A Power System Fault Prediction Method Based on Digital Twins

Abstract

This invention relates to the field of power operation and maintenance monitoring technology, specifically a power system fault prediction method based on digital twins. The method includes: collecting synchronous electrical measurement data from each monitoring node of the target power system within a continuous time window to form an original measurement sequence; and building a digital twin model that operates synchronously with the physical power grid based on the original measurement sequence. Within the model, multi-timescale dynamic state estimation is performed on the measurement sequence to generate a feature set of the current operating state of the power system. The feature set is then imported into a pre-trained temporal convolutional network, which outputs a fault probability distribution vector. Components exceeding a threshold are selected to locate fault prediction nodes and generate fault warning indicators. This approach can deeply mine the correlation features of power time-series data, accurately depict the system's operating status, and achieve accurate prediction and node location of potential power system faults.

Description

Technical Field

[0001] This invention relates to the field of power operation and maintenance monitoring technology, and in particular to a power system fault prediction method based on digital twins. Background Technology

[0002] Power system fault prediction is a crucial aspect of power grid safety management. Currently, most power system fault prediction operations directly collect electrical measurement data from field monitoring nodes and rely on fixed analytical logic for superficial data processing. This lack of a simulation mapping platform that matches the actual power grid in real time results in data analysis detached from the system's actual operating conditions. Conventional analysis models only use a single time scale to extrapolate the state of measurement data, leading to a fixed and rigid analytical dimension. This makes it impossible to adapt the analytical scale to dynamic changes such as power system load fluctuations and operating condition switching. The resulting system operating status information is fragmented and incomplete, failing to fully reconstruct the true operating state of the power grid.

[0003] Existing fault prediction methods mostly rely on basic data analysis models to process electrical measurement data. However, the model structure is difficult to adapt to the correlation characteristics of continuous time-series measurement data. They can only roughly determine whether a fault has occurred based on fixed thresholds and cannot output refined fault probability information. Traditional methods cannot accurately trace the source of potential faults, but can only delineate abnormal areas over a wide area. The generation of early warning signals lacks quantitative judgment criteria, which easily leads to missed or false alarms. This makes it difficult to meet the actual needs of the power grid for refined fault prediction and early warning. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a power system fault prediction method based on digital twins.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a power system fault prediction method based on digital twins, comprising: Synchronous electrical measurement data of each monitoring node in the target power system within a continuous time window are collected as the raw measurement sequence. A digital twin model that operates synchronously with the target power system is constructed based on the original measurement sequence. In the digital twin model, multi-timescale dynamic state estimation is performed on the original measurement sequence to generate a current operating state feature set; The current operating state feature set is input into a pre-trained temporal convolutional network to output a fault probability distribution vector; Based on the components in the fault probability distribution vector that exceed a set threshold, fault prediction nodes are located and fault warning identifiers are generated.

[0006] As a further aspect of the present invention, the step of collecting synchronous electrical measurement data of each monitoring node in the target power system within a continuous time window as the original measurement sequence specifically includes: Synchronous phasor measurement devices are deployed at each bus node and both ends of each transmission line in the target power system. Each synchronous phasor measurement device is controlled to acquire four electrical parameters—voltage amplitude, voltage phase angle, current amplitude, and current phase angle—at the same sampling frequency and on the same time reference. All electrical parameters collected from all synchronous phasor measurement devices will be aligned according to timestamps and organized into a multi-channel timing matrix; The row index of the multi-channel time series matrix is ​​the time sampling point, and the column index is a combination of the monitoring node identifier and the electrical parameter type; Missing value detection is performed on the multi-channel time series matrix, and the entire row of data corresponding to the time sampling point with missing value is removed; A sliding window segmentation is performed on the multi-channel time series matrix after removing missing values ​​to obtain sub-matrices under multiple consecutive time windows; Each submatrix is ​​used as a sequence of original measurements.

[0007] As a further aspect of the present invention, the step of constructing a digital twin model that operates synchronously with the target power system based on the original measurement sequence specifically includes: Obtain the offline topology parameters and offline component electrical parameters corresponding to the target power system; An initial framework for establishing the node admittance matrix is ​​established based on the offline topology parameters; Assign an initial admittance value to each non-zero element in the node admittance matrix based on the electrical parameters of the offline components; Each of the original measurement sequences is input into the calculation engine corresponding to the node admittance matrix in chronological order according to the time window. Within each time window, the real-time admittance correction value for each line is inversely calculated using the voltage phasor and current phasor in the original measurement sequence; The real-time admittance correction value is superimposed on the initial admittance value of the node admittance matrix to form a time-varying node admittance matrix; The time-varying node admittance matrix and the topological connection relationship of the target power system are encapsulated together into an executable simulation model as the digital twin model.

[0008] As a further aspect of the present invention, the step of sequentially inputting each of the original measurement sequences into the calculation engine corresponding to the node admittance matrix according to the time window sequence specifically includes: buffering and sorting the continuously arriving original measurement sequences through a message queue, and providing input to the calculation engine in a first-in-first-out manner.

[0009] As a further aspect of the present invention, the step of performing multi-timescale dynamic state estimation on the original measurement sequence in the digital twin model to generate a current operating state feature set specifically includes: In the digital twin model, an extended Kalman filter is set for each monitoring node; The measurement value of the current time window in the original measurement sequence is used as the observation input and fed into the corresponding extended Kalman filter; A prediction step is performed within each extended Kalman filter to obtain the state prior estimate at the current time. An update step is performed within each extended Kalman filter to compute the residual between the state prior estimate and the observed input and to correct the state posterior estimate. Collect the posterior state estimates of all extended Kalman filter outputs into a set of state vectors within a time window; Perform cubic spline interpolation along the time dimension on the set of state vectors under multiple consecutive time windows to obtain continuous state trajectories with millisecond-level resolution; The slope of the state change of each monitoring node at the start and end points of the time window is extracted from the continuous state trajectory as a fast dynamic feature component. Extract the mean and variance of the state of each monitoring node within the entire time window from the continuous state trajectory as steady-state feature components; The fast dynamic feature components and the steady-state feature components are concatenated into a one-dimensional feature vector, which is used as the current running state feature set.

[0010] As a further aspect of the present invention, in the step of extracting the state change slope of each monitoring node at the start and end points of the time window as a fast dynamic feature component from the continuous state trajectory, the state change slope is calculated by linearly fitting a preset number of sampling points near the start and end points within the time window using the least squares method.

[0011] As a further aspect of the present invention, the step of inputting the current operating state feature set into a pre-trained temporal convolutional network to output a fault probability distribution vector specifically includes: Obtain the structural parameters of the pre-trained temporal convolutional network model; The temporal convolutional network model includes multiple dilated convolutional layers, each followed by a batch normalization layer and an activation function layer. Arrange the current running state feature set into a two-dimensional input tensor according to the time window order; The first dimension of the two-dimensional input tensor is the time window index, and the second dimension is the one-dimensional feature vector corresponding to each time window. The two-dimensional input tensor is sequentially passed through multiple dilated convolutional layers to extract fault association feature maps under different receptive fields at different times. In each batch normalization layer, the output fault association feature map is subjected to distribution normalization. In each activation function layer, a nonlinear mapping is performed on the standardized fault association feature map; The fault association feature map output from the last activation function layer is input into the fully connected layer to reduce the dimensionality to the dimension of the number of fault categories; The dimensionality reduction result of the fully connected layer is subjected to a normalized exponential function mapping to use the probability value of each fault category as the fault probability distribution vector.

[0012] As a further aspect of the present invention, the training steps of the temporal convolutional network model include: constructing samples using historical fault events and corresponding running state feature sets, and optimizing the model parameters using stochastic gradient descent combined with a focus loss function.

[0013] As a further aspect of the present invention, the step of locating fault prediction nodes and generating fault warning identifiers based on components exceeding a set threshold in the fault probability distribution vector specifically includes: A probability warning threshold is pre-set for each type of fault; Compare the magnitudes of each component in the fault probability distribution vector with the probability warning threshold of the fault type corresponding to that component; Collect all fault types whose component values ​​are greater than the corresponding probability warning threshold as a candidate warning fault set; The fault type with the largest component value is selected from the candidate early warning fault set as the main early warning fault type; Query the list of associated monitoring nodes corresponding to the main early warning fault type in the target power system; Mark the three monitoring nodes in the associated monitoring node list that are closest to the typical location of the fault type as fault prediction nodes; Generate an early warning identifier that includes the name of the main early warning fault type, the identifier of the fault prediction node, and the current time window index; The warning sign is output as the fault warning sign.

[0014] As a further aspect of the present invention, the step of pre-setting a probability warning threshold for each fault type specifically includes: calculating and dynamically adjusting the corresponding probability warning threshold based on the frequency of each fault type in historical samples and the cost loss function of missed and false alarms.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: A digital twin model operating synchronously with the target power system is constructed. Based on the collected original measurement sequences, a virtual replica of the physical power grid is achieved. Dynamic state estimation at multiple time scales is then performed within the virtual model environment. The virtual model can synchronously follow the operating conditions of the physical power grid and iterate its operational status. The multi-time-scale analysis mode can analyze the changing patterns of electrical measurement data across different time spans, progressively mining the implicit operational status information within the data, broadening the extraction range of state features, and ensuring that the generated operational state feature set closely matches the complex and ever-changing dynamic operation of the power grid, fully covering the system's operational details under different operating conditions.

[0016] Temporal convolutional networks adapt to the arrangement characteristics of continuous time-series electrical measurement data, capturing the correlation patterns between consecutive time-series data. They perform deep feature mining and temporal extrapolation on the operational state feature set, generating a quantified fault probability distribution vector. Corresponding components in the probability distribution vector are selected according to a set threshold, and corresponding monitoring nodes are located based on component matching relationships. Quantified values ​​are used as the basis for node anomaly judgment, refining the hierarchical division of fault identification, reducing the intervention of human experience in fault judgment, and making the matching judgment of fault location more closely aligned with the data's own changing patterns. The generation of early warning indicators is naturally triggered by the data quantification results. Attached Figure Description

[0017] Figure 1 This is a flowchart of the power system fault prediction method based on digital twins as described in this invention; Figure 2 A flowchart illustrating the process of collecting synchronous electrical measurement data as the raw measurement sequence; Figure 3 A flowchart illustrating the work involved in constructing a digital twin model that operates synchronously with the target power system. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] See Figure 1The specific implementation of the power system fault prediction method based on digital twins is as follows: Synchronous electrical measurement data of each monitoring node in the target power system under continuous time windows are collected as the original measurement sequence. Based on the original measurement sequence, a digital twin model that operates synchronously with the target power system is constructed. In the digital twin model, dynamic state estimation of the original measurement sequence at multiple time scales is performed to generate a current operating state feature set. The current operating state feature set is input into a pre-trained temporal convolutional network to output a fault probability distribution vector. Based on the components in the fault probability distribution vector that exceed a set threshold, the fault prediction node is located and a fault warning label is generated.

[0020] In one embodiment of the present invention, the acquisition process corresponding to the original measurement sequence is described in the following reference. Figure 2 Synchronous phasor measurement devices are deployed at each bus node and both ends of each transmission line in the target power system. Each synchronous phasor measurement device is controlled to collect four electrical parameters—voltage amplitude, voltage phase angle, current amplitude, and current phase angle—at the same sampling frequency and on the same time reference. All electrical parameters collected from all synchronous phasor measurement devices are aligned according to timestamps and organized into a multi-channel time series matrix. The row index of this multi-channel time series matrix is ​​the time sampling point, and the column index is a combination of the monitoring node identifier and the electrical parameter type. Missing value detection is performed on the multi-channel time series matrix, and the entire row of data corresponding to the time sampling point with missing value is removed. The multi-channel time series matrix after removing missing values ​​is divided into multiple sub-matrices under multiple consecutive time windows, and each sub-matrix is ​​used as an original measurement sequence.

[0021] In practical implementation, a 110 kV substation power system is used as the target power system. This target power system includes 6 bus nodes and 8 transmission lines. A synchronous phasor measurement device (TPD) is deployed at each bus node and at both ends of each transmission line. All TPDs receive time synchronization signals from the Global Positioning System (GPS). Each TPD is controlled to collect four electrical parameters—voltage amplitude, voltage phase angle, current amplitude, and current phase angle—at the same time reference, with a sampling frequency of 100 frames per second. The units for voltage amplitude are kilovolts, voltage phase angle is degrees, current amplitude is kiloamperes, and current phase angle is degrees. All electrical parameters collected from all TPDs are aligned according to timestamps and organized into a multi-channel time-series matrix. The row index of this multi-channel time-series matrix represents the time sampling point, and the column index represents the combination of the monitoring node identifier and the electrical parameter type. The elements of the multi-channel time-series matrix are represented as follows:

[0022] in: Indicates the first The index number of each time sampling point Indicates the first The identification number of each monitoring node, Indicates the electrical parameter type and the set of values ​​is These correspond to voltage amplitude, voltage phase angle, current amplitude, and current phase angle, respectively. Indicates the first The time sampling point starts from the th time point The data collected by the monitoring node is the first The numerical values ​​of electrical parameters. In some embodiments, missing value detection is performed on the multi-channel timing matrix. Each row of the multi-channel timing matrix is ​​traversed. If any element in a row has a null value or is marked as invalid data, the data for the entire time sampling point corresponding to that row is determined to have missing values, and the entire row of data is removed from the multi-channel timing matrix. Optionally, missing value detection uses a flag bit judgment method. Each frame of sampled data is written to the multi-channel timing matrix simultaneously with a data validity flag bit. A data validity flag bit of 1 indicates that all electrical parameters at that sampling point are valid, and a data validity flag bit of 0 indicates that a missing value exists. It can be understood that removing the entire row of data with missing values ​​yields a complete multi-channel timing matrix without missing values. In specific implementation, a sliding window is performed on the multi-channel time-series matrix after removing missing values ​​to obtain multiple sub-matrices under consecutive time windows. The length of the sliding window is set to 100 consecutive time sampling points, and the sliding step size is 50 time sampling points. Starting from the first row of the multi-channel time-series matrix after removing missing values, 100 consecutive rows are extracted each time to form a sub-matrix, and then 50 rows are slid forward to continue extracting, repeating this operation until all rows have been traversed. In some embodiments, each extracted sub-matrix is ​​used as a raw measurement sequence, and each raw measurement sequence corresponds to the synchronous measurement data of all monitoring nodes on all electrical parameter types within a consecutive time window. Optionally, each raw measurement sequence is stored in the form of a three-dimensional array, with the first dimension being the sampling point index within the time window, the second dimension being the monitoring node identifier, and the third dimension being the electrical parameter type. It can be understood that multiple raw measurement sequences are arranged in the order of the time windows to form a time series queue of raw measurement sequences, which are used in chronological order by the subsequent digital twin model construction steps.

[0023] In one embodiment of the present invention, the construction process of the corresponding digital twin model is described in [reference needed]. Figure 3The system acquires offline topology parameters and offline component electrical parameters corresponding to the target power system. An initial framework for the node admittance matrix is ​​established based on the offline topology parameters. An initial admittance value is assigned to each non-zero element in the node admittance matrix based on the offline component electrical parameters. Each original measurement sequence is sequentially input into the calculation engine corresponding to the node admittance matrix according to the time window sequence. The continuously arriving original measurement sequences are buffered and sorted using a message queue, and input is provided to the calculation engine in a first-in-first-out manner. Within each time window, the real-time admittance correction value of each line is back-calculated using the voltage and current phasors in the original measurement sequence. The real-time admittance correction value is superimposed on the initial admittance value of the node admittance matrix to form a time-varying node admittance matrix. The time-varying node admittance matrix and the topology connection relationship of the target power system are encapsulated together into an executable simulation model as a digital twin model.

[0024] In specific implementation, a 110 kV substation power system containing 6 bus nodes and 8 transmission lines is used as the target power system. Offline topology parameters and offline component electrical parameters corresponding to this target power system are obtained. The offline topology parameters include the connection relationships between bus nodes and the endpoint information of each transmission line. The offline component electrical parameters include the resistance per unit length, reactance per unit length, susceptance per unit length, and leakage impedance of the transformer windings for each transmission line. Based on the offline topology parameters, a 6x6 node admittance matrix initial framework is established. The positions of non-zero elements in this initial framework correspond to the self-admittance positions between two bus nodes with direct electrical connections and within each bus node itself. In some embodiments, an initial admittance value is assigned to each non-zero element in the node admittance matrix based on the offline component electrical parameters. For the i-th bus node, the initial self-admittance value is equal to the sum of the admittances of all branches connected to that bus. For the branch connecting the i-th bus node and the j-th bus node, the initial mutual admittance value is equal to the negative of the branch's admittance. Optionally, when assigning the initial admittance value, the transmission line adopts the π-type equivalent circuit parameters, and the transformer adopts the model of an ideal transformer with series leakage reactance. It can be understood that after assigning the initial admittance value, a complex nodal admittance matrix initial framework is obtained, the dimension of which is equal to the total number of bus nodes.

[0025] In practical implementation, each raw measurement sequence is sequentially input into the calculation engine corresponding to the node admittance matrix according to the time window order. A message queue buffers and sorts the continuously arriving raw measurement sequences, providing input to the calculation engine in a first-in, first-out manner. The message queue is implemented using Kafka message middleware. Each raw measurement sequence is encapsulated into a message body containing a time window identifier and all synchronous electrical measurement data within that time window. In some embodiments, within each time window, the real-time admittance correction value for each line is inversely calculated using the voltage and current phasors from the raw measurement sequences. The calculation formula for this inverse calculation is as follows:

[0026] in: Indicates that the index is in the time window. Connect within the time window The first bus node and the first Real-time admittance correction value of the line at each bus node. This indicates the sequence obtained from the original measurement sequence from the first... The flow from the first bus node to the... Current phasors of each bus node and These represent the numbers obtained from the original measurement sequence. The first bus node and the first Voltage phasors of each bus node Indicates the connection of the first The first bus node and the first The initial admittance value of the line at the first bus node. Optionally, for transformer branches, the influence of the turns ratio needs to be considered during the reverse calculation, and the transformer branch should be converted into an equivalent admittance form before correction. It can be understood that the real-time admittance correction value reflects the changes in the parameters of power system components with operating conditions, including the change in resistance caused by temperature changes and the change in inductive reactance caused by reactance saturation. In specific implementation, the real-time admittance correction value is superimposed on the initial admittance value of the node admittance matrix to form a time-varying node admittance matrix. For the connection of the first bus node... The first bus node and the first For branches of each bus node, the corresponding mutual admittance elements in the time-varying node admittance matrix are updated to... The self-admittance of each bus node is updated to the negative of the sum of the updated mutual admittances of all branches connected to that bus.

[0027] In some embodiments, the time-varying node admittance matrix and the topological connectivity of the target power system are jointly encapsulated into an executable simulation model as a digital twin model. The encapsulation process adopts the functional mock cell standard format, writing the time-varying node admittance matrix into the model file in a sparse matrix storage format, and simultaneously writing the topological connectivity in the form of an edge list into the same model file. Optionally, the executable simulation model has an external input interface for updating all non-zero elements of the time-varying node admittance matrix in each time window. It can be understood that the digital twin model runs synchronously with the target power system, and the time-varying node admittance matrix in the digital twin model is updated in real time based on the current original measurement sequence within each time window.

[0028] In one embodiment of the present invention, corresponding to the process of multi-timescale dynamic state estimation and generation of operating state feature sets, an extended Kalman filter is set for each monitoring node in the digital twin model. The measurement value of the current time window in the original measurement sequence is used as the observation input and fed into the corresponding extended Kalman filter. A prediction step is performed in each extended Kalman filter to obtain the state prior estimate at the current time. An update step is performed in each extended Kalman filter to calculate the residual between the state prior estimate and the observation input and to correct the state posterior estimate. The state posterior estimates output by all extended Kalman filters are collected into a state vector under a time window. The set of state vectors under multiple consecutive time windows is subjected to cubic spline interpolation along the time dimension to obtain a continuous state trajectory with millisecond-level resolution. The state change slope of each monitoring node at the start and end points of the time window is extracted from the continuous state trajectory as a fast dynamic feature component. The state change slope is calculated by linearly fitting a preset number of sampling points near the start and end points within the time window using the least squares method. The state mean and variance of each monitoring node within the entire time window are extracted from the continuous state trajectory as steady-state feature components. The fast dynamic feature components and steady-state feature components are concatenated into a one-dimensional feature vector as the current running state feature set.

[0029] In specific implementation, a 110 kV substation power system with six bus nodes is used as the target power system. In the digital twin model, an extended Kalman filter is set for each monitoring node, resulting in six extended Kalman filters corresponding to each bus node. The measured values ​​of the current time window in the original measurement sequence are used as observation inputs and fed into the corresponding extended Kalman filters. The state variables of each extended Kalman filter are set as the real and imaginary parts of the voltage phasor of that bus node, and the observation variables are set as the voltage amplitude, voltage phase angle, and current amplitude and phase angle of each branch connected to that bus node obtained from the original measurement sequence. In some embodiments, a prediction step is performed within each extended Kalman filter to obtain the state prior estimate for the current time. The prediction step recursively extrapolates the state posterior estimate from the previous time step to the current time step using the state transition matrix based on the power system dynamic model, and calculates the state prior estimate error covariance matrix. Optionally, the state transition matrix is ​​approximated using a constant impedance load model, assuming that the state variables of the bus node remain unchanged between adjacent time sampling points. It is understandable that the state prior estimate reflects the prediction result of the electrical state of the bus node at the current moment based solely on the system dynamic model.

[0030] In specific implementation, an update step is performed within each extended Kalman filter to calculate the residual between the state prior estimate and the observed input, and to correct the state posterior estimate. The update step calculates the Kalman gain matrix, uses the difference between the observed input and the state prior estimate as the residual, multiplies the residual by the Kalman gain matrix, and adds it to the state prior estimate to obtain the state posterior estimate. Simultaneously, the state posterior estimate error covariance matrix is ​​updated. In some embodiments, the state posterior estimates output by all extended Kalman filters are collected into a set of state vectors within a time window. Each time window contains 100 consecutive sampling points. The collected state posterior estimates are arranged in the order of the sampling points to form a 6-row, 100-column set of state vectors, where each column corresponds to the state posterior estimates of all 6 bus nodes at a sampling time. Optionally, the state posterior estimates are stored in complex form, where the real part represents the real part of the voltage phasor, and the imaginary part represents the imaginary part of the voltage phasor. It can be understood that the set of state vectors within a time window contains the electrical states of all monitoring nodes after filtering correction at each sampling time within that time window.

[0031] In specific implementation, cubic spline interpolation is performed along the time dimension on the set of state vectors under multiple consecutive time windows to obtain continuous state trajectories with millisecond-level resolution. The interpolation target resolution is set to 1 millisecond, and the original sampling frequency is 100 frames per second, i.e., a sampling interval of 10 milliseconds. The cubic spline interpolation function generates 9 interpolation points between every two adjacent sampling points. In some embodiments, the state change slope of each monitoring node at the start and end points of the time window is extracted from the continuous state trajectory as a fast dynamic feature component. The state change slope is calculated by linearly fitting a preset number of sampling points near the start and end points within the time window using the least squares method. The preset number of sampling points is 10% of the total number of sampling points within the time window, i.e., the first 10 original sampling points and their corresponding interpolation points are taken near the start point, and the last 10 original sampling points and their corresponding interpolation points are taken near the end point. The objective function expression for the linear fitting is:

[0032] in: This represents the slope of the extracted state change. This represents the total number of sampling points involved in the fitting and , Indicates the first Time coordinates of each sampling point Indicates the first The posterior state estimate of each sampling point. Optionally, the state change slope is calculated by separately calculating the slope of the change of the real part of the voltage phasor and the slope of the change of the change of the imaginary part of the voltage phasor, and the two are combined as the fast dynamic characteristic component of the monitoring node. It can be understood that the fast dynamic characteristic component reflects the degree of abrupt change of the electrical quantity of the monitoring node at the boundary of the time window, and is sensitive to the rapid changes of electrical quantity when a fault occurs.

[0033] In specific implementation, the mean and variance of the state of each monitoring node within the entire time window are extracted from the continuous state trajectory as steady-state feature components. The state mean is equal to the arithmetic mean of the posterior state estimates of all sampling points and interpolation points within the time window, and the state variance is equal to the sum of the squares of the differences between the posterior state estimates of each sampling point and interpolation point and the mean, divided by the total number of sampling points. In some embodiments, the fast dynamic feature components and steady-state feature components are concatenated into a one-dimensional feature vector as the current operating state feature set. For a power system with 6 bus nodes, each monitoring node provides a state change slope component and a steady-state feature component. The fast dynamic feature component includes two values: the real part slope of voltage and the imaginary part slope of voltage. The steady-state feature component includes two values: the state mean and the state variance. Therefore, each monitoring node contributes 4 feature values, and the 6 monitoring nodes obtain a total of 24 feature values. Optionally, the concatenation order is arranged in ascending order according to the monitoring node number, and within each monitoring node, the order is arranged as follows: real part slope of voltage, imaginary part slope of voltage, state mean, and state variance. It can be understood that the current operating state feature set encapsulates the dynamic response characteristics and steady-state operating characteristics of the target power system within the corresponding time window in the form of a one-dimensional vector.

[0034] In one embodiment of the present invention, corresponding to the process of outputting a fault probability distribution vector by a temporal convolutional network, the structural parameters of a pre-trained temporal convolutional network model are obtained. This temporal convolutional network model includes multiple dilated convolutional layers, each followed by a batch normalization layer and an activation function layer. The current running state feature set is arranged into a two-dimensional input tensor according to the time window order. The first dimension of this two-dimensional input tensor is the time window index, and the second dimension is the one-dimensional feature vector corresponding to each time window. The two-dimensional input tensor is sequentially passed through multiple dilated convolutional layers to extract fault association feature maps under different temporal receptive fields. In each batch normalization layer, the output fault association feature map is processed by distribution normalization. In each activation function layer, the normalized fault association feature map is processed by nonlinear mapping. The fault association feature map output by the last activation function layer is input into the fully connected layer to reduce the dimensionality to the dimension of the number of fault categories. The dimensionality reduction result output by the fully connected layer is processed by normalized exponential function mapping to use the probability value of each fault category as the fault probability distribution vector. The training steps of the temporal convolutional network model include constructing samples using historical fault events and corresponding operating state feature sets, and optimizing the model parameters using stochastic gradient descent combined with the focus loss function.

[0035] In the specific implementation, a 110 kV substation power system with 6 bus nodes is used as the target power system. The number of fault categories is set to 5 types: three-phase short-circuit fault, single-phase ground fault, two-phase short-circuit fault, two-phase ground fault, and open-circuit fault. The structural parameters of a pre-trained temporal convolutional network model are obtained. The temporal convolutional network model contains 4 dilated convolutional layers, each followed by a batch normalization layer and an activation function layer. The dilation factors of the dilated convolutional layers are 1, 2, 4, and 8, respectively, and the kernel size is set to 3. The number of output channels for each dilated convolutional layer is set to 64. In some embodiments, the current operating state feature set is arranged into a two-dimensional input tensor according to the time window order. The first dimension of the two-dimensional input tensor is the time window index, and the second dimension is the one-dimensional feature vector corresponding to each time window. The one-dimensional feature vector generated by each time window has a dimension of 24. A total of 100 consecutive time windows are collected, therefore the shape of the two-dimensional input tensor is 100 rows and 24 columns. Optionally, the time window indexes are arranged in ascending order of time, with the oldest time window in the first row and the newest time window in the last row. This can be understood as the two-dimensional input tensor encapsulating the evolution of the target power system's operating state over 100 consecutive time windows.

[0036] In specific implementations, the two-dimensional input tensor is sequentially passed through multiple dilated convolutional layers to extract fault-related feature maps under different temporal receptive fields. The first dilated convolutional layer uses a dilation factor of 1 to convolve the input tensor, resulting in an output feature map with a size of 100 rows and 64 columns. The second dilated convolutional layer takes the output of the first layer as input and uses a dilation factor of 2, maintaining the same output feature map size. The third and fourth dilated convolutional layers process the same data sequentially. Each dilated convolutional layer maintains the temporal length of the feature map through padding operations in the temporal dimension. In some embodiments, the output fault-related feature map is subjected to distribution normalization in each batch normalization layer. The batch normalization layer calculates the mean and variance of all samples in each feature channel, normalizing the feature map values ​​to a zero-mean, unit-variance distribution and introducing learnable scaling and translation factors. Optionally, the initial value of the scaling factor in the batch normalization layer is set to 1, and the initial value of the translation factor is set to 0. It can be understood that distribution normalization accelerates the convergence speed during model training and reduces internal covariate bias.

[0037] In specific implementations, a nonlinear mapping is performed on the standardized fault association feature map in each activation function layer. The activation function layer uses a modified linear unit function, which sets all negative elements in the standardized fault association feature map to zero, while leaving positive elements unchanged. In some embodiments, the fault association feature map output from the last activation function layer is input into a fully connected layer to reduce the dimensionality to the dimension of the number of fault categories. The fault association feature map output from the last activation function layer has a size of 100 rows and 64 columns. The fully connected layer reduces the second dimension from 64 to 5, and the size of the output two-dimensional tensor becomes 100 rows and 5 columns. Optionally, the weight matrix of the fully connected layer is initialized with a normal distribution, and the bias vector is initialized to zero. It can be understood that the fully connected layer maps the high-dimensional fault association features extracted by the dilated convolutional layer to the fault category space. In practice, the dimensionality reduction result output by the fully connected layer is mapped by a normalized exponential function to use the probability value of each fault category as the fault probability distribution vector. The normalized exponential function converts each row of the 100-row, 5-column two-dimensional tensor output by the fully connected layer into a probability distribution vector. Each probability distribution vector contains 5 components, each component corresponds to the probability value of a fault category, and the sum of all components equals 1.

[0038] In practical implementation, the training steps of the temporal convolutional network model include constructing samples using historical fault events and corresponding operational state feature sets. Each sample uses the operational state feature set of 100 consecutive time windows before the historical fault occurred as the input feature vector, and the actual fault type label after the fault occurred as the supervision signal. The label adopts one-hot encoding, and its dimension is the same as the number of fault categories. In some embodiments, stochastic gradient descent combined with a focal loss function is used to optimize the model parameters. The expression of the focal loss function is:

[0039] in: This represents the probability value predicted by the model for the true fault category. The weighting coefficients represent the balance between the importance of positive and negative samples and range from 0 to 1. The focus parameter is set to 2.0. Optionally, the initial learning rate for stochastic gradient descent is set to 0.001, the momentum parameter to 0.9, and the weight decay coefficient to 0.0001. It can be understood that the focus loss function reduces the loss contribution weight of easily classified samples, causing the model training process to focus more on difficult-to-classify faulty samples.

[0040] In one embodiment of the present invention, corresponding to the fault prediction node location and fault warning identifier generation process, a probability warning threshold is pre-set for each fault type. The probability warning threshold is calculated and dynamically adjusted based on the frequency of each fault type in historical samples and the cost loss function of missed and false alarms. The relationship between each component in the fault probability distribution vector and the probability warning threshold of the corresponding fault type is compared one by one. All fault types with component values ​​greater than the corresponding probability warning threshold are collected as a candidate warning fault set. The fault type with the largest component value is selected from the candidate warning fault set as the main warning fault type. The list of associated monitoring nodes corresponding to the main warning fault type in the target power system is queried. The three monitoring nodes in the associated monitoring node list that are closest to the typical occurrence location of the fault type are marked as fault prediction nodes. A warning identifier containing the name of the main warning fault type, the identifier of the fault prediction node, and the current time window index is generated. The warning identifier is output as a fault warning identifier.

[0041] In specific implementation, a 110 kV substation power system containing 6 busbar nodes and 8 transmission lines is used as the target power system. Five fault types are set: three-phase short-circuit fault, single-phase ground fault, two-phase short-circuit fault, two-phase ground short-circuit fault, and open-circuit fault. A probability warning threshold is pre-set for each fault type: 0.85 for three-phase short-circuit fault, 0.70 for single-phase ground fault, 0.80 for two-phase short-circuit fault, 0.75 for two-phase ground short-circuit fault, and 0.65 for open-circuit fault. In some embodiments, the corresponding probability warning threshold is calculated and dynamically adjusted based on the frequency of each fault type in historical samples and the cost-loss function of missed and false alarms. The cost-loss function of missed and false alarms is defined as:

[0042] in: Indicates the total cost or loss. Indicates the total number of fault types, with a value of 5, and the subscript. Indicates the first Index of various fault types Indicates the first The cost coefficient for missed reports of each type of fault. Indicates the first The frequency of missed reports for each type of fault Indicates the first The false alarm cost coefficient for each type of fault. Indicates the first The false alarm frequency for each fault type. Optionally, the underreporting cost coefficient is set according to the degree of equipment damage caused by the fault type: the underreporting cost coefficient for three-phase short-circuit faults is 10, for single-phase ground faults it is 3, for two-phase short-circuit faults it is 7, for two-phase ground short-circuit faults it is 6, for open-circuit faults it is 5, and the false alarm cost coefficient is uniformly set to 1. It can be understood that the probability warning threshold that minimizes the total cost loss during dynamic adjustment is used as the final setting value.

[0043] In specific implementation, the probability distribution vector of each component is compared with the probability warning threshold of the corresponding fault type. The fault probability distribution vector output by the temporal convolutional network model contains five components, corresponding to the probability values ​​of three-phase short-circuit faults, single-phase ground faults, two-phase short-circuit faults, two-phase ground short-circuit faults, and open-circuit faults, respectively. Each probability value is compared with the probability warning threshold of the corresponding fault type. In some embodiments, all fault types with component values ​​greater than the corresponding probability warning threshold are collected as a candidate warning fault set. If the probability value of a three-phase short-circuit fault is greater than 0.85, it is added to the candidate warning fault set; if the probability value of a single-phase ground fault is greater than 0.70, it is added to the candidate warning fault set, and so on, traversing all five fault types. Optionally, if no component value is greater than the corresponding probability warning threshold, no warning is generated in this prediction, and the system waits for the fault probability distribution vector input in the next time window. It can be understood that the candidate warning fault set contains all potential fault types that meet the warning conditions.

[0044] In specific implementation, the fault type with the largest component value is selected from the candidate early warning fault set as the primary early warning fault type. Each fault type in the candidate early warning fault set is traversed, and its corresponding probability value is compared. The fault type with the highest probability value is determined as the primary early warning fault type. In some embodiments, the associated monitoring node list corresponding to the primary early warning fault type in the target power system is queried. An associated monitoring node list is pre-established for each fault type. The associated monitoring node list for a three-phase short-circuit fault includes all six bus nodes; the associated monitoring node list for a single-phase ground fault includes the measurement points corresponding to specific phases of the three phases of each bus node; the associated monitoring node lists for two-phase short-circuit faults and two-phase ground short-circuit faults include the bus nodes involved in the two short-circuit phases; and the associated monitoring node list for a line break fault includes the bus nodes at both ends of the broken line. Optionally, the associated monitoring node list is stored in a database table, with the fault type code as the primary key and the monitoring node identifier as the foreign key. It can be understood that the associated monitoring node list determines the range of monitoring nodes that may be affected by each fault type.

[0045] In specific implementation, the three monitoring nodes closest to the typical location of the fault type in the associated monitoring node list are marked as fault prediction nodes. For a three-phase short-circuit fault, the typical location of the fault is the midpoint of the transmission line or a bus node. The spatial distance from each associated monitoring node to the typical location of the fault is calculated, and the top three monitoring nodes are selected as fault prediction nodes according to the distance from smallest to largest. For a single-phase ground fault, the typical location of the fault is a point on the transmission line. Three monitoring nodes are selected: the bus nodes at both ends of the line corresponding to the fault phase and the intermediate node closest to the fault point. In some embodiments, a warning identifier is generated, which includes the name of the main warning fault type, the identifier of the fault prediction node, and the current time window index. The warning identifier adopts a structured string format, such as "main_fault_type:three-phase short-circuit fault|predicted_nodes:Node_01,Node_02,Node_03|window_index:00087". Optionally, the warning identifier also includes the probability value of the corresponding main warning fault type from the fault probability distribution vector. It can be understood that the warning identifier encapsulates the core information of fault prediction, making it easier for maintenance personnel to quickly locate the fault. In practice, the warning sign is output as a fault warning sign. The fault warning sign is sent to the human-machine interface of the power system monitoring master station, and at the same time, an audible and visual alarm signal is triggered.

[0046] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A power system fault prediction method based on digital twins, characterized in that, include: Synchronous electrical measurement data of each monitoring node in the target power system within a continuous time window are collected as the raw measurement sequence. A digital twin model that operates synchronously with the target power system is constructed based on the original measurement sequence. In the digital twin model, multi-timescale dynamic state estimation is performed on the original measurement sequence to generate a current operating state feature set; The current operating state feature set is input into a pre-trained temporal convolutional network to output a fault probability distribution vector; Based on the components in the fault probability distribution vector that exceed a set threshold, fault prediction nodes are located and fault warning identifiers are generated.

2. The power system fault prediction method based on digital twins according to claim 1, characterized in that, The steps for collecting synchronous electrical measurement data from various monitoring nodes in the target power system within a continuous time window as the raw measurement sequence specifically include: Synchronous phasor measurement devices are deployed at each bus node and both ends of each transmission line in the target power system. Each synchronous phasor measurement device is controlled to acquire four electrical parameters—voltage amplitude, voltage phase angle, current amplitude, and current phase angle—at the same sampling frequency and on the same time reference. All electrical parameters collected from all synchronous phasor measurement devices will be aligned according to timestamps and organized into a multi-channel timing matrix; The row index of the multi-channel time series matrix is ​​the time sampling point, and the column index is a combination of the monitoring node identifier and the electrical parameter type; Missing value detection is performed on the multi-channel time series matrix, and the entire row of data corresponding to the time sampling point with missing value is removed; A sliding window segmentation is performed on the multi-channel time series matrix after removing missing values ​​to obtain sub-matrices under multiple consecutive time windows; Each submatrix is ​​used as a sequence of original measurements.

3. The power system fault prediction method based on digital twins according to claim 2, characterized in that, The steps of constructing a digital twin model that operates synchronously with the target power system based on the original measurement sequence specifically include: Obtain the offline topology parameters and offline component electrical parameters corresponding to the target power system; An initial framework for establishing the node admittance matrix is ​​established based on the offline topology parameters; Assign an initial admittance value to each non-zero element in the node admittance matrix based on the electrical parameters of the offline components; Each of the original measurement sequences is input into the calculation engine corresponding to the node admittance matrix in chronological order according to the time window. Within each time window, the real-time admittance correction value for each line is inversely calculated using the voltage phasor and current phasor in the original measurement sequence; The real-time admittance correction value is superimposed on the initial admittance value of the node admittance matrix to form a time-varying node admittance matrix; The time-varying node admittance matrix and the topological connection relationship of the target power system are encapsulated together into an executable simulation model as the digital twin model.

4. The power system fault prediction method based on digital twins according to claim 3, characterized in that, The step of sequentially inputting each of the original measurement sequences into the computation engine corresponding to the node admittance matrix according to the time window order specifically includes: buffering and sorting the continuously arriving original measurement sequences through a message queue, and providing input to the computation engine in a first-in-first-out manner.

5. The power system fault prediction method based on digital twins according to claim 3, characterized in that, The step of performing multi-timescale dynamic state estimation on the original measurement sequence in the digital twin model to generate a current operating state feature set specifically includes: In the digital twin model, an extended Kalman filter is set for each monitoring node; The measurement value of the current time window in the original measurement sequence is used as the observation input and fed into the corresponding extended Kalman filter; A prediction step is performed within each extended Kalman filter to obtain the state prior estimate at the current time. An update step is performed within each extended Kalman filter to compute the residual between the state prior estimate and the observed input and to correct the state posterior estimate. Collect the posterior state estimates of all extended Kalman filter outputs into a set of state vectors within a time window; Perform cubic spline interpolation along the time dimension on the set of state vectors under multiple consecutive time windows to obtain continuous state trajectories with millisecond-level resolution; The slope of the state change of each monitoring node at the start and end points of the time window is extracted from the continuous state trajectory as a fast dynamic feature component. Extract the mean and variance of the state of each monitoring node within the entire time window from the continuous state trajectory as steady-state feature components. The fast dynamic feature components and the steady-state feature components are concatenated into a one-dimensional feature vector, which is used as the current running state feature set.

6. The power system fault prediction method based on digital twins according to claim 5, characterized in that, In the step of extracting the state change slope of each monitoring node at the start and end points of the time window as a fast dynamic feature component from the continuous state trajectory, the state change slope is calculated by linearly fitting a preset number of sampling points near the start and end points within the time window using the least squares method.

7. The power system fault prediction method based on digital twins according to claim 5, characterized in that, The step of inputting the current operating state feature set into a pre-trained temporal convolutional network to output a fault probability distribution vector specifically includes: Obtain the structural parameters of the pre-trained temporal convolutional network model; The temporal convolutional network model includes multiple dilated convolutional layers, each followed by a batch normalization layer and an activation function layer. Arrange the current running state feature set into a two-dimensional input tensor according to the time window order; The first dimension of the two-dimensional input tensor is the time window index, and the second dimension is the one-dimensional feature vector corresponding to each time window. The two-dimensional input tensor is sequentially passed through multiple dilated convolutional layers to extract fault association feature maps under different receptive fields at different times. In each batch normalization layer, the output fault association feature map is subjected to distribution normalization. In each activation function layer, a nonlinear mapping is performed on the standardized fault association feature map; The fault association feature map output from the last activation function layer is input into the fully connected layer to reduce the dimensionality to the dimension of the number of fault categories; The dimensionality reduction result of the fully connected layer is subjected to a normalized exponential function mapping to use the probability value of each fault category as the fault probability distribution vector.

8. The power system fault prediction method based on digital twins according to claim 7, characterized in that, The training steps of the temporal convolutional network model include: constructing samples using historical fault events and corresponding running state feature sets, and optimizing the model parameters using stochastic gradient descent combined with a focus loss function.

9. The power system fault prediction method based on digital twins according to claim 7, characterized in that, The steps of locating fault prediction nodes and generating fault warning identifiers based on components exceeding a set threshold in the fault probability distribution vector specifically include: A probability warning threshold is pre-set for each type of fault; Compare the magnitudes of each component in the fault probability distribution vector with the probability warning threshold of the fault type corresponding to that component; Collect all fault types whose component values ​​are greater than the corresponding probability warning threshold as a candidate warning fault set; The fault type with the largest component value is selected from the candidate early warning fault set as the main early warning fault type; Query the list of associated monitoring nodes corresponding to the main early warning fault type in the target power system; Mark the three monitoring nodes in the associated monitoring node list that are closest to the typical location of the fault type as fault prediction nodes; Generate an early warning identifier that includes the name of the main early warning fault type, the identifier of the fault prediction node, and the current time window index; The warning sign is output as the fault warning sign.

10. The power system fault prediction method based on digital twins according to claim 9, characterized in that, The step of pre-setting a probability warning threshold for each fault type specifically includes: calculating and dynamically adjusting the corresponding probability warning threshold based on the frequency of each fault type in historical samples and the cost loss function of missed and false alarms.

Images