, with the angle increasing along the circumference corresponding to an increase in the number of iterations. The polar radius represents the training accuracy, with the value increasing from the center outwards, corresponding to an accuracy from 0% to 100%. In the initial stage, the polar radius is relatively short, indicating that the accuracy is low in the early stages of model training, and the model is in the initial stage of learning fault features. In the intermediate iteration stage, the polar radius grows rapidly, reflecting that the model is continuously learning during the iteration process, the accuracy is greatly improved, and the ability to capture fault features is constantly enhanced. In the later stage, the polar radius approaches its maximum value, and the curve gradually smooths out, indicating that after sufficient iterations, the accuracy tends to stabilize, the model has learned the rules of fault identification well, and the training process has good convergence.

[0285] Example 3:

[0286] See Figure 6 A transmission line fault diagnosis system based on a QRNN-Bi-GRU model, which is applied to the method described in Example 1, the system comprising:

[0287] Fault model construction module 1 is used to construct a QRNN-Bi-GRU fault identification and classification model based on QRNN quantum recurrent neural network and Bi-GRU bidirectional cyclic gate unit; the QRNN-Bi-GRU fault identification and classification model includes a QRNN feature extraction module, a Bi-GRU time series modeling module and a fully connected classification module.

[0288] Fault determination module 2 is used to acquire transmission line fault data and perform quantum entanglement and feature extraction based on the QRNN feature extraction module to obtain multi-dimensional quantum enhanced features;

[0289] Based on the multidimensional quantum enhancement feature, bidirectional temporal modeling is performed through the Bi-GRU temporal modeling module to obtain bidirectional hidden states;

[0290] Based on the bidirectional hidden state, a linear transformation is performed through a fully connected classification module to output the probability distribution of various faults, thus completing the fault determination of transmission lines.

[0291] Furthermore, for the specific implementation steps of the above modules, please refer to the corresponding description in Example 1, which will not be repeated here.

[0292] Example 4:

[0293] See Figure 7 A transmission line fault diagnosis device based on the QRNN-Bi-GRU model, the device including a processor 3 and a memory 4;

[0294] The memory 4 is used to store computer program code 41 and transmit the computer program code 41 to the processor 3;

[0295] The processor 3 is used to execute the transmission line fault diagnosis method based on the QRNN-Bi-GRU model described in Embodiment 1 or 2 according to the instructions in the computer program code 41.

[0296] This embodiment also includes a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are executed on a computer, the transmission line fault diagnosis method based on the QRNN-Bi-GRU model described in Embodiment 1 or 2 is implemented.

[0297] Generally, computer instructions for implementing the method of the present invention can be carried on any combination of one or more computer-readable storage media. Non-transitory computer-readable storage media can include any computer-readable medium except for the signal itself which is temporarily propagating.

[0298] Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0299] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. In particular, Python, suitable for neural network computation, and platform frameworks based on TensorFlow, PyTorch, etc., can be used. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or connected to an external computer (e.g., via the Internet using an Internet service provider).

[0300] For details regarding the aforementioned equipment and non-transitory computer-readable storage media, please refer to the specific description of a transmission line fault diagnosis method based on the QRNN-Bi-GRU model and its beneficial effects, which will not be repeated here.

[0301] Although embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for fault diagnosis of transmission lines based on a QRNN-Bi-GRU model, characterized in that, include: A QRNN-Bi-GRU fault identification and classification model is constructed based on QRNN quantum recurrent neural network and Bi-GRU bidirectional cyclic gate unit; the QRNN-Bi-GRU fault identification and classification model includes a QRNN feature extraction module, a Bi-GRU time series modeling module and a fully connected classification module. Data on power transmission line faults are acquired, and quantum entanglement and feature extraction are performed based on the QRNN feature extraction module to obtain multidimensional quantum enhanced features. Based on the multidimensional quantum enhancement feature, bidirectional temporal modeling is performed through the Bi-GRU temporal modeling module to obtain bidirectional hidden states; Based on the bidirectional hidden state, a linear transformation is performed through a fully connected classification module to output the probability distribution of various faults, thus completing the fault determination of transmission lines.

2. The transmission line fault diagnosis method based on the QRNN-Bi-GRU model according to claim 1, characterized in that: The QRNN feature extraction module adopts an sQRNN interleaved quantum recurrent neural network, which is composed of multiple sets of QRB quantum recurrent blocks stacked alternately. A single QRB quantum loop block integrates a data encoding unit, an Ansatz variable quantum circuit unit, and a portion of the quantum measurement unit; The Bi-GRU temporal modeling module includes a forward GRU subnetwork, a backward GRU subnetwork, and a hidden state splicing layer; An attention mechanism layer is embedded after the output layer of the Bi-GRU time series modeling module.

3. The transmission line fault diagnosis method based on the QRNN-Bi-GRU model according to claim 1, characterized in that: The acquisition of transmission line fault data, based on the QRNN feature extraction module, involves quantum entanglement and feature extraction to obtain multidimensional quantum enhanced features, specifically including: Obtain the fault feature vector of the fault condition, and normalize the fault feature vector to... After grading, angle encoding is performed, and its expression is as follows: ； ； in: Let be the unitary matrix of the revolving door; , These are the original and normalized values ​​of the fault feature vector, respectively. This represents the initial state of a qubit; The number of qubits; It is the quantum state tensor product; The angle-encoded input data is fed into an Ansatz variable quantum circuit. Quantum entanglement between the data register Reg.D and the historical information register Reg.H is achieved through a three-layer structure of single-quantum gates and double-quantum gates within the Ansatz variable quantum circuit, thus extracting the correlation features between the input data and historical information. Each layer contains a single quantum gate. and two quantum gates Its expression is as follows: ； ； in: , , , All of these are rotation angles corresponding to the rotation gates of a single quantum bit; , respectively acting on the first The and the first Pauli-Z operator for 1 qubit; To bypass Pauli revolving door with an axle; To bypass Pauli revolving door with an axle; for Pauli revolving door; After completing quantum entanglement and feature extraction, a projection measurement with the Pauli-Z operator as the observable is performed on the first qubit of the data register Reg.D, and a single classical eigenvalue is output. The single classical eigenvalue is used as the one-dimensional feature corresponding to a single QRB quantum loop block. All QRB quantum loop blocks are traversed in turn, and the one-dimensional features are summarized to obtain a multi-dimensional quantum enhancement eigenvector.

4. The transmission line fault diagnosis method based on the QRNN-Bi-GRU model according to claim 3, characterized in that: The method of obtaining bidirectional hidden states by performing bidirectional temporal modeling through the Bi-GRU temporal modeling module based on multidimensional quantum enhancement features specifically includes: The first in the multidimensional quantum enhancement eigenvector Input features at each time step The hidden state of the previous time step Calculate the reset door separately and Update Gate The expression for controlling the retention and updating of historical information is as follows: ； ； in: It is the sigmoid activation function; , , , Both are weight matrices; , All are bias vectors; The first in the multidimensional quantum enhancement eigenvector Input features at each time step Concealed state after resetting the door The hidden states are generated through fusion, and their expression is as follows: ； in: This is a candidate hidden state; It is the hyperbolic tangent activation function; , Both are weight matrices; This is element-wise multiplication; It is the bias vector; Based on the generated candidate hidden states, the multidimensional quantum enhancement feature vectors are traversed in forward GRU order and backward GRU order to obtain bidirectional hidden states. .

5. The transmission line fault diagnosis method based on the QRNN-Bi-GRU model according to claim 4, characterized in that: The method, based on bidirectional hidden states, performs linear transformation through a fully connected classification module to output the probability distribution of various fault types, thereby completing the fault determination of transmission lines. Specifically, this includes: Based on the attention mechanism, feature mapping is performed on the bidirectional hidden state to generate query vector, key vector, and value vector. Then, based on the query vector, key vector, and value vector, the attention weights are obtained, and their expressions are as follows: ； ； in: , , These are the query vector, key vector, and value vector, respectively. , , Both are weight matrices; Attention weights; The number of time steps; The dimension of the key vector; We perform a weighted summation based on the attention weights to obtain the input feature representation after processing by the attention mechanism, as shown in the following expression: ； in: For input feature representation; Represent the input features The input fully connected layer undergoes a linear transformation, and the probability distribution of various fault types is calculated using the cross-entropy loss function, as shown in the following expression: ； ； ； in: For each The predicted probabilities of various faults obtained by performing the Softmax operation; The original score for the fault characteristics; For the first The original score for the type of fault; For the first Fault category tags; This is the weight matrix; The total number of samples; For sample index; This is the cross-entropy loss function.

6. The transmission line fault diagnosis method based on the QRNN-Bi-GRU model according to claim 2, characterized in that: The acquisition of the fault feature vector of the fault condition specifically includes: A fault simulation model of a transmission line is constructed, historical fault signals of the transmission line are collected, and based on variational mode decomposition, the historical fault signals are decomposed into... The expression for each IMF modal component is as follows: ； ； in: For the first One IMF modal component; , These are the instantaneous amplitude and phase, respectively. The center frequency; The unit impulse function; This is a convolution operation; This is a fault signal; The preset number of decomposition modes; The modal component number; Key features are extracted from the IMF modal components, including energy percentage, kurtosis factor, and permutation entropy, and their expressions are as follows: ； ； ； in: , , These are energy percentage, kurtosis factor, and permutation entropy, respectively. Signal duration; For mathematical expectation; To reconstruct the probability of the arrangement pattern in the phase space; Total number of patterns; For the first One IMF modal component; The key features are concatenated to obtain the fault feature vector, which is expressed as follows: ； in: This is the fault feature vector.

7. The transmission line fault diagnosis method based on the QRNN-Bi-GRU model according to claim 1, characterized in that: The construction of the fault simulation model for the transmission line specifically includes: The available distributed parameters for the fault simulation model are set as follows: ； in: For any position on the line; , These are voltage and current vectors, respectively; For resistance; It is an inductor; Electrical conductivity; For capacitors; For frequency; The propagation constant; Characteristic impedance; For any point on the line, the voltage and current vectors at the beginning and end of the line are as follows: ； in: It is a hyperbolic sine function; It is a hyperbolic cosine function; To simulate the transient state of a transmission line fault, the transient process control equations for the transmission line are established, as follows: ； ； in: For the location of the line place, time The instantaneous voltage; For the location of the line place, time The instantaneous current; Solve the transient process control equations of the transmission line to obtain the traveling wave propagating along the line, the expression of which is as follows: ； ； in: A traveling wave propagating along the path; , These are forward and reverse traveling waves, respectively. The speed of travel wave propagation; Considering the environmental factors affecting actual transmission lines, an environmental correction is added, the expression of which is as follows: ； in: , , These are the corrected resistors, inductors, and capacitors for transmission lines covered by ice and snow. This represents the original cross-sectional area of ​​the transmission line; The cross-sectional area of ​​the power transmission line after it has been covered with ice and snow; This refers to the altitude above the ground. This refers to the amount of downward sag. The expressions for setting the grounding resistance and ground modulus attenuation coefficient of transmission lines are as follows: ； ； in: This is the grounding resistance attenuation coefficient; This is the ground model attenuation coefficient; This represents the actual soil resistivity. Reference resistivity; A lightning strike signal is constructed using a double exponential wave to simulate a lightning fault. The double exponential function is as follows: ； in: This represents the instantaneous value of the lightning strike current. This refers to the amplitude of the lightning strike current. The decay time constant; The step time constant; It is a constant coefficient.

8. The transmission line fault diagnosis method based on the QRNN-Bi-GRU model according to claim 1, characterized in that: After constructing the QRNN-Bi-GRU fault identification and classification model, the following is also included: The hyperparameters of the QRNN-Bi-GRU fault identification and classification model are iteratively optimized based on the IKOA-improved Kepler optimization algorithm, and the optimal hyperparameters are output. The IKOA improved Kepler optimization algorithm specifically includes: To improve the global search phase of the KOA algorithm, a perturbation mechanism is introduced, as shown in the following expression: ； ； in: For the first The updated global optimal position of each planet; , The first The positions of the planets before and after the update; This is the fitness value; The position of the sun is randomly selected from an external archive library; The disturbance coefficient; A random number within the range [0, 1]; This represents the number of iterations. This represents the maximum number of iterations. For the exploration and development transition phase of the KOA algorithm, a dynamic decay factor is introduced to improve the adjustment parameters, and its expression is as follows: ； ； ； in: To adjust the parameters; It is a dynamic decay factor; These are random values ​​based on a normal distribution. A random value between [0, 1]; These are the parameters for cyclic control. To control the initial value of the loop; This is the modulo operation; An external archive mechanism is introduced to retain multiple non-dominated solutions and provide alternative locations for randomized optimal solutions, as detailed below: If the solution No inferior solution on all objective functions And solve It performs strictly better than the solution on at least one objective function. The term "explanation" is sometimes used to describe the process of solving problems. Dominate , recorded as The basis for judgment is: ； in: To solve In the The values ​​that can be obtained on the objective function; To solve In the The values ​​that can be obtained on the objective function; The solution filtering mechanism is optimized by removing redundant solutions based on the density of the region where the solution resides. Its expression is as follows: ； in: Density; The number of objective functions; For the first The normalized value of the fitness value of each objective function.

9. The transmission line fault diagnosis method based on the QRNN-Bi-GRU model according to claim 8, characterized in that: The improved Kepler optimization algorithm based on IKOA iteratively optimizes the hyperparameters of the QRNN-Bi-GRU fault identification and classification model, outputting the optimal hyperparameters, specifically including: The parameters to be optimized are combined and encoded into planetary position vectors using the IKOA improved Kepler optimization algorithm. ,in The total number of parameters; the combination of parameters to be optimized includes the angle encoding rotation angle of the QRNN quantum recurrent neural network, the circuit parameters and measurement parameters of the Ansatz variable quantum circuit; and the gating parameters, hidden layer parameters, and training parameters of the Bi-GRU. Set up an optimization multi-objective function, the expression of which is as follows: ； ； ； ； in: , , All are optimization objectives; , These represent the number of correctly classified samples and the total number of samples, respectively. , These represent the circuit depth and required quantum coherence time for the Ansatz variable quantum circuit, respectively. , All are coefficient weights; , These represent the running time of the QRNN quantum recurrent neural network and the backpropagation time of the Bi-GRU bidirectional recurrent gate unit, respectively. Set initialization parameters and randomly generate initial position vectors for each planet. ; Calculate the multi-objective function values ​​for each planet. Based on the positional relationship between the planet and the Sun, the gravitational force between the planet and the Sun and the planet's orbital velocity are calculated, and the expressions are as follows: ； ； ； in: The gravitational pull between the planet and the sun; , These are the masses of the Sun and the planets, respectively. It is the gravitational coefficient; The Euclidean distance between the planet and the Sun; It is a very small positive number; , These are the current planetary orbital speed and the updated planetary orbital speed, respectively. For random numbers The controllable planetary orbit direction coefficient; The global search mechanism based on the IKOA-improved Kepler optimization algorithm performs planetary position updates, and the non-dominated solutions generated during the iteration process are stored in the external archive based on the external archive mechanism and solution filtering mechanism. When the algorithm iterates to the preset maximum number of iterations, it outputs the Pareto front solution set, calculates the distance of each solution in the solution set to the ideal point, selects the solution with the smallest distance as the optimal solution, and obtains the optimal hyperparameters.

10. A transmission line fault diagnosis system based on a QRNN-Bi-GRU model, characterized in that, The system is applied to the method according to any one of claims 1-9, the system comprising: The fault model construction module (1) is used to construct a QRNN-Bi-GRU fault identification and classification model based on QRNN quantum recurrent neural network and Bi-GRU bidirectional cyclic gate unit; the QRNN-Bi-GRU fault identification and classification model includes a QRNN feature extraction module, a Bi-GRU time series modeling module and a fully connected classification module. The fault determination module (2) is used to acquire transmission line fault data, and to perform quantum entanglement and feature extraction based on the QRNN feature extraction module to obtain multidimensional quantum enhanced features; Based on the multidimensional quantum enhancement feature, bidirectional temporal modeling is performed through the Bi-GRU temporal modeling module to obtain bidirectional hidden states; Based on the bidirectional hidden state, a linear transformation is performed through a fully connected classification module to output the probability distribution of various faults, thus completing the fault determination of transmission lines.

Images