Power system fault intelligent diagnosis method, device, equipment and medium

Abstract

The application relates to an intelligent fault diagnosis method, device, equipment and medium for a power system. The method comprises the following steps: receiving multi-source data in the operation of a power system; synchronizing the multi-source data to obtain time-space synchronized power parameters; based on the power parameters, obtaining real-time fault estimation results through a fault migration model; based on the real-time fault estimation results, obtaining a cascading fault risk corresponding to the power system through a multi-layer LSTM model; and based on the cascading fault risk, obtaining a fault diagnosis report corresponding to the power system. The method can focus on the cascading fault propagation path, predict the fault diffusion trend, and improve the safety of the power system.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent monitoring, and particularly relates to a power system fault intelligent diagnosis method, device, equipment and medium. BACKGROUND

[0002] With the development of power system technology, power system fault diagnosis technology appears. Traditional power system fault diagnosis methods mainly rely on offline constructed expert knowledge base and threshold criterion, and realize fault positioning through multi-source data matching. Specifically, first, based on the voltage, current, power and other steady-state measurement data collected by the SCADA (resource collection and monitoring) system, combined with the protection device action signal, the rule reasoning or fuzzy logic is used to judge the fault area. For example, when it is detected that the current of a line exceeds the preset threshold and the corresponding circuit breaker acts, the line is determined as the fault point. The traditional power system fault diagnosis method has the problems of relying on artificial experience summary, only focusing on a single fault point, lacking attention to the cascading fault propagation path, and being difficult to predict the fault diffusion trend. SUMMARY

[0003] Therefore, it is necessary to provide a power system fault intelligent diagnosis method, device, equipment and medium capable of saving human resources, paying attention to the cascading fault propagation path, and predicting the fault diffusion trend in view of the above technical problems.

[0004] In a first aspect, the application provides a power system fault intelligent diagnosis method, comprising:

[0005] receiving multi-source data in power system operation;

[0006] synchronizing the multi-source data to obtain time and space synchronized power parameters;

[0007] based on the power parameters, obtaining real-time fault estimation results through a fault migration model;

[0008] based on the real-time fault estimation results, obtaining the cascading fault risk of the power system through a multi-layer LSTM model;

[0009] based on the cascading fault risk, obtaining a fault diagnosis report corresponding to the power system.

[0010] Further, based on the power parameters, the real-time fault estimation results are obtained through the fault migration model, comprising:

[0011] extracting power parameter features to obtain a fault time sequence segment;

[0012] extracting the voltage value of each node in the fault time sequence segment to obtain a voltage time sequence;

[0013] based on the voltage time sequence, the correlation coefficient matrix is obtained through the following formula:

[0014]

[0015] wherein, ρ ij is the correlation coefficient of nodes i and j, representing the linear correlation of the two residual errors, the correlation coefficient matrix is the set of ρ ij , V i (t) is the voltage value of node i at time t, φ k is the k-th order coefficient of the autoregressive model, p is the order of the autoregressive model, ∈ t is the residual error of the model at time t, ∈ i , ∈ j is the residual error sequence of node i and node j, Cov(∈ i ,∈ j ) is the covariance of the residual error sequence of node i and node j, is the standard deviation of the residual error sequence of node i and node j;

[0016] Discard the node pairs with correlation coefficient matrix less than the threshold value, and determine the remaining correlation coefficient matrix as the fault migration result;

[0017] Based on the fault migration result, the state transition matrix is obtained in combination with the historical fault statistical library;

[0018] Based on the state transition matrix, the real-time fault estimation result is obtained.

[0019] Further, based on the fault migration result, the state transition matrix is obtained in combination with the historical fault statistical library, including:

[0020] Based on the historical fault statistical library, the historical fault times of each path in the fault migration result are counted to obtain a historical fault statistical table;

[0021] Based on the historical fault statistical table, the initial transition probability matrix is calculated by the following formula:

[0022]

[0023] wherein, N ij is the historical number of times of fault migration from node i to node j, N i is the total number of faults of node i, is the initial transition probability matrix, representing the historical statistical probability of fault migration to node j when node i has a fault;

[0024] Based on the topology database and the initial transition probability matrix, the weighted transition probability matrix is calculated by the following formula:

[0025]

[0026] wherein, is the weighted transition probability matrix, Z ij is the line impedance from node i to node j, extracted from the topology database, reflecting the physical connection characteristics of the power grid, Z base is the impedance weight factor;

[0027] The weighted transition probability matrix is normalized to obtain a state transition matrix.

[0028] Further, based on the fault migration result, in combination with the historical fault statistical library, after obtaining the state transition matrix, it further includes:

[0029] Detecting a real-time switch state change signal and correcting the state transition matrix based on the real-time switch state change signal to obtain a real-time state matrix;

[0030] Based on the real-time state matrix, a new transition probability is obtained through a Q-learning algorithm;

[0031] Based on the new transition probability, the state transition matrix is updated to obtain an updated state transition matrix.

[0032] Further, based on the state transition matrix, a real-time fault estimation result is obtained, including:

[0033] Based on the state transition matrix, the predicted state is calculated through the following formula:

[0034]

[0035] wherein, is the prior state estimation, indicating the predicted state at time t without considering the current observation, and M is the state transition matrix, is the prior state estimation at the previous time, is the prior state covariance matrix, and α t-1 is the posterior covariance matrix at the previous time, and Q is the engineering noise covariance matrix;

[0036] Based on the predicted state, based on the Kalman gain, the corrected predicted state is calculated through the following formula:

[0037]

[0038] wherein, K t is the Kalman gain, H is the observation matrix, R is the observation noise covariance matrix, and z t is the actual observation value, indicating the measurement data at time t, is the posterior state estimation, indicating the corrected predicted state, and α t is the posterior covariance matrix, and I represents the identity matrix;

[0039] Based on the corrected prediction state and the state transition matrix, a real-time fault estimation result is obtained.

[0040] Further, based on the real-time fault estimation result, a cascading fault risk corresponding to the power system is obtained through a multi-layer LSTM model, including:

[0041] Features are extracted from the real-time fault estimation result to obtain a fault diffusion feature vector;

[0042] The fault diffusion feature vector is sorted by time to obtain an input sequence matrix;

[0043] Based on a three-layer LSTM structure, the input sequence matrix is obtained through the following formula:

[0044]

[0045] wherein, is the predicted value, X in is the input sequence matrix, is the hidden state of the lth layer LSTM at time step t, LSTM l is the calculation function of the lth layer LSTM, W is the weight matrix of the output layer, and b is the bias vector of the output layer;

[0046] Based on the predicted value, a Seq2Seq structure is used to recursively generate subsequent predictions to obtain the cascading fault risk corresponding to the power system.

[0047] Further, based on the predicted value, a Seq2Seq structure is used to recursively generate subsequent predictions to obtain the cascading fault risk, and further includes:

[0048] Align the predicted value and the true value to obtain a distribution sample pair;

[0049] Based on the distribution sample pair, the Wasserstein distance is calculated through the following formula:

[0050]

[0051] γ∈Π(μ,v)

[0052] wherein, is the Wasserstein distance, μ and v are two probability distributions, Π(μ,v) is a set of joint probability distributions satisfying the edge distribution μ and v, γ represents the transportation process from μ to v, ||x-y|| is the distance between two points in space, and inf is the lower limit;

[0053] The Wasserstein distance is used as a loss function, and the updated LSTM structure is obtained by backpropagation.

[0054] In a second aspect, the application further provides an intelligent power system fault diagnosis device, comprising:

[0055] a receiving module configured to receive multi-source data in power system operation;

[0056] a synchronizing module configured to synchronize the multi-source data to obtain spatiotemporally synchronized power parameters;

[0057] a migration module configured to obtain real-time fault estimation results based on the power parameters through a fault migration model;

[0058] a risk module configured to obtain a cascading fault risk of the power system based on the real-time fault estimation results through a multi-layer LSTM model;

[0059] a diagnosis module configured to obtain a fault diagnosis report of the power system based on the cascading fault risk.

[0060] In a third aspect, the application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method provided in the first aspect of the application when executing the computer program.

[0061] In a fourth aspect, the application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the method provided in the first aspect of the application.

[0062] The above intelligent power system fault diagnosis method, device, equipment and medium receive multi-source data in power system operation, synchronize the multi-source data to obtain spatiotemporally synchronized power parameters, obtain real-time fault estimation results based on the power parameters through a fault migration model, obtain a cascading fault risk of the power system based on the real-time fault estimation results through a multi-layer LSTM model, and obtain a fault diagnosis report of the power system based on the cascading fault risk. The technical means realize closed-loop management from data collection to risk warning, significantly improve the reliability and self-healing ability of the power system, save human resources, focus on the cascading fault propagation path, and predict the fault diffusion trend. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the application or the related art, the drawings needed to be used in the embodiment or related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0064] Figure 1 The flowchart of the intelligent power system fault diagnosis method of the application;

[0065] Figure 2 Figure 1 is a diagram of an intelligent fault diagnosis device for a power system of the present application. DETAILED DESCRIPTION

[0066] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0067] In one embodiment, as shown in Figure 1 Figure 1, an intelligent fault diagnosis method for a power system is provided, and the present embodiment is exemplified by the method applied to a terminal. It should be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and can be realized through the interaction of the terminal and the server. In the present embodiment, the method includes the following steps:

[0068] Step 101, receiving multi-source data in the operation of the power system.

[0069] Specifically, the power system completes the dynamic process of power generation, power transmission, power transformation, power distribution and power consumption in normal or abnormal state, including device state, load change, power grid topology and other information. Multi-source data is heterogeneous data from different collection devices or systems, which can include data acquisition and monitoring systems, synchronous phasor measurement units, weather sensors, device logs, user power consumption records, etc.

[0070] Step 102, synchronizing multi-source data to obtain time-space synchronized power parameters.

[0071] Specifically, time-space synchronization is the consistency of data in time and space dimensions. Power parameters are physical quantities used to quantify the state of the power system, which can include voltage, current, frequency, power factor, harmonic distortion rate, etc. Through time-space alignment, the differences in sampling frequency of different devices are solved, and redundant or conflicting data in power parameters are eliminated through Kalman filtering or weighted average.

[0072] Step 103, based on the power parameters, obtaining real-time fault estimation results through a fault migration model.

[0073] Specifically, the fault migration model is a mathematical model describing the propagation of faults from the occurrence point to other devices, which can be a combination of autoregressive model and residual calculation. Real-time fault estimation results are online diagnosis of fault type, location and severity, which have real-time performance. The synchronized power parameters are input, and the fault features are identified through the fusion algorithm of autoregressive model and residual.

[0074] At step 104, based on the real-time fault estimation result, a cascading fault risk corresponding to the power system is obtained through a multi-layer LSTM model.

[0075] Specifically, the multi-layer LSTM model is a long short-term memory network containing multiple hidden layers, which is suitable for capturing long-term dependencies in time series data. The cascading fault risk is the potential probability and impact range of the chain reaction caused by the initial fault, such as line overload, protection misoperation, voltage collapse, etc. Early warning of possible expanding faults. Combine the fault estimation result with the historical data to construct the time sequence feature, input the real-time fault estimation result into the LSTM network, and output the fault propagation path and risk level of the next N time steps.

[0076] At step 105, based on the cascading fault risk, a fault diagnosis report corresponding to the power system is obtained.

[0077] Specifically, the fault diagnosis report is a structured document or visual dashboard containing fault root cause, impact range, risk level, and treatment suggestion. Automatically extract key indicators (such as risk probability, affected load), combine natural language generation technology to output text, and intuitively display high-risk areas through topological coloring graph and heat map.

[0078] The power system fault intelligent diagnosis method provided by the embodiment of the application receives multi-source data in the operation of the power system; synchronizes the multi-source data to obtain spatiotemporally synchronized power parameters; based on the power parameters, obtains real-time fault estimation results through a fault migration model; based on the real-time fault estimation results, obtains a cascading fault risk corresponding to the power system through a multi-layer LSTM model; and obtains a fault diagnosis report corresponding to the power system based on the cascading fault risk. The steps realize closed-loop management from data acquisition to risk warning, significantly improve the reliability and self-healing ability of the power system, save human resources, pay attention to the cascading fault propagation path, and predict the fault diffusion trend.

[0079] In one of the embodiments, based on the power parameters, the real-time fault estimation result is obtained through the fault migration model, including:

[0080] At step 201, the power parameter features are extracted to obtain a fault time sequence segment.

[0081] Specifically, feature extraction is to select key information related to faults from original power parameters. The fault time sequence segment is a continuous time sequence data segment extracted within a specific time window, such as 10 seconds before the fault to 30 seconds after the fault. The data is extracted by sliding the fixed time window, and the abnormal segment is marked by threshold method or clustering algorithm.

[0082] At step 202, the voltage value of each node in the fault time sequence segment is extracted to obtain a voltage time sequence.

[0083] Specifically, the node is a device connection point in the power grid topology, and the voltage time series is a continuous record of the change of the node voltage over time. The voltage time series of each node is extracted from the fault time series segment. The voltage values are normalized by a reference value to eliminate the dimension effect.

[0084] Step 203, based on the voltage time series, the correlation coefficient matrix is obtained by the following formula:

[0085]

[0086] Where, ρ ij is the correlation coefficient of nodes i and j, indicating the linear correlation of the residuals of the two, and the correlation coefficient matrix is the set of ρ ij , V i (t) is the voltage value of node i at time t, φ k is the k-th order coefficient of the autoregressive model, p is the order of the autoregressive model, ∈ t is the residual of the model at time t, ∈ i , ∈ j is the residual sequence of node i and node j, Cov(∈ i ,∈ j ) is the covariance of the residual sequence of node i and node j, is the standard deviation of the residual sequence of node i and node j.

[0087] Specifically, the autoregressive model is a linear model for predicting current values with historical data, the correlation coefficient matrix is a linear correlation matrix of the residual sequences between nodes, the AR model coefficients are solved by least squares or Yule-Walker equation, the residuals are generated by the difference between the actual voltage value and the predicted value of the AR model, and the covariance and standard deviation of the node residual sequence are calculated.

[0088] Step 204, eliminate the node pairs with correlation coefficient matrix less than the threshold value, and determine the remaining correlation coefficient matrix as the fault migration result.

[0089] Specifically, the node pair refers to the combination of two different nodes, specifically the two nodes corresponding to the row index i and the column index j in the correlation coefficient matrix, i.e. two devices or location points in the power grid, used to quantify the correlation between the two in fault propagation. The elements below the threshold value are removed. The remaining node pairs constitute the potential paths of fault propagation.

[0090] Step 205, based on the fault migration result, combine the historical fault statistical library to obtain the state transition matrix.

[0091] Specifically, the historical fault statistics library is a database recording historical fault types, propagation paths, and occurrence frequencies. The state transition matrix represents the conditional probability of node j failure triggered by node i failure. The current fault migration result is weighted and fused with the historical statistical probability to improve the prediction credibility.

[0092] In step 206, based on the state transition matrix, real-time fault estimation results are obtained.

[0093] Specifically, the real-time fault estimation result can include fault source positioning, propagation path prediction, and impact range assessment. The state transition probability of the future k steps is calculated through the Markov chain model. The high-risk node list and the propagation path are output. The fault evolution trend is dynamically predicted.

[0094] The embodiment quantifies the fault propagation probability through AR model residual correlation analysis, combines historical statistics and real-time data, and takes into account real-time and accuracy, which is suitable for online diagnosis of complex power grids. The threshold screening reduces the false positive rate and improves the practicability.

[0095] In one embodiment, based on the fault migration result, the state transition matrix is obtained by combining the historical fault statistics library, including:

[0096] In step 301, based on the historical fault statistics library, the historical fault times of each path in the fault migration result are counted to obtain a historical fault statistics table.

[0097] Specifically, the historical fault statistics library is a structured database storing historical fault events, which can include fields such as fault nodes, migration paths, occurrence times, and durations. Each path in the fault migration result is a path from one node to another. The current fault migration path is matched with the same node pair record in the historical library. The historical number of times of fault migration from node i to node j and the total number of faults of node i are counted.

[0098] In step 302, based on the historical fault statistics table, the initial transition probability matrix is calculated by the following formula:

[0099]

[0100] where N ij is the historical number of times of fault migration from node i to node j, N i is the total number of faults of node i, is the initial transition probability matrix, representing the historical statistical probability of fault migration to node j when node i fails.

[0101] Specifically, the initial transition probability matrix represents the failure propagation probability based on historical data only. The historical statistics are converted into probability form, which intuitively reflects the failure propagation tendency between nodes. If the total failure number of node i is zero, the value of the initial transition probability matrix is set to zero. If the numerator is zero, Laplace smoothing is introduced to avoid zero probability.

[0102] In step 303, based on the topology database and the initial transition probability matrix, the weighted transition probability matrix is calculated by the following formula:

[0103]

[0104] wherein, is the weighted transition probability matrix, Z ij is the line impedance from node i to node j, which is extracted from the topology database and reflects the physical connection characteristics of the power grid, Z base is the impedance weight factor.

[0105] Specifically, the topology database is a database that stores the physical connection relationship of the power grid, including line impedance, node coordinates, etc. The impedance weight factor is a reference value for normalizing impedance. The line impedance of the node pair is read from the topology database. The smaller the impedance, the larger the weighting factor, and the higher the propagation probability. The physical characteristics of the power grid are integrated into the probability model, making the propagation probability more consistent with the actual electrical rules.

[0106] In step 304, the weighted transition probability matrix is normalized to obtain the state transition matrix.

[0107] Specifically, the row normalization formula is used for calculation to ensure that the sum of each row of the matrix is 1, which meets the definition of the state transition probability of Markov chain.

[0108] This embodiment improves the prediction credibility by considering both historical rules (data-driven) and power grid physical characteristics (model-driven). Normalization ensures the legality of the probability and supports long-term evolution analysis of Markov chain.

[0109] In one of the embodiments, based on the fault migration results, after obtaining the state transition matrix in combination with the historical fault statistics library, it further includes:

[0110] In step 401, a real-time switch state change signal is detected, and the state transition matrix is corrected based on the real-time switch state change signal to obtain a real-time state matrix.

[0111] Specifically, the real-time switch state change signal is the real-time opening / closing state change of devices such as circuit breakers and disconnectors (such as line opening / closing), which is usually from the SCADA system or protection device action signal. The real-time state matrix is a state transition matrix dynamically adjusted based on the current power grid topology.

[0112] For example, if the line is disconnected, the impedance of the relevant node pair is set to infinity (unreachable), and the weighted transition probability is recalculated.

[0113] At step 402, based on the real-time state matrix, a new transition probability is obtained through the Q-learning algorithm.

[0114] Specifically, Q-learning is a model-free reinforcement learning algorithm that optimizes the decision-making strategy by iteratively updating the Q value. The new transition probability is a dynamically adjusted transition probability that combines real-time operating data and Q-learning exploration results. The grid state (node voltage, load rate, switch state, etc.) is encoded as a discrete or continuous state space.

[0115] At step 403, based on the new transition probability, the state transition matrix is updated to obtain an updated state transition matrix.

[0116] Specifically, the updated state transition matrix is a probability matrix that combines real-time topology correction and Q-learning exploration results. The historical state transition matrix and the Q-learning new transition probability are combined by weighted averaging.

[0117] This embodiment uses reinforcement learning strategy to update the state transition matrix in real time, which improves the dynamic adaptability of fault prediction. It supports complex fault scenarios of new power grids.

[0118] In one embodiment, based on the state transition matrix, real-time fault estimation results are obtained, including:

[0119] At step 501, based on the state transition matrix, the predicted state is calculated by the following formula:

[0120]

[0121] where, is the prior state estimation, represents the predicted state at time t without considering the current observation, and M is the state transition matrix, is the prior state estimation at the previous time, is the prior state covariance matrix, and α t-1 is the posterior covariance matrix at the previous time, and Q is the engineering noise covariance matrix.

[0122] Specifically, the state transition matrix is a matrix describing the probability of fault propagation between nodes, the prior state estimation is a predicted state based only on historical state transition rules (without considering current observation data), representing the fault probability distribution of each node at time t. The prior covariance matrix is a measure of the uncertainty of the predicted state, containing the covariance information of the prediction error. The engineering noise covariance matrix represents the noise covariance of uncontrollable factors in modeling fault propagation. Predict the future trend of fault propagation. The initial time can be set as a uniform distribution or based on the real-time fault estimation result, and the state at the last time is distributed to the next time according to the transition probability through matrix operation.

[0123] Step 502, based on the predicted state, based on the Kalman gain, the modified predicted state is calculated by the following formula:

[0124]

[0125] Where K is the Kalman gain, H is the observation matrix, R is the observation noise covariance matrix, z t is the actual observation value, z t is the measurement data at time t, is the posterior state estimation, which represents the modified predicted state, and a t is the posterior covariance matrix, and I represents the identity matrix.

[0126] Specifically, the observation matrix is a matrix that maps the state space to the observation space, and the observation noise covariance matrix is the covariance of the measurement device error. Dynamically fuse real-time observation data to correct the prediction bias. The residual represents the difference between the observed value and the predicted value, and if the observation noise is small, the Kalman gain will increase, and the observation data will be more trusted.

[0127] For example, the voltage of node B is observed to be abnormal, but the model predicts that its fault probability is low, and the Kalman gain will increase the probability weight of node B.

[0128] Step 503, based on the modified predicted state and the state transition matrix, the real-time fault estimation result is obtained.

[0129] Specifically, the fault probability of each node is extracted from the posterior state estimation, a probability threshold is set, high-risk nodes are marked and an alarm signal is generated, the source of the fault is traced back based on the state transition matrix, and the real-time fault location, probability and propagation path are output.

[0130] This embodiment improves the prediction noise resistance by modifying the model, real-time fuses SCADA / PMU observation data, improves the estimation accuracy, and gradually approaches the real fault state through iterative correction of Kalman filtering, providing a reliable basis for operation and maintenance decision-making.

[0131] In one of the embodiments, based on the real-time fault estimation result, the corresponding cascading failure risk of the power system is obtained through a multi-layer LSTM model, including:

[0132] Step 601, features are extracted from the real-time fault estimation result to obtain a fault propagation feature vector.

[0133] Specifically, the fault propagation feature vector is a multi-dimensional vector representing the dynamic of fault propagation, which can include the following dimensions: node fault probability from the real-time fault estimation result; topological connection strength representing the electrical distance or transition probability between nodes; time window statistics, which are the mean, variance, and trend in the sliding window. The multi-dimensional heterogeneous data is encoded into a unified dimensional feature vector to adapt to the input requirements of LSTM.

[0134] Step 602, the fault propagation feature vector is sorted by time to obtain an input sequence matrix.

[0135] Specifically, the input sequence matrix is a set of feature vectors arranged in chronological order. Continuous sequences are intercepted at a fixed time step, and linear interpolation or forward padding is used for outage data. The input structure with time dependence is formed.

[0136] Step 603, based on a three-layer LSTM structure, the input sequence matrix is obtained through the following formula:

[0137]

[0138] wherein, is the predicted value, X in is the input sequence matrix, is the hidden state of the l-th layer LSTM at time step t, LSTM l (·) is the calculation function of the l-th layer LSTM, W is the weight matrix of the output layer, and b is the bias vector of the output layer.

[0139] Specifically, the multi-layer LSTM is a stacked LSTM structure, each layer gradually abstracts the time sequence features (the bottom layer captures short-term dependencies, and the high layer captures long-term patterns). The hidden state transmits time sequence information. The first layer inputs the original feature sequence and outputs the primary time sequence feature. The second layer further integrates the context information across time steps. The third layer extracts high-order time sequence patterns.

[0140] Step 604, based on the predicted value, a Seq2Seq structure is used to recursively generate subsequent predictions to obtain the corresponding cascading failure risk of the power system.

[0141] Specifically, the Seq2Seq structure is an encoder-decoder architecture, the encoder compresses the input sequence into a context vector, and the decoder recursively generates multi-step predictions to generate a cascading failure risk sequence of multiple time steps.

[0142] This embodiment captures minute-level fault propagation through long-term dependency modeling, realizes a closed loop from real-time data to multi-step risk prediction, and significantly improves the resilience of the power grid.

[0143] In one embodiment, based on the predicted value, the Seq2Seq structure is used to recursively generate subsequent predictions, and the cascading fault risk is obtained, which further includes:

[0144] Step 701, align the predicted value and the true value to obtain a distribution sample pair.

[0145] Specifically, the predicted value is a future multi-step cascading fault risk probability sequence output by the Seq2Seq model, and the true value is the actual observed fault propagation result (such as node overload events recorded by SCADA). The distribution sample pair is a matching data pair obtained by aligning the predicted sequence and the true sequence by time, which is used to measure the distribution difference between the two. Through sliding window interception and dynamic time warping, time alignment is ensured to compare the predicted and true data under the same time reference, avoiding misalignment evaluation.

[0146] Step 702, based on the distribution sample pair, the Wasserstein distance is calculated by the following formula:

[0147]

[0148] γ∈Π(μ,v)

[0149] wherein, is the Wasserstein distance, μ and v are two probability distributions, Π(μ,v) is a set of joint probability distributions satisfying the edge distribution μ and v, γ represents the transportation process from μ to v, ||x-y|| is the distance between two points in space, and inf is the lower limit.

[0150] Specifically, the Wasserstein distance is used to measure the difference between two probability distributions μ (predicted distribution) and v (true distribution), defined as the minimum transportation cost, which quantifies the overall difference between the predicted distribution and the true distribution.

[0151] Step 703, take the Wasserstein distance as the loss function, and back-propagate to obtain the updated LSTM structure.

[0152] Specifically, the Wasserstein distance is used as the training target to drive the optimization of model parameters. Through the chain rule, the gradient is transmitted to the LSTM parameters, forcing the LSTM to generate more accurate predictions closer to the true distribution, and improving the accuracy of multi-step predictions.

[0153] By introducing the Wasserstein distance as the loss function, the model can better capture the statistical characteristics of fault propagation in a complex power system, providing a more reliable decision basis for cascade failure prevention.

[0154] It should be understood that, although each step in the flowchart involved in each embodiment as described above is shown in sequence according to the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately executed with at least part of other steps or steps or stages in other steps.

[0155] Based on the same inventive concept, the embodiments of the present application also provide an intelligent power system fault diagnosis device for implementing the above-mentioned intelligent power system fault diagnosis method. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more intelligent power system fault diagnosis device embodiments provided below can refer to the limitations of the intelligent power system fault diagnosis method described above, which will not be repeated here.

[0156] In one exemplary embodiment, as shown in Figure 2 An intelligent power system fault diagnosis device 800 is provided, comprising:

[0157] The receiving module 801 is configured to receive multi-source data in the operation of the power system;

[0158] The synchronization module 802 is configured to synchronize the multi-source data to obtain time-space synchronized power parameters;

[0159] The migration module 803 is configured to obtain real-time fault estimation results based on the power parameters through a fault migration model;

[0160] The risk module 804 is configured to obtain the corresponding cascade failure risk of the power system based on the real-time fault estimation results through a multi-layer LSTM model;

[0161] The diagnosis module 805 is configured to obtain a fault diagnosis report corresponding to the power system based on the cascade failure risk.

[0162] Further, the migration module 803 is further configured to:

[0163] Extracting power parameter features to obtain a fault time sequence segment;

[0164] Extracting the voltage value of each node in the fault time sequence segment to obtain a voltage time sequence;

[0165] Based on the voltage time sequence, the correlation coefficient matrix is obtained by the following formula:

[0166]

[0167] Wherein, ρ ij is the correlation coefficient of nodes i and j, indicating the linear correlation of the residuals of the two, and the correlation coefficient matrix is a set of ρ ij , V i (t) is the voltage value of node i at time t, φ k is the k-th order coefficient of the autoregressive model, p is the order of the autoregressive model, ∈ t is the residual of the model at time t, ∈ i , ∈ j is the residual sequence of node i and node j, Cov(∈ i ,∈ j ) is the covariance of the residual sequence of node i and node j, is the standard deviation of the residual sequence of node i and node j;

[0168] Eliminating the node pairs with correlation coefficient matrix less than the threshold value, and determining the remaining correlation coefficient matrix as the fault migration result;

[0169] Based on the fault migration result, the state transition matrix is obtained by combining the historical fault statistical library;

[0170] Based on the state transition matrix, the real-time fault estimation result is obtained.

[0171] Further, the migration module 803 is also used for:

[0172] Based on the historical fault statistical library, the historical fault times of each path in the fault migration result are counted to obtain a historical fault statistical table;

[0173] Based on the historical fault statistical table, the initial transition probability matrix is calculated by the following formula:

[0174]

[0175] Wherein, N ij is the historical number of times of fault migration from node i to node j, N i is the total number of faults of node i, is the initial transition probability matrix, indicating the historical statistical probability of fault migration to node j when node i has a fault;

[0176] Based on the topology database and the initial transition probability matrix, a weighted transition probability matrix is calculated by the following formula:

[0177]

[0178] wherein, is the weighted transition probability matrix, Z ij is the line impedance from node i to node j, extracted from the topology database, reflecting the physical connection characteristics of the power grid, Z base is the impedance weight factor;

[0179] The weighted transition probability matrix is normalized to obtain a state transition matrix.

[0180] Further, the migration module 803 is further configured to:

[0181] detect a real-time switch state change signal, and correct the state transition matrix based on the real-time switch state change signal to obtain a real-time state matrix;

[0182] based on the real-time state matrix, a new transition probability is obtained by a Q-learning algorithm;

[0183] based on the new transition probability, the state transition matrix is updated to obtain an updated state transition matrix.

[0184] Further, the migration module 803 is further configured to:

[0185] based on the state transition matrix, a predicted state is calculated by the following formula:

[0186]

[0187] wherein, is the prior state estimation, indicating the predicted state at time t without considering the current observation, M is the state transition matrix, is the prior state estimation at the previous time, is the prior state covariance matrix, α t-1 is the posterior covariance matrix at the previous time, Q is the engineering noise covariance matrix;

[0188] based on the predicted state, a corrected predicted state is calculated based on the Kalman gain by the following formula:

[0189]

[0190] wherein, K t is the Kalman gain, H is the observation matrix, R is the observation noise covariance matrix, z t is the actual observation value, indicating the measurement data at time t, is the corrected predicted state, and α t is the posterior covariance matrix, and I is the identity matrix.

[0191] Based on the corrected predicted state and the state transition matrix, a real-time fault estimation result is obtained.

[0192] Further, the risk module 804 is further configured to:

[0193] Extracting features from the real-time fault estimation result, a fault propagation feature vector is obtained.

[0194] The fault propagation feature vector is sorted by time to obtain an input sequence matrix.

[0195] Based on the three-layer LSTM structure, the input sequence matrix is obtained by the following formula:

[0196]

[0197] wherein, is the predicted value, X in is the input sequence matrix, is the hidden state of the l-th layer LSTM at time step t, LSTM l is the calculation function of the l-th layer LSTM, W is the weight matrix of the output layer, and b is the bias vector of the output layer.

[0198] Based on the predicted value, a Seq2Seq structure is used to recursively generate subsequent predictions to obtain the cascading fault risk corresponding to the power system.

[0199] Further, the system further comprises a feedback module configured to:

[0200] Aligning the predicted value and the true value to obtain a distribution sample pair.

[0201] Based on the distribution sample pair, the Wasserstein distance is calculated by the following formula:

[0202]

[0203] γ∈Π(μ,v)

[0204] wherein, is the Wasserstein distance, μ and v are two probability distributions, Π(μ, v) is a set of joint probability distributions satisfying the edge distribution μ and v, γ represents the transport process from μ to v, ||x-y|| is the distance between two points in space, and inf is the lower limit.

[0205] The Wasserstein distance is used as a loss function, and the updated LSTM structure is obtained by backpropagation.

[0206] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the power system fault intelligent diagnosis method as described above when executing the computer program.

[0207] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method embodiments described above.

[0208] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts are described in the part of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present disclosure according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0209] The above-described embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the application. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application.

Claims

1. A power system fault intelligent diagnosis method, characterized in that, The method comprises: receiving multi-source data in power system operation; synchronizing the multi-source data to obtain spatio-temporally synchronized power parameters; based on the power parameters, obtaining real-time fault estimation results through a fault transfer model; based on the real-time fault estimation results, obtaining the cascading fault risk of the power system through a multi-layer LSTM model; based on the cascading fault risk, obtaining a fault diagnosis report corresponding to the power system; wherein, based on the real-time fault estimation results, the cascading fault risk of the power system is obtained through a multi-layer LSTM model, comprising: extracting features from the real-time fault estimation results to obtain a fault diffusion feature vector; sorting the fault diffusion feature vector by time to obtain an input sequence matrix; based on a three-layer LSTM structure, the input sequence matrix is obtained through the following formula: ； ； ； ； wherein, is a predicted value, is an input sequence matrix, is a hidden state of the l-th layer LSTM at time step t, is a computation function of the l-th layer LSTM, W is a weight matrix of the output layer, b is a bias vector of the output layer, and l is a numerical index representing 1, 2, 3. based on the predicted value, using the Seq2Seq structure to recursively generate subsequent predictions to obtain the cascading fault risk of the power system.

2. The intelligent diagnosis method of power system faults according to claim 1, characterized in that, based on the power parameters, the real-time fault estimation results are obtained through a fault transfer model, comprising: extracting features from the power parameters to obtain a fault time sequence segment; extracting the voltage value of each node in the fault time sequence segment to obtain a voltage time sequence; based on the voltage time sequence, the correlation coefficient matrix is obtained through the following formula: ； ； wherein, is the correlation coefficient of nodes i and j, indicating the linear correlation of the two residuals, and the correlation coefficient matrix is is the set of is the voltage value of node i at time t, is the kth order coefficient of the autoregressive model, and p is the order of the autoregressive model, is the residual of the model at time t, is the residual sequence of node i and node j, is the covariance of the residual sequence of node i and node j, is the standard deviation of the residual sequence of node i and node j; eliminate the node pairs with correlation coefficient matrix less than the threshold value, and determine the remaining correlation coefficient matrix as the fault transfer result; based on the fault transfer result, a state transition matrix is obtained by combining a historical fault statistical library; based on the state transition matrix, the real-time fault estimation results are obtained.

3. The method of claim 2, wherein, based on the fault transfer result, a state transition matrix is obtained by combining a historical fault statistical library, comprising: based on the historical fault statistical library, the historical fault frequency of each path in the fault transfer result is counted to obtain a historical fault statistical table; based on the historical fault statistical table, the initial transition probability matrix is calculated through the following formula: ； wherein, is the number of historical failures from node i to node j, is the total number of failures of node i, is the initial transition probability matrix, representing the historical statistical probability of migrating to node j when node i fails. based on the topology database and the initial transition probability matrix, the weighted transition probability matrix is calculated through the following formula: ； wherein, is the weighted transition probability matrix, is the line impedance from node i to node j, extracted from the topology database, reflecting the physical connection characteristics of the power grid, is the impedance weight factor, is the initial transition probability matrix; the weighted transition probability matrix is normalized to obtain the state transition matrix.

4. The method of claim 2, wherein, after obtaining the state transition matrix based on the fault transfer result and combining the historical fault statistical library, it further comprises: detecting a real-time switch state change signal and correcting the state transition matrix based on the real-time switch state change signal to obtain a real-time state matrix; based on the real-time state matrix, a new transition probability is obtained through a Q-learning algorithm; based on the new transition probability, the state transition matrix is updated to obtain an updated state transition matrix.

5. The method of claim 2, wherein, based on the state transition matrix, the real-time fault estimation results are obtained, comprising: based on the state transition matrix, the predicted state is calculated through the following formula: ； ； wherein, is a prior state estimate, represents a predicted state at time t without considering the current observation, and M is a state transition matrix, is a prior state estimate at the previous time, is a prior state covariance matrix, is a posterior covariance matrix at the previous time, and Q is an engineering noise covariance matrix; based on the predicted state, based on the Kalman gain, the corrected predicted state is calculated through the following formula: ； ； ； wherein, is the Kalman gain, H is the observation matrix, R is the observation noise covariance matrix, is the actual observation value, represents the measurement data at time t, is the posterior state estimate, represents the modified predicted state, is the posterior covariance matrix, I represents the identity matrix; based on the corrected predicted state and the state transition matrix, the real-time fault estimation results are obtained.

6. The method of claim 1, wherein, The method further comprises: aligning the predicted value and the true value to obtain a distribution sample pair; calculating the Wasserstein distance based on the distribution sample pair by using the following formula: ； where, is the Wasserstein distance, is the two probability distributions, is the joint probability distribution set satisfying the marginal distribution is the joint probability distribution set satisfying the marginal distribution denotes the transport process from to is the distance between two points in space, and inf is the lower bound.​ using the Wasserstein distance as a loss function to obtain an updated LSTM structure through back propagation.

7. An intelligent diagnosis apparatus for power system faults, characterized by, The device comprises: a receiving module configured to receive multi-source data in power system operation; a synchronization module configured to synchronize the multi-source data to obtain time-space synchronized power parameters; a migration module configured to obtain real-time fault estimation results based on the power parameters through a fault migration model; a risk module configured to obtain a cascading failure risk of the power system based on the real-time fault estimation results through a multi-layer LSTM model; a diagnosis module configured to obtain a fault diagnosis report of the power system based on the cascading failure risk. The risk module is further configured to: extract features from the real-time fault estimation results to obtain a fault diffusion feature vector; sort the fault diffusion feature vector by time to obtain an input sequence matrix; obtain a predicted value based on the input sequence matrix through a three-layer LSTM structure by using the following formula: ； ； ； ； wherein, is a predicted value, is an input sequence matrix, is a hidden state of the l-th layer LSTM at time step t, is a computation function of the l-th layer LSTM, W is a weight matrix of the output layer, b is a bias vector of the output layer, and l is a numerical index representing 1, 2, 3. obtain the cascading failure risk of the power system by recursively generating subsequent predictions based on the predicted value using a Seq2Seq structure.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

Images