A power grid bottom guarantee communication fault pre-judgment method based on an intelligent algorithm

Abstract

The application provides a power grid bottom guarantee communication fault pre-judgment method based on an intelligent algorithm, and belongs to the technical field of power grid bottom guarantee communication.The application carries out wavelet packet decomposition and singular value decomposition noise reduction processing on an initial state parameter set through a dynamic topological graph structure, adopts principal component analysis and local linear embedding manifold learning for feature dimension reduction, inputs a compact feature vector into a graph attention time sequence fusion prediction model to output a fault risk score, uses an adaptive integrated extreme learning machine for secondary verification to obtain a refined fault probability distribution, establishes a double-layer game optimization model to solve an optimal communication resource allocation scheme, generates virtual samples to update model parameters according to the fault probability distribution and type judgment result, adjusts node transmission power and channel allocation strategies according to an early warning level, and triggers communication path reconstruction, so that the technical problem of low accuracy of power grid bottom guarantee communication fault pre-judgment is solved.

Description

Technical Field

[0001] This invention belongs to the field of power grid backup communication technology, specifically, it relates to a power grid backup communication fault pre-judgment method based on intelligent algorithms. Background Technology

[0002] As a critical infrastructure of the power system, the power grid backup communication system relies on traditional fault prediction technologies, primarily employing threshold-based rule-based judgment methods and single machine learning model prediction methods. These methods monitor operating parameters such as voltage fluctuations and current distortions, triggering alarms when these parameters exceed preset thresholds. However, due to the complex operating environment and strong coupling between nodes in the power grid backup communication system, traditional threshold-based judgment methods struggle to capture the inter-node correlations, and single prediction models cannot effectively handle the dynamic topology and temporal dependencies of the communication network. This results in numerous false alarms and missed alarms during fault prediction. Furthermore, existing technologies lack mechanisms for optimizing communication resource allocation, failing to adjust resource allocation strategies promptly after detecting fault risks, leading to rapid fault propagation and a wide impact. In short, existing technologies suffer from low accuracy in fault prediction during power grid backup communication. Summary of the Invention

[0003] In view of this, the present invention provides a method for predicting power grid backup communication faults based on intelligent algorithms, which can solve the technical problem of low accuracy in predicting power grid backup communication faults in the prior art.

[0004] This invention is implemented as follows: It provides a method for predicting power grid backup communication faults based on intelligent algorithms. This includes collecting operational data from each node of the power grid backup communication system and modeling the communication network as a dynamic topology graph. Voltage fluctuation amplitude, current harmonic distortion rate, signal transmission delay, bit error rate, and link load rate of each node are extracted as an initial state parameter set. Wavelet packet decomposition and singular value decomposition are performed on the initial state parameter set to obtain a denoised state parameter set. Principal component analysis and local linear embedding manifold learning are then performed on the denoised state parameter set to obtain a compact feature vector set. This compact feature vector set is input into a graph attention temporal fusion prediction model to output a fault risk score vector. When the fault risk score exceeds a threshold, an adaptive ensemble extreme learning machine is activated for secondary verification to obtain a refined fault probability distribution. A communication resource allocation game optimization model is established to solve for the optimal communication resource allocation scheme. The fault type is determined based on the refined fault probability distribution, and virtual fault samples are generated to update the model parameters. The transmit power and channel allocation strategy of each node are adjusted according to the optimal communication resource allocation scheme, and a fault warning level classification rule is established to trigger the communication path reconstruction process.

[0005] The dynamic topology graph structure refers to abstracting each communication device in the power grid backup communication network as a node of the graph, and the physical or logical communication connections between devices as edges of the graph. The weight of the edge is determined by the ratio of link bandwidth and real-time traffic.

[0006] Wherein, the voltage fluctuation amplitude is the ratio of the difference between the maximum and minimum node voltages during the monitoring period to the rated voltage, and the current harmonic distortion rate is the square root of the sum of the squares of the effective values ​​of each harmonic current divided by the effective value of the fundamental current.

[0007] Specifically, the wavelet packet decomposition uses the Daubechies4 wavelet basis function to perform a five-level wavelet packet transform on the signal to decompose the signal into 32 frequency bands. The singular value decomposition performs singular value decomposition on the coefficient matrix of each frequency band, retaining the left singular vectors corresponding to the top 10 largest singular values ​​to form the signal subspace.

[0008] The process includes extracting the signal subspace features through singular value decomposition, subtracting the noise power spectrum estimate from the noise power spectrum estimate in the frequency domain using spectral subtraction, and then performing an inverse transform to obtain the denoised signal. The signal-to-noise ratio gain between the reconstructed signal and the original signal is calculated, and when the signal-to-noise ratio gain is less than 12, the number of empirical mode decomposition layers is increased to eight.

[0009] Specifically, the principal component analysis dimensionality reduction involves converting related variables into linearly independent principal components through orthogonal transformation and retaining principal components with a cumulative contribution rate of 95%. The mutual information criterion is used to calculate the correlation between features and to delete redundant features with mutual information values ​​greater than 0.85.

[0010] The graph attention temporal fusion prediction model includes a graph attention layer that calculates the attention weight matrix between nodes, a long short-term memory layer that extracts temporal dependency features from the historical state sequence of each node, a graph pooling layer that aggregates neighborhood node information through summation pooling, a gated fusion layer that uses a gating mechanism to fuse local and global features, and a temporal convolution layer that uses dilated convolution to capture dependencies over long time spans.

[0011] The attention weights in the graph attention layer are calculated by concatenating the node feature vectors and mapping them to scalars through a single-layer feedforward neural network. The scalars are then normalized using the LeakyReLU activation function and softmax to obtain the attention coefficients. The attention coefficients are modulated based on the product of the signal transmission delay and the link load rate in the noise reduction state parameter set.

[0012] The threshold of 0.72 is determined by selecting the failure risk score corresponding to the maximum Youden index by plotting the subject operating characteristic curve on the validation set. The Youden index is the difference between the true positive rate and the false positive rate.

[0013] The adaptive ensemble extreme learning machine consists of 5 base learners with 100, 150, 200, 250 and 300 hidden layer nodes respectively. The activation functions are Sigmoid, hyperbolic tangent, ReLU, Swish and Gaussian radial basis functions respectively. The inverse of the root mean square error of each base learner is normalized on the validation set and used as the ensemble weight coefficient.

[0014] Before the adaptive ensemble extreme learning machine performs secondary verification, the input weights and hidden layer biases are optimized using a differential evolution algorithm. The population size is 50, the mutation factor is 0.8, the crossover probability is 0.9, and the number of iterations is 100. The incremental learning mechanism updates the new sample features and the hidden layer output matrix using recursive least squares after receiving new samples.

[0015] The communication resource allocation game optimization model includes an upper-layer model and a lower-layer model. The upper-layer model aims to minimize the overall network fault propagation rate, while the lower-layer model aims to maximize the communication reliability of each node. It takes refined fault probability distribution, link load rate, and signal transmission delay as input parameters and obtains the optimal communication resource allocation scheme by solving the two-layer game equilibrium.

[0016] The two-layer game equilibrium is solved by an iterative algorithm. The upper-layer model optimizes the overall network fault propagation rate based on the current communication reliability of each node, and the lower-layer model adjusts the communication reliability of each node based on the optimized overall network fault propagation rate. The iteration terminates when the change in the objective function value between two consecutive iterations is less than 0.001 or the number of iterations reaches 50.

[0017] The virtual fault sample generation adopts synthetic minority class oversampling technology. For each fault sample, five nearest neighbor fault samples are found in the feature space. A point is randomly selected on the line connecting the fault sample and a randomly selected nearest neighbor fault sample as a synthetic sample. The weight coefficient of the fault sample is set to 50 times that of the normal sample using cost-sensitive learning.

[0018] The fault types include four categories: communication interruption, signal attenuation, protocol anomaly, and equipment overload. The dominant fault type is determined based on the maximum probability of each category in the refined fault probability distribution.

[0019] Specifically, the fault warning level classification rule is to mark a finely distributed fault probability in the range of 0.4 to 0.6 as a low-risk warning, in the range of 0.6 to 0.8 as a medium-risk warning, and in the range of greater than 0.8 as a high-risk warning. For high-risk warning nodes, a communication path reconstruction process is triggered to transfer communication traffic to a backup link.

[0020] This invention addresses the technical problem of low accuracy in predicting power grid communication failures by constructing a graph attention-based temporal fusion prediction model combined with a two-layer verification mechanism using an adaptive ensemble extreme learning machine (LAM), and introducing a communication resource allocation game theory optimization model for dynamic resource adjustment. Specifically, the graph attention-based temporal fusion prediction model models the communication network as a dynamic topological graph structure. It captures spatial association features between nodes through a graph attention layer and extracts temporal dependency features through a long short-term memory layer, achieving simultaneous modeling of the network topology and its temporal evolution. The LAM employs a diverse combination of base learners to perform secondary verification of the initial prediction results, reducing the misjudgment rate of a single model. The communication resource allocation game theory optimization model finds the optimal balance between minimizing the fault propagation rate and maximizing node reliability through a two-layer game equilibrium solution, achieving dynamic optimization of resource allocation. In summary, this invention solves the technical problem of low accuracy in predicting power grid communication failures mentioned in the background section. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the structure of the graph attention temporal fusion prediction model.

[0023] Figure 3 The time-series evolution diagram of fault risk scoring for each node.

[0024] Figure 4 Reconstruct the topology for communication paths of high-risk nodes.

[0025] Figure 5 A comparison chart of node reliability before and after communication path reconstruction. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1 The diagram shown is a flowchart of a power grid backup communication fault pre-judgment method based on intelligent algorithms provided by the present invention. This method includes the following steps:

[0028] S1. Collect the operating data of each node in the power grid backup communication system, model the communication network as a dynamic topology graph structure, where nodes represent communication devices and edges represent communication links, and extract the voltage fluctuation amplitude, current harmonic distortion rate, signal transmission delay, bit error rate and link load rate of each node as the initial state parameter set.

[0029] S2. Perform wavelet packet five-level decomposition on the initial state parameter set, extract the signal subspace features by singular value decomposition of the coefficients of each frequency band after decomposition, reconstruct the signal after suppressing background noise by spectral subtraction, calculate the signal-to-noise ratio gain between the reconstructed signal and the original signal, and increase the number of empirical mode decomposition layers to eight when the signal-to-noise ratio gain is less than 12 to obtain the denoised state parameter set.

[0030] S3. Perform principal component analysis and linear dimensionality reduction on the noise reduction state parameter set, retain the principal components with a cumulative contribution rate of 95%, calculate the correlation between each feature using the mutual information criterion, delete redundant features with mutual information values ​​greater than 0.85, and maintain the intrinsic structure of the data through local linear embedding manifold learning to obtain a compact feature vector set.

[0031] S4. Input the compact feature vector set into the graph attention temporal fusion prediction model. The model outputs the fault risk score vector of each node. When the fault risk score of any node is greater than the threshold of 0.72, start the adaptive ensemble extreme learning machine for secondary verification to obtain the refined fault probability distribution.

[0032] S5. Establish a game-theoretic optimization model for communication resource allocation. The upper-level model aims to minimize the overall network fault propagation rate, while the lower-level model aims to maximize the communication reliability of each node. The refined fault probability distribution, link load rate, and signal transmission delay are used as input parameters. The optimal communication resource allocation scheme is obtained by solving the two-level game equilibrium.

[0033] S6. Determine the fault type based on the refined fault probability distribution. When the fault probability is greater than 0.68 and the fault type is communication interruption, generate virtual fault samples to expand the training database. Use cost-sensitive learning to set the weight coefficient of the fault samples to 50 times that of the normal samples and update the parameters of the graph attention temporal fusion prediction model.

[0034] S7. Adjust the transmit power and channel allocation strategy of each node according to the optimal communication resource allocation scheme, establish a fault warning level classification rule, mark the fine fault probability distribution in the range of 0.4 to 0.6 as low risk warning, in the range of 0.6 to 0.8 as medium risk warning, and greater than 0.8 as high risk warning, trigger the communication path reconstruction process for high risk warning nodes, and transfer the communication traffic to the backup link.

[0035] The dynamic topology graph structure abstracts each communication device in the power grid backup communication network as a node of the graph, and the physical or logical communication connections between devices as edges of the graph. The weight of the edge is determined by the ratio of link bandwidth and real-time traffic. The dynamic topology graph structure is dynamically updated over time to reflect changes in network status.

[0036] The voltage fluctuation amplitude is the ratio of the difference between the maximum and minimum node voltages during the monitoring period to the rated voltage, expressed as a percentage (%). The current harmonic distortion rate is the square root of the sum of the squares of the effective values ​​of all harmonic currents divided by the effective value of the fundamental current, expressed as a percentage (%). The signal transmission delay is the propagation time of a data packet from the source node to the target node, expressed as milliseconds (ms). The bit error rate is the number of transmitted erroneous bits divided by the total number of transmitted bits, dimensionless. The link load rate is the current link traffic divided by the link's rated capacity, expressed as a percentage (%).

[0037] The five-level wavelet packet decomposition uses the Daubechies4 wavelet basis function to perform a five-level wavelet packet transform on the signal, decomposing it into 32 frequency bands, each containing signal components within a different frequency range. The singular value decomposition (SVD) process for extracting signal subspace features involves performing SVD on the coefficient matrix of each frequency band, retaining the left singular vectors corresponding to the top 10 largest singular values ​​to form the signal subspace. The spectral subtraction method subtracts the estimated noise power spectrum from the power spectrum of the noisy signal in the frequency domain, followed by an inverse transform to obtain the denoised signal. The signal-to-noise ratio (SNR) gain is the ratio of the denoised signal power to the noise power divided by the ratio of the original signal power to the noise power, expressed in dB. The empirical mode decomposition (EMD) is a time-frequency analysis method that adaptively decomposes a non-stationary signal into several eigenmode functions and residual components.

[0038] The principal component analysis transforms correlated variables into linearly independent principal components through orthogonal transformation, with the principal components arranged in descending order of variance contribution rate. The cumulative contribution rate is the sum of the variances of the top few principal components divided by the total variance. Mutual information is an information-theoretic index that measures the degree of interdependence between two random variables; a larger value indicates a stronger correlation. Locally linear embedding manifold learning is a nonlinear dimensionality reduction algorithm that preserves the local neighborhood structure of the data, mapping high-dimensional data to a low-dimensional space by optimizing and reconstructing the weight matrix.

[0039] like Figure 2The graph attention temporal fusion prediction model described herein has the following structure: the input layer receives compact feature vectors from each node; the graph attention layer calculates the attention weight matrix between nodes; the long short-term memory layer extracts temporal dependency features from the historical state sequences of each node; the graph pooling layer aggregates neighboring node information through summation pooling; the gated fusion layer uses a gating mechanism to fuse local and global features; the temporal convolutional layer uses dilated convolution to capture dependencies over long time spans; and the output layer generates fault risk scores for each node. The training dataset for this graph attention temporal fusion prediction model is established by collecting operational data from the power grid backup communication system over the past three years, marking the actual fault timestamps and node locations, using data from the 72 hours prior to the fault as input samples, and fault occurrence details as labels. The training and validation sets are divided at an 8:2 ratio, and the fault samples are augmented using synthetic minority oversampling technology to a ratio of 1:5 to the number of normal samples. The training steps for the graph attention temporal fusion prediction model are as follows: initialize the model parameters to follow a normal distribution with a mean of 0 and a standard deviation of 0.01, update the parameters using an adaptive moment estimation optimization algorithm, set the initial learning rate to 0.001 and adjust it according to an exponential decay strategy, set the batch size to 64, conduct 200 training epochs, use the loss function as a weighted sum of cross-entropy loss and focus loss with a weight ratio of 3:7, evaluate the model performance on the validation set every 10 epochs, and save the model parameters with the minimum validation loss.

[0040] The attention weights in the graph attention layer are calculated by assigning weights to nodes. and nodes The feature vectors are concatenated and mapped to scalars through a single-layer feedforward neural network. Attention coefficients are obtained by applying the LeakyReLU activation function and softmax normalization. These attention coefficients are modulated based on the product of signal transmission delay and link load rate in the noise reduction state parameter set. The modulation factor is the reciprocal of the product normalized to the 0-1 range. The Long Short-Term Memory (LSTM) layer contains three gating units: an input gate, a forget gate, and an output gate. The hidden state dimension is 128, and the time step is 72. The gating coefficients of the gated fusion layer are determined by the cosine similarity between the local and global feature vectors. The similarity is calculated by dividing the dot product of the two vectors by the product of their magnitudes. The temporal convolutional layer uses a causal convolution structure with dilation rates of 1, 2, 4, and 8, a kernel size of 3, and zero-padding.

[0041] The threshold of 0.72 is determined by plotting the subject operating characteristic curve on the validation set and selecting the failure risk score corresponding to the maximum Youden index, which is the difference between the true positive rate and the false positive rate.

[0042] The adaptive ensemble extreme learning machine consists of five base learners with 100, 150, 200, 250, and 300 hidden layer nodes respectively. The activation functions used are the Sigmoid, hyperbolic tangent, ReLU, Swish, and Gaussian radial basis functions, respectively. The root mean square error of each base learner is calculated on the validation set, and the inverse of the error is normalized and used as the ensemble weight coefficient. The input weights and hidden layer biases are optimized using a differential evolution algorithm with a population size of 50, a mutation factor of 0.8, a crossover probability of 0.9, and 100 iterations. Upon receiving a new sample, the incremental learning mechanism recursively updates the new sample features and the hidden layer output matrix using least squares, with a forgetting factor of 0.95 in the update formula. The Q-statistic is used to evaluate the differences among the base learners; a Q-statistic less than 0.3 indicates sufficient diversity. The final output is a refined fault probability distribution obtained by weighted averaging the fault probability predictions of each base learner.

[0043] The communication resource allocation game optimization model includes an upper-level model and a lower-level model. The upper-level model aims to minimize the overall network failure propagation rate, while the lower-level model aims to maximize the communication reliability of each node.

[0044] The objective function of the upper-layer model is used to calculate the minimization index of the overall network fault propagation rate. The inputs include refined fault probability distribution, link load rate, and signal transmission delay. The output is the overall network fault propagation rate. The objective function of the upper-layer model is expressed as follows: The overall network fault propagation rate divided by the baseline fault propagation rate equals the sum of the products of the refined fault probability distribution of each node and the number of edges connected to that node, divided by the total number of network nodes multiplied by the baseline fault probability, plus the product of the link load rate and signal transmission delay of each link, divided by the number of links multiplied by the baseline load delay product, and then multiplied by the sum of the squares of the fault risk scores of each node, divided by the total number of network nodes multiplied by the squares of the baseline fault risk scores. The constraints of the upper-layer model include that the sum of the communication bandwidth allocated to each node does not exceed the total system bandwidth, the transmit power of each node does not exceed the maximum allowable power, and the normalized value of the overall network fault propagation rate does not exceed 1.

[0045] The objective function of the lower-level model is used to calculate the maximization index of communication reliability for each node. The inputs include a refined fault probability distribution, link load rate, and allocated bandwidth for each node. The output is the communication reliability of each node. The objective function of the lower-level model is expressed as follows: The communication reliability of each node divided by the baseline communication reliability equals 1 minus the refined fault probability distribution of the node, multiplied by the allocated bandwidth of the node divided by the baseline bandwidth, then divided by 1, plus the sum of the link load rates of all adjacent edges of the node divided by the number of adjacent edges of the node multiplied by the baseline load rate, and the product is then multiplied by the fault risk score of each node divided by the baseline fault risk score. The constraints of the lower-level model include that the allocated bandwidth of each node is greater than the minimum service bandwidth, the normalized value of the communication reliability of each node is not less than 0.6, and the channel occupancy rate of each node does not exceed 0.9.

[0046] The coupling term between the objective functions of the upper-level and lower-level models is the fault risk score of each node. This node fault risk score simultaneously affects the calculation results of both the overall network fault propagation rate and the communication reliability of each node. The two-layer game equilibrium is solved using an iterative algorithm. The upper-level model optimizes the overall network fault propagation rate based on the current communication reliability of each node, while the lower-level model adjusts the communication reliability of each node based on the optimized overall network fault propagation rate. The iteration terminates when the change in the objective function value between two consecutive iterations is less than 0.001 or the number of iterations reaches 50.

[0047] The optimal communication resource allocation scheme includes the optimal transmission power, optimal allocated bandwidth, and optimal channel allocation strategy for each node. The optimal transmission power is the transmission power value that maximizes the communication reliability of each node in the equilibrium solution of the two-layer game. The optimal allocated bandwidth is the bandwidth allocation value that satisfies the constraints of the upper-layer model and the lower-layer model in the equilibrium solution of the two-layer game. The optimal channel allocation strategy is to select the communication channel with the least interference based on the link load rate and signal transmission delay of each node.

[0048] The baseline fault propagation rate is the historical average network fault propagation rate, expressed in times / hour. The baseline fault probability is the historical average network fault probability, dimensionless. The baseline load-delay product is the historical average link load rate multiplied by the average signal transmission delay, expressed in %·ms. The baseline fault risk score is the historical average network fault risk score, dimensionless. The baseline communication reliability is the historical average network communication reliability, dimensionless. The baseline bandwidth is the standard communication bandwidth designed for the system, expressed in Mbps. The baseline load rate is the standard link load rate designed for the system, expressed in %.

[0049] The virtual fault sample generation employs a synthetic minority class oversampling technique. For each fault sample, five nearest neighbor fault samples are found in the feature space. A point is randomly selected on the line connecting the fault sample and a randomly selected nearest neighbor fault sample as a synthetic sample. The number of synthetic samples is adaptively determined based on the class imbalance ratio. The cost-sensitive learning adds weight coefficients to the fault samples in the loss function, making the model focus more on the classification accuracy of fault samples during training.

[0050] The fault types include four categories: communication interruption, signal attenuation, protocol anomaly, and equipment overload. The dominant fault type is determined based on the maximum probability of each category in the refined fault probability distribution. The communication path reconstruction process calculates the shortest or highest-traffic path of the backup path based on the dynamic topology graph structure, and redistributes the communication traffic of high-risk nodes to the backup link through a dynamic routing protocol, with a reconstruction time of no more than 500ms.

[0051] Optionally, the present invention also provides a method for implementing a power grid backup communication fault pre-judgment system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions execute the above-described method when running in the computer.

[0052] The specific implementation methods of the above steps are described in detail below.

[0053] The specific implementation of step S1 involves first acquiring operational data in real time through data acquisition devices deployed at key nodes of the power grid backup communication system. These devices record raw data such as node voltage, current, signal quality, and network traffic with a sampling period of 100ms. The communication network is abstracted into a mathematical graph theory model, storing the network topology in the form of an adjacency matrix. The matrix element values ​​are calculated from the ratio of link bandwidth to real-time traffic; a larger ratio indicates more sufficient link margin. The raw voltage data is then used to calculate its value within a 10... The voltage fluctuation amplitude is obtained by dividing the difference between the maximum and minimum values ​​within the time window by the rated voltage value. The current signal is decomposed into fundamental and harmonic components using a Fast Fourier Transform (FFT). The effective values ​​of the 2nd to 50th harmonic currents are extracted, and the square root of the sum of the squares of all harmonic effective values ​​is divided by the effective value of the fundamental current to obtain the current harmonic distortion rate. The signal transmission delay is obtained by adding a timestamp to the data packet header and the difference between the timestamps at the receiving and transmitting ends. The bit error rate is obtained by calculating the ratio of the number of erroneous bits detected by the receiver to the total number of transmitted bits per unit time. The link load rate is obtained by monitoring the real-time traffic of each link and dividing it by the link's rated capacity. These five types of parameters together constitute the initial state parameter set, providing a data foundation for subsequent fault feature extraction.

[0054] The specific implementation of step S2 involves using a wavelet packet decomposition algorithm to perform multi-scale time-frequency analysis on the initial state parameter set. The Daubechies4 wavelet is chosen as the basis function because it has good compact support and regularity, and is highly sensitive to abrupt changes in the communication signal. A five-level decomposition is performed on the signal, uniformly dividing the spectrum into 32 sub-bands. Each sub-band corresponds to a signal component within a certain frequency range. This decomposition effectively separates fault characteristic signals from background noise in the frequency domain. Singular value decomposition is performed on the wavelet coefficient matrix of each band. The magnitude of the singular values ​​reflects the distribution of signal energy. Retaining the left singular vectors corresponding to the top 10 largest singular values ​​captures the main features of the signal while filtering out noise components. Spectral subtraction is applied to estimate the noise power spectrum in the frequency domain and subtract it from the power spectrum of the noisy signal. This method is based on the assumption that noise and signal are statistically independent, and noise reduction is achieved through power spectrum subtraction. The quotient of the ratio of signal power to noise power before and after noise reduction is the signal-to-noise ratio gain. When the gain is less than 12, it indicates that the noise suppression effect is not ideal. At this time, the number of empirical mode decomposition layers is increased to eight layers for adaptive supplementary noise reduction. Empirical mode decomposition can decompose non-stationary signals into several stationary intrinsic mode functions. Each function represents an oscillation mode at a different time scale. Eight-layer decomposition can more finely separate fault signals and noise, and finally obtain a noise reduction state parameter set with a significantly improved signal-to-noise ratio.

[0055] Step S3 is implemented by using principal component analysis (PCA) to linearly reduce the dimensionality of the denoised state parameter set. This algorithm projects the original high-dimensional features onto the principal component direction with the largest variance by calculating the eigenvalues ​​and eigenvectors of the covariance matrix. Principal components are selected by sorting the eigenvalues ​​from largest to smallest, and the cumulative variance contribution rate is calculated until it reaches 95%, at which point selection stops. The number of retained principal components is typically 10% to 30% of the original feature dimension, thus preserving most of the information while significantly reducing the feature space dimensionality. The mutual information criterion is used to measure the nonlinear correlation between features; a larger mutual information value indicates more redundant information carried by the two features. A threshold of 0.85 is set to delete one of the redundant feature pairs with mutual information exceeding this value. Further dimensionality reduction is achieved by applying a locally linear embedding manifold learning algorithm. This algorithm assumes that the data is distributed on a low-dimensional manifold in a high-dimensional space. By finding the nearest neighbors of each data point and calculating and reconstructing the weight matrix, the local neighborhood structure of the data remains unchanged in the low-dimensional space. Manifold learning can discover the inherent geometric structure of the data, ensuring that the compact feature vector set after dimensionality reduction maintains the topological relationship of the original data, providing high-quality input for subsequent models.

[0056] The specific implementation of step S4 involves organizing the compact feature vector set according to the time series and feeding it into the graph attention temporal fusion prediction model. The model first calculates the attention weights between node pairs in the graph attention layer; the weight magnitude reflects the correlation strength of fault propagation between nodes. The Long Short-Term Memory (LSTM) layer receives the feature sequence of each node's historical 72 time steps. An input gate controls the degree of new information acceptance, a forget gate determines the retention ratio of historical information, and an output gate adjusts the output of the current hidden state. The synergistic effect of these three gates enables the model to capture temporal dependencies spanning long periods. The graph pooling layer sums and aggregates the feature vectors within the node's neighborhood to obtain a global feature representation containing local topological information. The gated fusion layer calculates the cosine similarity between local and global features as the fusion weight. Higher similarity assigns greater weight to local features, while lower similarity relies more on global features. This adaptive fusion mechanism balances local details and global trends. The temporal convolutional layer employs a causal convolutional structure with increasing dilation rates. Four convolutional layers with dilation rates of 1, 2, 4, and 8 can expand the receptive field to 15 time steps without increasing the number of parameters, effectively capturing long-term temporal dependencies. The model output layer generates a fault risk score between 0 and 1 for each node. When the score is greater than 0.72, the node is considered to be in a high-risk state. At this time, an adaptive ensemble extreme learning machine is started for secondary verification. The extreme learning machine consists of 5 base learners with different numbers of hidden layer nodes and activation function types. The root mean square error of each base learner on the validation set is normalized by the inverse and used as the ensemble weight. The outputs of each base learner are fused by weighted average to obtain a more accurate and refined fault probability distribution.

[0057] The specific implementation of step S5 involves constructing a two-layer game optimization model to solve for the optimal allocation strategy of communication resources. The upper-layer model aims to minimize the overall network fault propagation rate. It normalizes the sum of the refined fault probability distribution of each node and the product of the number of connected edges, then adds the normalized value of the product of the link load rate and signal transmission delay of each link, and finally multiplies it by the normalization coefficient of the sum of squares of the fault risk scores to obtain a dimensionless evaluation index for the overall network fault propagation rate. The lower-layer model aims to maximize the communication reliability of each node. It subtracts the refined fault probability distribution of the node from 1, multiplies it by the normalized value of the node's allocated bandwidth, divides it by the normalization correction term of the average load rate of adjacent edge links, and finally multiplies it by the normalization coefficient of the fault risk score to obtain a dimensionless evaluation index for the communication reliability of each node. The upper-layer objective function and the lower-layer objective function are coupled through the fault risk scores of each node. The fault risk scores affect both the calculation of the overall network fault propagation rate and the evaluation of the communication reliability of each node, forming a strategic relationship between the two sides of the game. An iterative algorithm is used to solve the equilibrium solution of the two-layer game. The upper-layer model fixes the variables of the lower-layer model and optimizes its own objective function to obtain the optimal network resource allocation. The lower-layer model adjusts the resource allocation of each node based on the optimization results of the upper layer to maximize communication reliability. The two-layer model is optimized alternately until the change in the objective function is less than 0.001 or converges after 50 iterations. The solution at convergence is the game equilibrium point. The corresponding transmit power, allocated bandwidth and channel selection of each node constitute the optimal communication resource allocation scheme.

[0058] The specific implementation of step S6 involves selecting the category with the highest probability as the dominant fault type based on the probability values ​​of four types of faults—communication interruption, signal attenuation, protocol anomaly, and equipment overload—from the refined fault probability distribution. When the probability of a communication interruption fault exceeds 0.68, a synthetic minority class oversampling technique is used to generate virtual fault samples. In the feature space, five fault samples with the closest Euclidean distance are found as nearest neighbors for each fault sample. New samples are generated by interpolation along the line connecting the original sample and a randomly selected nearest neighbor sample at a random ratio. The number of synthetic samples is adaptively adjusted to 1:5 based on the ratio of fault samples to normal samples. In the loss function of the model training, fault samples are assigned a weight coefficient 50 times that of normal samples, which increases the penalty for classification errors of fault samples during gradient descent optimization, thereby improving the sensitivity to minority class fault identification. The graph attention temporal fusion prediction model is retrained using the expanded training dataset, and the model's weight and bias parameters are updated to enable the model to more accurately identify newly emerging fault modes.

[0059] The specific implementation of step S7 involves adjusting the operating parameters of each node based on the optimal communication resource allocation scheme obtained in step S5, setting the transmission power of each node to the optimal transmission power value, reallocating the communication bandwidth to the optimal allocated bandwidth value, and selecting the optimal channel allocation strategy based on the interference level of each channel. Warning levels are divided according to the numerical range of the refined fault probability distribution: nodes with a probability between 0.4 and 0.6 are marked as low-risk warnings and only logs are recorded; nodes with a probability between 0.6 and 0.8 are marked as medium-risk warnings and the monitoring frequency is increased to three times the normal value; nodes with a probability greater than 0.8 are marked as high-risk warnings and the communication path reconstruction process is immediately triggered. The path reconstruction process uses Dijkstra's shortest path algorithm or the maximum flow algorithm based on the dynamic topology graph structure to calculate the backup path from the faulty node to the target node. By modifying the dynamic routing table, communication traffic is transferred from the high-risk node to the backup link. The entire path reconstruction process is completed within 500ms to ensure communication continuity, thereby achieving early prediction and proactive protection against faults in the power grid backup communication system.

[0060] It should be noted that one of the key technical approaches of this invention is the use of a multi-scale denoising technique combining wavelet packet decomposition and singular value decomposition. Five-level wavelet packet decomposition separates the aliased fault signal from noise across 32 frequency bands. Singular value decomposition further extracts the principal components of the signal subspace in each frequency band. Spectral subtraction suppresses noise in the frequency domain. When the denoising effect is insufficient, the number of empirical mode decomposition layers is adaptively increased for supplementary processing. Compared to the limitations of traditional single-filtering methods that can only handle stationary noise, this multi-scale denoising technique can effectively address non-stationary strong noise such as electromagnetic interference and harmonic pollution in power grid communication systems. Even under harsh conditions with a signal-to-noise ratio as low as -10dB, it can accurately extract weak fault characteristic signals, significantly improving the sensitivity and reliability of early fault warning.

[0061] The second key technical approach involves constructing a graph attention-based temporal fusion prediction model to achieve network-level collaborative fault prediction. The model models the communication network as a dynamic graph structure. A graph attention mechanism learns the dynamic correlation weights of fault propagation between nodes. A long short-term memory network extracts temporal dependency features from the historical state sequences of each node over 72 time steps. A gated fusion layer adaptively balances local topology information with the global network state, and a temporal convolutional layer expands the temporal receptive field through dilated convolution to capture long-term evolutionary trends. Compared to traditional machine learning methods that model each node independently while ignoring inter-node correlations, this model fully utilizes the network topology and temporal evolution patterns, enabling it to predict the fault risk of each node 72 hours before a fault occurs, providing maintenance personnel with ample time for preventative maintenance.

[0062] The third key technical approach involves establishing a two-layer game-theoretic optimization model for communication resource allocation to achieve dynamic optimization of network resources. The upper-layer model optimizes the global resource allocation strategy with the goal of minimizing the overall network fault propagation rate, while the lower-layer model adjusts local resource allocation parameters with the goal of maximizing the communication reliability of each node. The two models are coupled through fault risk scoring of each node, and an iterative algorithm is used to solve for the game equilibrium point to obtain the optimal communication resource allocation scheme. Compared with the shortcomings of traditional static resource allocation methods that cannot dynamically adjust according to network conditions, the game-theoretic optimization model can optimize the transmit power, communication bandwidth, and channel allocation of each node in real time during network topology changes and fault risk evolution, minimizing fault propagation risk while ensuring communication service quality.

[0063] The synergistic effect of the three key technological approaches described above forms a complete closed-loop system for predicting and anticipating power grid backup communication faults. Multi-scale noise reduction technology provides high-quality input features for the graph attention time-series fusion prediction model, while the refined fault probability distribution output by the model provides a decision-making basis for the game-theoretic optimization model. The resource allocation scheme optimized by the game-theoretic model further improves the model's prediction environment by reducing the network fault propagation rate. These three aspects mutually reinforce each other, forming a positive feedback loop. Compared to the fragmented architecture of signal preprocessing, fault prediction, and resource scheduling in existing technologies, this collaborative system achieves integrated intelligent decision-making across the entire process from data acquisition, feature extraction, fault prediction to resource optimization and allocation. It reduces the fault warning time of the power grid backup communication system from several hours in traditional methods to 72 hours, and lowers the false alarm rate to less than one-fifth of that of traditional methods, significantly improving the security and reliability of the power grid backup communication system.

[0064] It should be noted that this invention also solves the following technical problem: existing methods for predicting power grid communication faults suffer from a weak ability to identify fault samples due to imbalanced training sample classes. This invention expands the training database by generating virtual fault samples using synthetic minority class oversampling technology. Cost-sensitive learning is used to set the weight coefficients of fault samples to 50 times that of normal samples, making the model more focused on the classification accuracy of fault samples during training. Simultaneously, an adaptive ensemble extreme learning machine is used to construct a heterogeneous base learner with diverse activation functions and hidden layer nodes of different sizes. Differential evolutionary algorithms are used to optimize input weights and hidden layer biases, and an incremental learning mechanism enables online updates of model parameters, effectively improving the model's accuracy in identifying minority class fault samples and resolving the negative impact of sample imbalance on predictive performance.

[0065] Specifically, the principle of this invention is as follows: This invention can solve the technical problem of low accuracy in predicting communication faults in power grids by employing a prediction framework combining graph neural networks and time series analysis, as well as a resource allocation mechanism based on two-layer game optimization. Traditional methods treat each node in isolation, ignoring the topological characteristics of the communication network. The graph attention mechanism, by calculating the attention weight matrix between nodes, can adaptively learn the influence relationships between nodes and capture the propagation patterns of faults in the network. Long Short-Term Memory (LSTM) networks selectively retain key historical information through gating mechanisms, solving the problem of insufficient modeling ability for long-term dependencies in traditional time series models. The adaptive ensemble extreme learning machine uses multiple heterogeneous base learners and optimizes parameters through differential evolution algorithms, enhancing the model's generalization ability and robustness. The two-layer game optimization model couples fault prediction results with resource allocation decisions. The upper-layer model optimizes global fault propagation control, while the lower-layer model optimizes local node reliability. Iteratively solving for Nash equilibrium achieves closed-loop feedback between prediction and control, ensuring the accuracy of fault prediction in principle.

[0066] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0067] The specific implementation methods of steps S4, S6 and S7 are the same as those described above, and will not be repeated in detail here.

[0068] The specific implementation of step S1 involves collecting operational data from each node of the power grid backup communication system, modeling the communication network as a dynamic topology graph structure, where nodes represent communication devices and edges represent communication links, and extracting the voltage fluctuation amplitude, current harmonic distortion rate, signal transmission delay, bit error rate, and link load rate of each node as the initial state parameter set. The formula for calculating the voltage fluctuation amplitude is as follows:

[0069] ;

[0070] In the formula, Voltage fluctuation amplitude, in percentage (%) To monitor the maximum node voltage during the specified period, the unit is... ; The minimum node voltage during the monitoring period is expressed in units of 1. ; Rated voltage, unit: The formula for calculating the harmonic distortion rate of current is expressed as follows:

[0071] ;

[0072] In the formula, The current harmonic distortion rate is expressed in % (%). For the first The effective value of the second harmonic current, in units of , The value range is from the 2nd to the 50th harmonic; This is the effective value of the fundamental current, in units of . The formula for calculating link load rate is as follows:

[0073] ;

[0074] In the formula, Link load rate, in percentage (%) Current link traffic, in Mbps; This refers to the rated capacity of the link, measured in Mbps. The formula for calculating the bit error rate is as follows:

[0075] ;

[0076] In the formula, Bit error rate, dimensionless; The number of transmitted error bits, dimensionless; The total number of transmitted bits is dimensionless. The formula for calculating the edge weights in a dynamic topology graph structure is as follows:

[0077] ;

[0078] In the formula, Let be the weight of the edge, which is dimensionless; Link bandwidth, in Mbps; This represents the current link traffic, measured in Mbps.

[0079] The specific implementation of step S2 involves performing a five-level wavelet packet decomposition on the initial state parameter set, extracting signal subspace features from the singular value decomposition of the coefficients in each frequency band after decomposition, reconstructing the signal after suppressing background noise using spectral subtraction, and calculating the signal-to-noise ratio (SNR) gain between the reconstructed signal and the original signal. When the SNR gain is less than 12, the number of empirical mode decomposition layers is increased to eight, resulting in the denoised state parameter set. The formula for calculating the denoised signal using spectral subtraction is as follows:

[0080] ;

[0081] In the formula, The frequency domain representation of the noise-reduced signal, in units of ; The frequency domain representation of the noisy signal, in units of ; This is an estimate of the noise power spectrum, in units of... ; Frequency, unit: The formula for calculating the signal-to-noise ratio gain is as follows:

[0082] ;

[0083] In the formula, Signal-to-noise ratio gain, in dB; The signal power after noise reduction is expressed in watts (W). The noise power after noise reduction is expressed in watts (W). The signal power before noise reduction is expressed in watts (W). This represents the noise power before noise reduction, expressed in W.

[0084] The specific implementation of step S3 involves performing principal component analysis (PCA) on the denoised state parameter set for linear dimensionality reduction, retaining principal components with a cumulative contribution rate of 95%, calculating the correlation between features using the mutual information criterion, deleting redundant features with mutual information values ​​greater than 0.85, and preserving the inherent structure of the data through local linear embedding manifold learning to obtain a compact feature vector set. The formula for calculating the cumulative contribution rate is expressed as follows:

[0085] ;

[0086] In the formula, The cumulative contribution rate is dimensionless. For the first The variance of each principal component is dimensionless. The number of principal components to be retained is dimensionless; The total number of principal components is dimensionless. The formula for calculating mutual information is as follows:

[0087] ;

[0088] In the formula, Features With features Mutual information between them, in units of ; Features The space of values; Features The space of values; Features The possible values ​​of ; Features The possible values ​​of ; It is a joint probability distribution, dimensionless; Features The marginal probability distribution of is dimensionless; Features The marginal probability distribution is dimensionless.

[0089] The specific implementation of step S5 involves establishing a communication resource allocation game optimization model. The upper-layer model aims to minimize the overall network failure propagation rate, while the lower-layer model aims to maximize the communication reliability of each node. Refined failure probability distribution, link load rate, and signal transmission delay are used as input parameters. The optimal communication resource allocation scheme is obtained by solving the two-layer game equilibrium. The objective function of the upper-layer model is expressed as follows:

[0090] ;

[0091] In the formula, This represents the overall network fault propagation rate, expressed in times per hour. The baseline fault propagation rate is expressed in times per hour. For nodes A refined fault probability distribution, dimensionless; For nodes The number of connected edges, dimensionless; The total number of network nodes, dimensionless; The baseline failure probability is dimensionless. For the first in the network Link load rate of each link, in percentage (%) For the first in the network The signal transmission delay of a link, measured in milliseconds; The total number of links in the network, dimensionless; The baseline load factor is expressed as % The reference signal transmission delay is expressed in milliseconds (ms). For nodes Fault risk score, dimensionless; The baseline fault risk score is dimensionless. The objective function of the lower-level model is expressed as follows:

[0092] ;

[0093] In the formula, For nodes The communication reliability is dimensionless. The baseline communication reliability is dimensionless. For nodes The allocated bandwidth is measured in Mbps. This is the baseline bandwidth, measured in Mbps. For nodes The number of adjacent edges, dimensionless; For nodes The The link load rate of each adjacent edge, in units of %.

[0094] The formula for calculating the attention weights in the graph attention layer is as follows:

[0095] ;

[0096] In the formula, For nodes and nodes The attention coefficient between them is dimensionless; This represents the weight vector of a single-layer feedforward neural network. For nodes eigenvectors; For nodes Feature vectors of the links between them; This represents a vector concatenation operation; For nodes The set of neighboring nodes; For nodes With nodes The signal transmission delay of the link between them, in milliseconds; The reference signal transmission delay is expressed in milliseconds (ms). For nodes With nodes Link load rate of the links between them, in percentages (%) The baseline load factor is expressed as % It is the minimum value of the normalized delay-load product of all links in the entire network, and is dimensionless. It is the maximum value of the normalized delay-load product of all links in the entire network, and is dimensionless. To prevent small constants from being divided by zero, a value of is usually taken. , dimensionless. Among them, and The calculation formulas are respectively and , Let be the set of all edges in the network.

[0097] The formula for calculating the gating coefficient of the gating fusion layer is as follows:

[0098] ;

[0099] In the formula, For nodes The gating coefficient is dimensionless; For nodes Local eigenvectors; For nodes The global feature vector; This represents the vector dot product operation; Represents the magnitude of a vector. Vector magnitude The calculation formula is: ,in The first local feature vector One portion, The dimension of the local feature vector is dimensionless; The calculation formula is: ,in The first global feature vector One portion, The dimension of the global feature vector is dimensionless.

[0100] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0101] To verify the effectiveness of this invention, technicians built a numerical simulation analysis environment to simulate the fault prediction process of a regional power grid backup communication system under high load operation. The communication system comprises 120 communication nodes connected via fiber optic and wireless links, forming a redundant network topology. The simulation system continuously collected 72 hours of operating data at a sampling frequency of 10Hz. Each node collected five types of state parameters: voltage fluctuation amplitude, current harmonic distortion rate, signal transmission delay, bit error rate, and link load rate. The initial state parameters collected ranged from 1.2% to 8.7%, current harmonic distortion rate from 3.5% to 15.8%, signal transmission delay from 12ms to 89ms, and bit error rate from [missing data]. to The link load rate ranges from 35% to 92%. Technicians model the communication network as a dynamic topology graph structure, where the weight of the edges is determined by the ratio of link bandwidth to real-time traffic, with bandwidth ranging from 100Mbps to 1000Mbps, and real-time traffic calculated based on the link load rate.

[0102] Technicians performed a five-level wavelet packet decomposition on the initial state parameter set, using the Daubechies4 wavelet basis function to decompose the signal into 32 frequency bands. Singular value decomposition was then performed on the coefficient matrix of each frequency band, retaining the left singular vectors corresponding to the top 10 largest singular values ​​to form the signal subspace. After suppressing background noise using spectral subtraction, the signal was reconstructed. The calculated signal-to-noise ratio gain between the reconstructed and original signals was 9.6 dB. Since this value was less than 12, the technicians increased the number of empirical mode decomposition levels to eight, ultimately obtaining the denoised state parameter set. After denoising, the standard deviation of voltage fluctuation amplitude decreased from 1.85% to 0.62%, the standard deviation of current harmonic distortion rate decreased from 2.43% to 0.91%, and the standard deviation of signal transmission delay decreased from 18.7 ms to 7.3 ms.

[0103] Technicians performed principal component analysis (PCA) on the denoised state parameter set for linear dimensionality reduction, obtaining 15 principal components through orthogonal transformation. The cumulative contribution rate of the first 8 principal components reached 95.3%. Subsequently, the mutual information criterion was used to calculate the correlation between features. It was found that the mutual information value between the 3rd and 7th principal components was 0.88, and the mutual information value between the 5th and 11th principal components was 0.91. The technicians removed these redundant features with mutual information values ​​greater than 0.85. Local linear embedding manifold learning was used to preserve the intrinsic structure of the data, mapping the 8-dimensional principal component data to a 5-dimensional compact feature vector set. During the mapping process, the local neighborhood structure of the original data was preserved, and the number of neighborhood points was set to 12.

[0104] Technicians input a compact feature vector set into a graph attention temporal fusion prediction model, which was trained on historical data from the past three years. The training set contained 8640 normal samples and 1728 faulty samples. Figure 3 As shown, the model outputs fault risk score vectors for each node. The fault risk scores for nodes numbered 37, 54, 68, 89, and 103 are 0.78, 0.81, 0.75, 0.86, and 0.79, respectively, all exceeding the threshold of 0.72. Technicians then used an adaptive ensemble extreme learning machine (ESM) for secondary validation on these five high-risk nodes. This ESM consists of five base learners with hidden layer nodes of 100, 150, 200, 250, and 300. The root mean square errors of each base learner on the validation set are 0.042, 0.038, 0.035, 0.041, and 0.039, respectively. After normalizing the reciprocals of these errors, the ensemble weight coefficients are 0.213, 0.235, 0.256, 0.218, and 0.229, respectively. By fusing the predictions from each base learner using a weighted average, the refined fault probability distributions for nodes 37, 54, 68, 89, and 103 are obtained as 0.71, 0.76, 0.69, 0.82, and 0.73, respectively.

[0105] Technicians established a game-theoretic optimization model for communication resource allocation. The upper-level model aims to minimize the overall network failure propagation rate, while the lower-level model aims to maximize the communication reliability of each node. Using refined failure probability distribution, link load rate, and signal transmission delay as input parameters, an iterative algorithm was used to solve for the two-level game equilibrium. Initially, the normalized value of the overall network failure propagation rate was 0.68, and the normalized value of the average communication reliability of each node was 0.71. After 32 iterations, the change in the objective function value between two consecutive iterations decreased to 0.0008, satisfying the termination condition. At this point, the normalized value of the overall network failure propagation rate decreased to 0.42, and the normalized value of the average communication reliability of each node increased to 0.83. In the optimal communication resource allocation scheme, the optimal transmit powers of nodes 37, 54, 68, 89 and 103 are 18.3W, 21.5W, 16.7W, 23.8W and 19.2W respectively, and the optimal allocated bandwidths are 256Mbps, 312Mbps, 228Mbps, 348Mbps and 271Mbps respectively.

[0106] Technicians determined the fault type based on the refined fault probability distribution. Node 89 had a fault probability of 0.82 and its dominant fault type was communication interruption, meeting the conditions for generating virtual fault samples. Using synthetic minority class oversampling, the technicians found the five nearest neighbor fault samples in the feature space for this node's fault sample, synthesizing 128 virtual fault samples to expand the training database. Cost-sensitive learning was then used to set the weight coefficients of the fault samples to 50 times that of normal samples, updating the parameters of the graph attention temporal fusion prediction model. The updated model improved the fault detection accuracy on the validation set from 87.3% to 93.6%, and reduced the false negative rate from 12.7% to 6.4%.

[0107] As shown in Table 1, technicians adjust the transmit power and channel allocation strategy of each node according to the optimal communication resource allocation scheme.

[0108] Table 1. Communication Resource Adjustment Plan for High-Risk Nodes

[0109]

[0110] Technicians established a fault warning level classification rule, marking faults with a probability distribution between 0.4 and 0.6 as low-risk, between 0.6 and 0.8 as medium-risk, and greater than 0.8 as high-risk. Nodes 37, 54, 68, and 103 were marked as medium-risk, and node 89 was marked as high-risk. A communication path reconstruction process was triggered for node 89, calculating an alternative path based on the dynamic topology graph structure, such as... Figure 4As shown, the backup path passing through nodes 12, 45, and 67 is selected as the shortest path, with a path length of 4 hops and an estimated transmission latency of 35ms. The communication traffic of node 89 is redistributed to the backup link using a dynamic routing protocol. The reconfiguration process takes 472ms, which does not exceed the 500ms limit. After reconfiguration, the link load rate of node 89 decreases from 88% to 62%, the signal transmission latency decreases from 78ms to 41ms, and the fault risk score decreases from 0.86 to 0.58.

[0111] like Figure 5 As shown, technicians compared the changes in communication reliability of each node before and after communication path reconstruction. Before reconstruction, the high-risk area involved 17 nodes, with an average normalized communication reliability value of 0.64. After reconstruction, the high-risk area was reduced to 3 nodes, and the average normalized communication reliability value increased to 0.81. Throughout the simulation, the system successfully predicted 8 potential communication failures, issuing warnings for 6 within 24 hours before the actual failures and 2 within 48 hours before the actual failures. By adjusting the communication resource allocation scheme and reconstructing the communication path in advance, the system avoided 5 communication interruption incidents and shortened the recovery time for 3 failures.

[0112] The advancements of this invention over traditional methods lie in its multi-level feature extraction and dimensionality reduction strategy. By combining wavelet packet decomposition with singular value decomposition to extract signal subspace features, it effectively suppresses background noise interference with state parameters, resulting in more accurate feature extraction. The combination of principal component analysis and locally linear embedding manifold learning preserves the inherent structure of the data, avoiding the loss of nonlinear relationships inherent in traditional linear dimensionality reduction methods. The graph attention temporal fusion prediction model can simultaneously capture spatial dependencies between nodes and long-term time series dependencies, exhibiting stronger fault prediction capabilities compared to traditional single-time series prediction models. The two-layer game optimization model achieves a synergistic balance between overall network optimization and individual node optimization, avoiding the unreasonable allocation of local resources that may result from traditional single-objective optimization. Cost-sensitive learning and virtual sample generation techniques address the problem of insufficient model training due to scarce fault samples, improving the model's ability to identify minority class faults. The communication path reconstruction mechanism proactively adjusts the network topology before a fault occurs, significantly shortening system downtime compared to traditional passive fault recovery methods.

[0113] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0114] Table 2. Variable Explanation Table (Part 1)

[0115]

[0116] Table 3. Variable Explanation Table (Part Two)

[0117]

[0118] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting communication faults in power grids based on intelligent algorithms, characterized in that, This process includes collecting operational data from each node of the power grid backup communication system and modeling the communication network as a dynamic topology graph. Voltage fluctuation amplitude, current harmonic distortion rate, signal transmission delay, bit error rate, and link load rate of each node are extracted as initial state parameter sets. Wavelet packet decomposition and singular value decomposition are used to denoise the initial state parameter sets, resulting in a denoised state parameter set. Principal component analysis and local linear embedding manifold learning are then applied to the denoised state parameter set to obtain a compact feature vector set. This compact feature vector set is input into a graph attention temporal fusion prediction model, which outputs a fault risk score vector. When the fault risk score exceeds a threshold, an adaptive ensemble extreme learning machine is activated for secondary verification to obtain a refined fault probability distribution. A communication resource allocation game optimization model is established to solve for the optimal communication resource allocation scheme. Fault types are determined based on the refined fault probability distribution, and virtual fault samples are generated to update model parameters. The transmit power and channel allocation strategies of each node are adjusted according to the optimal communication resource allocation scheme, and a fault warning level classification rule is established to trigger the communication path reconstruction process.

2. The method according to claim 1, characterized in that, The dynamic topology graph structure refers to abstracting each communication device in the power grid backup communication network as a node of the graph, and the physical or logical communication connections between devices as edges of the graph. The weight of the edge is determined by the ratio of link bandwidth and real-time traffic.

3. The method according to claim 2, characterized in that, The voltage fluctuation amplitude is the ratio of the difference between the maximum and minimum node voltages during the monitoring period to the rated voltage, and the current harmonic distortion rate is the square root of the sum of the squares of the effective values ​​of each harmonic current divided by the effective value of the fundamental current.

4. The method according to claim 3, characterized in that, The wavelet packet decomposition uses the Daubechies4 wavelet basis function to perform a five-level wavelet packet transform on the signal, decomposing the signal into 32 frequency bands. The singular value decomposition performs singular value decomposition on the coefficient matrix of each frequency band, retaining the left singular vectors corresponding to the top 10 largest singular values ​​to form the signal subspace.

5. The method according to claim 4, characterized in that, After extracting the signal subspace features by singular value decomposition, the method also includes using spectral subtraction to subtract the noise power spectrum estimate from the noise power spectrum estimate in the frequency domain and then performing an inverse transformation to obtain the denoised signal. The signal-to-noise ratio gain between the reconstructed signal and the original signal is calculated. When the signal-to-noise ratio gain is less than 12, the number of empirical mode decomposition layers is increased to eight.

6. The method according to claim 5, characterized in that, The principal component analysis dimensionality reduction specifically involves converting related variables into linearly independent principal components through orthogonal transformation and retaining principal components with a cumulative contribution rate of 95%. The mutual information criterion is used to calculate the correlation between features and delete redundant features with mutual information values ​​greater than 0.

85.

7. The method according to claim 6, characterized in that, The graph attention temporal fusion prediction model includes a graph attention layer that calculates the attention weight matrix between nodes, a long short-term memory layer that extracts temporal dependency features from the historical state sequence of each node, a graph pooling layer that aggregates neighboring node information through summation pooling, a gated fusion layer that uses a gating mechanism to fuse local and global features, and a temporal convolution layer that uses dilated convolution to capture dependencies over long time spans.

8. The method according to claim 7, characterized in that, The attention weights in the graph attention layer are calculated by concatenating the node feature vectors and mapping them to scalars through a single-layer feedforward neural network. The scalars are then activated by the LeakyReLU activation function and normalized by softmax to obtain the attention coefficients. The attention coefficients are modulated based on the product of the signal transmission delay and the link load rate in the noise reduction state parameter set.

9. The method according to claim 8, characterized in that, The threshold of 0.72 is determined by selecting the failure risk score corresponding to the maximum Youden index by plotting the receiver operating characteristic curve on the validation set. The Youden index is the difference between the true positive rate and the false positive rate.

10. The method according to claim 9, characterized in that, The adaptive ensemble extreme learning machine consists of 5 base learners with 100, 150, 200, 250, and 300 hidden layer nodes, respectively. The activation functions are Sigmoid, hyperbolic tangent, ReLU, Swish, and Gaussian radial basis functions, respectively. The inverse of the root mean square error of each base learner is normalized on the validation set and used as the ensemble weight coefficient.

Images