A fault monitoring method and device for a smart grid

Abstract

The application discloses a kind of fault monitoring method and device of smart grid, belong to electric power system wide-area monitoring and fault diagnosis technical field.Method includes: synchronous acquisition grid each node measurement data and pre-processing, obtain space-time data matrix;Extract time-frequency characteristics by adaptive fractional order S transform, wherein fractional order parameter is adjusted online by prediction network according to signal kurtosis;Time-frequency characteristics and grid topology admittance are input into physical perception graph neural differential equation network, and latent state continuous evolution is described using graph attention mechanism and ordinary differential equation, and reconstructed time-frequency characteristics, reconstructed electrical quantity and theoretical injection power are decoded and output;Calculate multi-dimensional node-level residual and obtain global fault indication score by fusion, determine fault location after fault is determined by adaptive threshold value.The application integrates adaptive time-frequency transform, physical constraint graph neural differential equation and multi-dimensional residual fusion organically, and improves the adaptability and reliability of wide-area grid fault detection and positioning under complex conditions.

Description

Technical Field

[0001] This invention relates to the field of wide-area monitoring and fault diagnosis technology for power systems, and specifically to a fault monitoring method and device for smart grids. Background Technology

[0002] With the continuous expansion of power system scale and the sustained increase in the penetration rate of new energy sources, the operation mode of power grid is becoming increasingly complex, and the forms of various disturbances and faults exhibit obvious non-stationarity and uncertainty. Wide-area monitoring systems based on synchronous phasor measurement devices provide a data foundation for real-time perception of the dynamic behavior of the power grid. However, how to quickly and accurately detect and locate faults from multi-node, high-dimensional synchronous measurement data remains a technical challenge to ensure the safe and stable operation of the power grid.

[0003] Currently, fault detection methods for wide-area power grids mainly fall into two categories: model-driven and data-driven. Model-driven methods rely on precise mathematical models of the power grid and preset thresholds, making them susceptible to topological changes and operating point offsets in actual operation, and threshold tuning is quite difficult. Data-driven methods utilize machine learning or deep learning techniques to automatically extract fault features from historical data, but they suffer from shortcomings in interpretability and generalization ability, especially when fault samples are scarce in actual power grids, making it difficult to guarantee the representativeness of training data. Furthermore, existing time-frequency analysis methods typically employ fixed-parameter transformation tools when extracting non-stationary disturbance features, making it difficult to achieve an adaptive balance between frequency resolution and time resolution.

[0004] In recent years, hybrid modeling methods that integrate physical knowledge with deep learning have gradually attracted attention. However, in the field of wide-area power grid fault detection, there is still a lack of effective solutions on how to organically integrate the topology of the power grid, electromechanical transient physical equations and continuous-time deep dynamic models, and achieve end-to-end collaborative optimization from feature extraction to fault determination.

[0005] To address the aforementioned issues, there is an urgent need for a fault monitoring method and device for smart grids to solve the problems associated with traditional methods. Summary of the Invention

[0006] The purpose of this invention is to provide a fault monitoring method and device for smart grids, which organically integrates adaptive time-frequency transformation, physical constraint graph neural differential equations and multidimensional residual fusion, thereby improving the adaptability and reliability of wide-area power grid fault detection and location under complex operating conditions.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A fault monitoring method for a smart grid, comprising: Step 1: Synchronously collect measurement data from each node of the power grid and preprocess it to obtain a spatiotemporal data matrix and electrical quantity vectors at corresponding times; Step 2: Based on the spatiotemporal data matrix, extract time-frequency features node by node through adaptive fractional S-transform to obtain the global time-frequency feature vector; Step 3: Input the global time-frequency feature vector and the power grid topology admittance parameters into the physical sensing graph neural differential equation network. Aggregate the potential states of neighboring nodes through the graph attention mechanism and use ordinary differential equations to describe the continuous time evolution of the potential states. The reconstructed time-frequency features, reconstructed voltage amplitude, and reconstructed phase angle are output through the dual-branch decoder. The theoretical injected power is calculated based on the reconstructed voltage amplitude, reconstructed phase angle, and admittance matrix. Step 4: Based on the reconstructed time-frequency characteristics, reconstructed voltage amplitude, reconstructed phase angle, theoretical injected power, and measurement data, calculate the node-level residuals in multiple dimensions, fuse them to obtain the global fault indication score, and compare it with the adaptive threshold to determine the fault.

[0008] Furthermore, in step 1, measurement data from each node of the power grid are synchronously collected and preprocessed to obtain a spatiotemporal data matrix and electrical quantity vectors at corresponding times, specifically: Measurement vectors containing voltage amplitude, voltage phase angle, active power and reactive power are obtained from the synchronous phasor measurement devices at each node of the power grid at a fixed sampling frequency. Zero-phase filtering is performed on the measurement vectors within each time window, and the validity of the filtered data is verified. The verified data is organized into a spatiotemporal data matrix, and the measurement values ​​at the end of the time window are extracted as electrical quantity vectors.

[0009] Furthermore, the validity verification includes amplitude over-limit detection and phase angle jump detection.

[0010] Furthermore, in step 2, the adaptive fractional S-transform employs a fractional Gaussian window function, and the fractional parameters of the adaptive fractional S-transform are adjusted online by a differentiable fractional parameter prediction network based on the local statistical characteristics of the signal.

[0011] Furthermore, in step 3, the physical perception graph neural differential equation network includes: A multi-scale temporal convolutional encoder is used to generate the initial latent state of a node based on the global time-frequency feature vectors of multiple frames; The graph attention message passing module uses the real and imaginary parts of the node admittance as edge features to calculate the attention weights between neighboring nodes and aggregate messages. The vector field function takes the latent state of a node, the aggregated message, and the time-frequency feature context as input, and outputs the rate of change of the latent state. Numerical ordinary differential equation solver that performs continuous-time integration of the latent state based on the vector field function; The dual-branch decoder decodes the reconstructed time-frequency characteristics, reconstructed voltage amplitude, and reconstructed phase angle from the latent state, respectively.

[0012] Furthermore, the training loss function of the physical perception graph neural differential equation network includes: time-frequency feature reconstruction loss, electrical quantity reconstruction loss, physical consistency loss, latent state smoothing regularization loss, and weight decay term; wherein, the physical consistency loss is constructed based on the deviation between the theoretical injected power and the measured power calculated by the reconstructed voltage amplitude and phase angle according to the AC power flow equation.

[0013] Furthermore, in step 4, the node-level residuals in multiple dimensions include: time-frequency feature reconstruction residuals, electrical quantity reconstruction residuals, and physical consistency residuals; the physical consistency residuals are calculated based on the normalized deviation between the measured active power, reactive power, and theoretical injected power. The global fault indication score is the maximum value of the local anomaly index of each node, and the local anomaly index is the weighted sum of the three-dimensional residuals; The adaptive threshold is dynamically updated based on the mean and variance of the exponentially weighted moving average of the global fault indication score.

[0014] Furthermore, step 4 also includes: After identifying the fault, the location of the fault is determined by the spatial distribution of the physical consistency residuals and the grid topology connection relationship.

[0015] The present invention also provides a fault monitoring device for a smart grid, applied to the aforementioned fault monitoring method for a smart grid, comprising: The data acquisition and preprocessing module is used to synchronously acquire measurement data from the synchronous phasor measurement devices at each node, perform zero-phase filtering and validity verification, and output spatiotemporal data matrix and electrical quantity vector. An adaptive fractional S-transform feature extraction module, connected to the data acquisition and preprocessing module, is used to perform the adaptive fractional S-transform online and generate a global time-frequency feature vector. The physical perception graph neural differential equation inference module is connected to the feature extraction module and is used to receive the global time-frequency feature vector and use the physical perception graph neural differential equation network to output the reconstructed time-frequency features, reconstructed voltage amplitude, reconstructed phase angle and theoretical injected power. The residual fault detection and location module is integrated and connected to the inference module and the data acquisition and preprocessing module. It is used to calculate multi-dimensional node-level residuals, generate global fault indication scores, perform adaptive threshold comparisons to realize fault determination, and perform fault location based on physical consistency residuals and power grid topology when a fault occurs. The auxiliary support unit includes at least a storage array, a time synchronization unit, and a management interface.

[0016] In summary, the present invention has at least one of the following beneficial technical effects: Firstly, by using an adaptive fractional S-transform feature extraction method, the adjustment of fractional parameters is incorporated into the end-to-end joint optimization framework, enabling the window shape of the time-frequency transform to be adaptively adjusted online according to the local kurtosis of the signal. This achieves higher frequency resolution in the steady oscillation segment and higher time resolution in the impulse transient segment, thereby improving the adaptability of feature extraction to complex power grid disturbance patterns.

[0017] Secondly, by embedding the admittance matrix of the power grid nodes as an edge feature into the graph attention message passing process, and introducing the AC power flow equation into the loss function as an explicit physical consistency constraint, the continuous evolution of the potential state is guided by both data and prior physical structure, which enhances the physical interpretability of the model and its generalization ability in topology change scenarios.

[0018] Third, by integrating the time-frequency feature reconstruction residual, electrical quantity reconstruction residual and physical consistency residual into a multi-dimensional anomaly measurement mechanism, and combining it with an exponentially weighted moving average adaptive threshold, the dependence on fixed thresholds is reduced, which helps to improve the robustness of fault detection under varying operating conditions.

[0019] Fourth, in the fault location stage, by constructing a centrality index that is dominated by the node's own accumulated physical residuals and supplemented by the neighbor residuals weighted by topological admittance, the interference of fault spillover effect on location accuracy can be effectively suppressed, and the accuracy of fault source identification can be improved. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0021] 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. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] This invention provides a fault monitoring device for smart grids, comprising: a data acquisition and preprocessing module, an adaptive fractional S-transform feature extraction module, a physical sensing graph neural differential equation inference module, a fusion residual fault detection and location module, and an auxiliary support unit. The modules interact with each other via a high-speed data bus and control signal lines to exchange data and control flows. The structure, connection relationships, and functions of each device are described in detail below: 1. Data Acquisition and Preprocessing Module The data acquisition and preprocessing module includes: Synchronization Phasor Measurement Unit (PMU): Employs a phasor measurement device conforming to the IEEE C37.118.2 standard, equipped with a GPS / BeiDou dual-mode timing module, achieving time synchronization accuracy superior to [previous standard]. μs. One unit is configured on each node, responsible for... The PMU synchronously acquires the three-phase instantaneous values ​​of bus voltage and line current at a fixed sampling frequency of Hz. The voltage amplitude, phase angle, frequency, and active / reactive power are calculated in real time using the DFT fundamental frequency extraction algorithm. The PMU transmits the time-stamped measurement frames through a process layer fiber optic network (IEC 61850-9-2). The data is transmitted to the data concentrator at a rate of frames per second.

[0023] Data concentrator: A real-time data processing board based on FPGA (such as the Xilinx Zynq UltraScale+), integrating... Each device has a gigabit SFP+ Ethernet interface, supporting IEC 61850 protocol parsing and IEEE C37.118.2 frame synchronization. The concentrator receives measurement data uploaded by each PMU, and implements the data alignment, zero-phase filtering, spatiotemporal data matrix splicing, and validity verification functions described in step 1 in the FPGA logic. The processed clean data is pushed to the GPU memory and system main memory via PCIe 4.0 x8 bus using DMA. Simultaneously, the device obtains absolute time from the GPS / BeiDou module through the IRIG-B interface to generate a stable time. Time step label.

[0024] The PMU is connected to the SFP+ port of the data concentrator via single-mode fiber (optical power budget meets requirements). (km transmission distance); the data concentrator uses a PCIe 4.0 x8 bus (unidirectional bandwidth approximately...) This module (GB / s) connects to the GPU accelerator card; it reads UTC time from an external GPS / BeiDou module via the I2C bus. This module performs data aggregation, filtering, and timing organization, providing a standard spatiotemporal data matrix for subsequent modules. and electrical quantity vector 2. Adaptive Fractional S-Transform Feature Extraction Module The adaptive fractional S-transform feature extraction module includes: High-performance GPU accelerator card: Employs NVIDIA A100 (or an equivalent FPGA accelerator card, such as Xilinx Alveo U280), possessing... GB HBM2e video memory and One CUDA core is used for online execution of adaptive fractional S-transform and fractional parameter prediction network inference. CUDA kernel functions are deployed on the GPU to implement batch FFT (cuFFT library), element-wise complex number operations, and frequency domain window function generation.

[0025] The parameter prediction network embeds the inference kernel: the trained lightweight feedforward network Compile to a TensorRT engine or FPGA HLS IP core, run at FP16 precision, and receive node kurtosis read from video memory. And the previous moment Real-time output of current fractional order parameters The delay for a single inference attempt does not exceed μs.

[0026] Shared memory buffer: partitioned within GPU video memory A fixed area (MB) is used to store the history of each node. Value circular queue (keeps the most recent value) Frames), intermediate results of kurtosis calculation and pre-generated Group frequency domain window coefficient This avoids redundant calculations online.

[0027] The data concentrator connects via the PCIe bus The voltage amplitude channel data is written to the GPU memory; the CUDA stream on the GPU sequentially executes the kurtosis calculation kernel function, parameter prediction network inference, frequency domain window generation kernel function, and fractional S-transform kernel function, generating the time-frequency amplitude spectrum through frequency domain multiplication and inverse FFT, and concatenating them node by node into a global time-frequency feature vector. This vector is simultaneously stored in the GPU memory for later use and mirrored to the system main memory via the PCIe bus for further use.

[0028] 3. Physical Perception Map Neural Differential Equation Reasoning Module The physical perception graph neural differential equation reasoning module includes: Neural network inference acceleration chip: Employs a dedicated NPU supporting mixed-precision inference (such as the Huawei Ascend 310 or the DLA on the NVIDIA Jetson Orin) to carry the physical perception map neural differential equation model trained and solidified in step 3 with INT8 / FP16 precision. Internally embedded is a multi-scale temporal convolutional encoder. ODE vector field function Dual-branch decoder and The total weights.

[0029] The RK4 ODE numerical solution coprocessor implements a fixed-point fourth-order Runge-Kutta integrator within the NPU using custom operators, completing the process once per substep. Forward reasoning and state accumulation, Each sub-step constitutes an integral of a complete frame period, and the integration delay is controlled within... Within ms.

[0030] Topological constant memory: On-chip SRAM area (capacity) MB) Storage of power grid node admittance matrix The real and imaginary parts of the value , Pre-computed edge feature vectors and neighbor index table , used for constant lookup in graph attention aggregation.

[0031] This module retrieves each frame from GPU memory or system main memory via the PCIe bridge. and measuring electrical quantities , The module's internal flow is as follows: Encoder (executed only on the initial frame) → ODE solver (executed every frame). Each RK4 substep (each substep includes graph attention aggregation and vector field function inference) → dual-branch decoder. Output reconstructs time-frequency features. Reconstructed voltage amplitude Reconstructing phase angle The theoretical injected power is calculated in real time by combining the admittance matrix in the topological constant memory. and All results are packaged and sent to the fault detection CPU via a high-speed SPI or PCIe interface.

[0032] 4. Integration of residual fault detection and location module The integrated residual fault detection and location module includes: Embedded multi-core CPU processor: Uses an ARM Cortex-A78 quad-core processor (or Intel Atom x6000 series), clock speed The CPU runs a real-time Linux operating system (PREEMPT_RT kernel) at GHz, responsible for executing residual calculation, EWMA adaptive threshold update, and fault location logic in step 4. The CPU incorporates the NEON SIMD instruction set for parallel processing. Vectorized residual operation for each node.

[0033] Statistical parameter memory: on-chip KB SRAM is used to store EWMA online statistics. , , Historical normal standard deviation of each node , , , And fault detection state machine and historical residual circular buffer (reserved) frame).

[0034] Alarm output interfaces: including The output of the circuit contact relay (maximum switching voltage) VAC / VDC) and One gigabit Ethernet port. The relay output is hardwired to the digital input module of the SCADA system or the local alarm light panel; the Ethernet port sends fault flags, fault indication scores, adaptive thresholds, fault node numbers and location results to the dispatch center via IEC 61850-8-1 MMS or GOOSE messages.

[0035] The CPU obtains reconstruction data from the inference module and GPU via the PCIe Gen3 x4 bus. , , , , and measurement data , , , , Online calculation of 3D node-level residuals , , And merge to obtain a global fault indication score With EWMA adaptive threshold Comparison. Once a fault is confirmed, the topological weighted centrality of the physical residuals is calculated, the fault node or region number is output, and the alarm interface is activated. The end-to-end delay from data input to alarm output is controlled within... Within milliseconds, it meets the real-time requirements of wide-area protection in power systems.

[0036] 5. Auxiliary support unit The auxiliary support unit includes: Storage array: using piece A TB NVMe SSD is used to build a RAID 1 ring buffer to... The write rate of MB / s continuously records raw PMU measurements, time-frequency characteristics, reconstruction output, and fault events, and the local storage capacity can store no less than [number missing] MB / s. Full data for the day. Automatically locks the relevant data segment and generates a COMTRADE format waveform file when a fault is triggered.

[0037] Time synchronization unit: Built-in high-stability temperature-controlled crystal oscillator (OCXO, frequency stability) The ppm) and IRIG-B decoding module, combined with GPS / BeiDou external reference sources, provide nanosecond-level time synchronization signals for all modules in the device, ensuring global time tag consistency.

[0038] Management port: A dedicated gigabit Ethernet port, physically isolated from the data plane, used for remote parameter download (including online updating of model weights), operation status monitoring, and operation and maintenance diagnosis.

[0039] like Figure 1 As shown, the present invention also provides a fault monitoring method for a smart grid, applied to the aforementioned fault monitoring device for a smart grid, comprising: Step 1: Synchronously collect measurement data from each node of the power grid and preprocess it to obtain a spatiotemporal data matrix and electrical quantity vectors at corresponding times; Step 2: Based on the spatiotemporal data matrix, extract time-frequency features node by node through adaptive fractional S-transform to obtain the global time-frequency feature vector; Step 3: Input the global time-frequency feature vector and the power grid topology admittance parameters into the physical sensing graph neural differential equation network. Aggregate the potential states of neighboring nodes through the graph attention mechanism and use ordinary differential equations to describe the continuous time evolution of the potential states. The reconstructed time-frequency features, reconstructed voltage amplitude, and reconstructed phase angle are output through the dual-branch decoder. The theoretical injected power is calculated based on the reconstructed voltage amplitude, reconstructed phase angle, and admittance matrix. Step 4: Based on the reconstructed time-frequency characteristics, reconstructed voltage amplitude, reconstructed phase angle, theoretical injected power, and measurement data, calculate the node-level residuals in multiple dimensions, fuse them to obtain the global fault indication score, and compare it with the adaptive threshold to determine the fault.

[0040] The following steps will be explained in detail with specific parameter settings: In step 1, measurement data from each node of the power grid are collected synchronously and preprocessed to obtain a spatiotemporal data matrix and electrical quantity vectors at corresponding times, specifically: This step is performed by the data acquisition and preprocessing module. The data acquisition and preprocessing module uses a fixed sampling interval. To the jurisdiction Each PMU sends a data frame request. This represents the preset total number of critical busbars and line nodes. The upload refresh rate of the PMU determines the frequency coverage required for subsequent time-frequency analysis to meet the transient detection needs of the power grid. Fixed ms, corresponding to sampling frequency Hz. Upon receiving a request, each PMU sends back a synchronization phasor data message containing the synchronization measurement value of that node under the same GPS / BDS timing pulse.

[0041] For nodes ,in In timestamp The measurement vector parsed from the PMU message at any given time is denoted as: ; The meanings and uses of each parameter are as follows: For nodes The fundamental positive sequence voltage amplitude, in per-unit value pu, is calculated by the DFT fundamental extraction algorithm after analog-to-digital conversion from the secondary side of the voltage transformer inside the PMU. It is used for the time-frequency feature extraction in step 2 and the voltage amplitude term in the physical consistency constraint in step 3. For nodes The fundamental positive-sequence voltage phase angle, in radians (rad), is synchronously aligned with a wide-area reference phase angle and calculated after being locked by the PMU's built-in timing module. It is used in the power flow equations of step 3 for calculating the phase angle difference. The structure; For nodes The injected active power, measured in per-unit (pu), is calculated in real-time by the power calculation module from the secondary signals of the current and voltage transformers associated with the node by the PMU. The direction of injection into the node is defined as positive, and this is used to provide the actual measured power value during training in step 3. And step 4, online calculation of physical residuals; For nodes The injected reactive power, measured in per-unit (pu), is calculated in parallel with the active power calculation. The injection node direction is assumed to be positive, and this is used to provide the actual measured power value during training in step 3. Step 4 involves online calculation of the physical residual.

[0042] The data acquisition and preprocessing module processes each timestamp. Received The measurement vectors are concatenated in node number order to form the wide-area measurement snapshot vector at the current moment: ; Subsequently, the data acquisition and preprocessing module uses data of length [missing information]. A sliding time window of each sampling point buffers and frames consecutively arriving snapshot vectors, whereby... The target sample points for each time-frequency transform processing, with a typical value of [value missing]. or Sliding step size Pick sampling points (i.e.) This allows for point-to-point sliding and maximum temporal resolution. After framing, the snapshot vectors within a historical time window are combined into a spatiotemporal data matrix: ; in This indicates the end time of the current sliding window, and is also the time label for outputting the time-frequency features in step 2. Compared with the original sampling time The alignment relationship is (No. The timestamps of each sampling point). The three dimensions of this matrix correspond to spatial nodes ( ), Measurement type ( aisle: , , , ), time sampling points ( ).

[0043] To eliminate high-frequency random impulses in measurement noise and maintain the relative phase relationship between channels, the data acquisition and preprocessing module performs processing on each node after framing. Each measurement channel undergoes zero-phase digital filtering along the time axis. Zero-phase filtering is achieved through forward-backward IIR filtering: first, the length is... The time series is passed forward through a second-order Butterworth low-pass filter (cutoff frequency is taken as...). (Hz, to retain the fundamental and low-order harmonic information of the power grid), then the filtered result is inverted in time and passed through the same filter again. Finally, the output is inverted to obtain the zero-phase filtered sequence. Taking the voltage amplitude channel as an example, the filtered output is denoted as... Since the equivalent amplitude-frequency response of the forward-reverse filter is the square of the amplitude-frequency response of the original filter, and the phase response is always zero, the relative phase relationship of each channel after filtering can be guaranteed to remain unchanged. All measurements after filtering still use the original symbols. , , , This is to maintain consistency in the symbol system.

[0044] The data acquisition and preprocessing module also needs to perform validity checks on the filtered data, specifically including: Amplitude exceeding limit detection: If any node voltage amplitude Exceeding The range of pu is marked as an invalid frame, the frame is discarded and the previous valid frame is reused; Phase jump detection: If any node phase angle The absolute value of the change between adjacent sampling points exceeds a preset threshold. (Pick (rad), then further calculate all at that moment The mean of the phase angle change at each node ;like and The ratio exceeds Furthermore, the deviation between the phase angle change at this node and the network average is greater than [missing value]. If the value is less than rad, it is considered an isolated bad point, and linear interpolation is used to replace the value; if the phase angle of most nodes in the entire network changes synchronously (exceeding rad), it is considered an isolated bad point. The phase angle change at each node exceeds If the value is 0, it is considered a change in the normal operating mode of the system, and the original value is retained.

[0045] After the above acquisition, framing, filtering, and verification processes, the data acquisition and preprocessing module outputs a clean spatiotemporal data matrix. The feature extraction module for the adaptive fractional S-transform is used as input for step 2. Simultaneously, the end time of the sliding window... Corresponding voltage amplitudes at each node (Right now Voltage phase angle Active power and reactive power Extracted separately as time points The measured electrical quantity vectors are used for constructing the training loss in step 3 and calculating the online residuals in step 4. Voltage amplitude The complete time window sequence is then fully incorporated into step 2 as the main analysis signal for time-frequency transformation.

[0046] In step 2, based on the spatiotemporal data matrix, time-frequency features are extracted node-by-node using an adaptive fractional S-transform to obtain the global time-frequency feature vector, specifically as follows: This step is performed by the adaptive fractional S-transform feature extraction module, which receives the clean spatiotemporal data matrix output from the data acquisition and preprocessing module in step 1. The system generates multi-resolution time-frequency features node by node and time window by time window, and finally splices them into a wide-area spatiotemporal feature sequence. This is for use by the continuous-time evolution network in step 3. The following is a detailed explanation of the content involved in step 2: 1. Mathematical definition and discretization calculation of fractional S-transform For any node Extract the time series segment of its voltage amplitude channel from the spatiotemporal data matrix, with a time window length of [missing information]. Each sampling point is aligned to the current time. That is, the processed signal is: ; in This represents the voltage amplitude after zero-phase filtering has been completed. For simplicity, the continuous-time independent variable will be written as follows: And set a signal A sequence of voltage magnitudes representing any node.

[0047] Define the window function of the generalized fractional S-transform as a fractional Gaussian window with frequency dependence and harmonic order parameters: ; in This is a time shift variable (in seconds). To analyze frequency (unit: Hz) This is a fractional-order parameter that controls the frequency modulation intensity of the window function. The characteristics of this window function are as follows: when At that time, it degenerates into a Gaussian window that is a standard S-transform. ; when At that time, the modulation law of frequency to window width is changed by Given. For low-frequency components, The value of is relatively large, and the window function decays faster over time, thus achieving higher time resolution; for high-frequency components, Smaller window size allows for wider window function expansion, resulting in higher frequency resolution. Parameters A decrease in will enhance the overall frequency modulation effect, and vice versa.

[0048] Signal The fractional S-transform expression is: ; in For the center time of the window function, The imaginary unit, This is the complex time spectrum.

[0049] In the discrete implementation, continuous integration is transformed into a fast frequency domain algorithm. Let the discrete signal... Length is Sampling interval s. First calculate Discrete Fourier Transform : ; The corresponding discrete frequency is Hz. For each positive frequency point Construct a frequency domain window function: ; in It is a translation variable in the frequency domain. Then, the discrete time is obtained by multiplying in the frequency domain and performing an inverse Fourier transform. Complex time spectrum at: ; in This represents the modulo operation to achieve cyclic shifting. This module selects... Analysis frequency points of logarithmic intervals The frequency range is to cover the main frequency bands of power grid disturbances. Hz to Hz (below the Nyquist frequency) (Hz), covering oscillations near the fundamental frequency to high-frequency transient components. Pick Discard the DC component ( ) and Nyquist frequency point ( ).

[0050] For each frequency point All time points The magnitudes of the complex time spectrum constitute the time-frequency amplitude spectrum: ; in Represented by voltage amplitude signal As input The calculated complex time spectrum.

[0051] 2. Fractional order parameters Adaptive online adjustment network The traditional S-transform has a fixed window function shape, which cannot adapt to the non-stationary characteristics of fault signals. This invention introduces a lightweight feedforward neural network as a fractional-order parameter prediction network. Its function is to: based on the local kurtosis of the signal within the current sliding window and the previous moment... Value, dynamically predicting the optimal value at the current moment. The parameters enable the window to automatically shrink during the impulsive transient phase to improve time resolution, and to widen during the steady oscillation phase to improve frequency resolution.

[0052] The basis for choosing kurtosis as a statistical prior is that: when the power grid is operating normally, the voltage waveform is close to sinusoidal, and the kurtosis is close to... When a fault transient occurs, the waveform contains obvious spikes or oscillations, and the kurtosis increases sharply. Kurtosis is defined as: ; in For the current node in the window The average voltage amplitude.

[0053] Prediction Networks The structure is defined as follows: The input layer dimension is The two scalar inputs are: the kurtosis of the current window. and the fractional-order parameters of the previous time step (System initial moment) Set as ); Hidden layer by It consists of 10 neurons, and the activation function is the Swish function. ,in For the Sigmoid function; The output layer is There are 10 neurons, with the sigmoid function as the activation function, which maps the output value to 10 ... The interval, after output bias After lower bound protection, the parameter is assigned as a candidate fractional parameter. .

[0054] The overall calculation expression is: ; in This is the weight matrix from the input layer to the hidden layer. This is the hidden layer bias vector. This is the weight matrix from the hidden layer to the output layer. This is the output layer bias scalar. This represents the vector concatenation operation. All of the above learnable parameters... Included in parameter set .

[0055] To suppress high-frequency parameter jitter, a first-order low-pass smoothing is applied to the network output: ; in Smoothing factor (empirical value, range of values) Smoothed Directly acting on the frequency domain window function Used to calculate the current time The time spectrum below.

[0056] Network parameters It is not trained independently, but is used as the overall loss function in step 3. Part of this is uniformly optimized using stochastic gradient descent. The gradient flow is transmitted backward from the loss in step 3 through the time-spectrum calculation link and the network output to... This end-to-end collaborative learning enables The adjustment direction directly serves the ultimate goal of fault detection, rather than relying on manual criteria.

[0057] 3. Construction of global time-frequency feature sequences Get the current time node Time-frequency amplitude spectrum Then, select the time index corresponding to the end time point of the window. (Right now The sampling positions corresponding to each time point are arranged along the frequency axis to obtain a... A dimensional vector, called the node time-frequency feature: ; For the entire network Each node performs the above operation, and the time-frequency feature vectors of all nodes are concatenated vertically to obtain the global multi-channel time-frequency features: ; make ,but This feature vector is expressed at frame rate. The Hz output is continuous, forming a continuous feature stream, which serves as the input driving signal for the neural differential equation in step 3.

[0058] 4. Engineering implementation details and workflow of the feature extraction module The adaptive fractional S-transform feature extraction module contains a frequency window coefficient buffer. After the update, pre-calculation Group frequency window The data is then stored in the GPU texture memory, and then the time-frequency spectrum of all nodes is calculated concurrently using batch matrix multiplication. , , In typical scenarios, the time consumed for single-frame feature extraction is controlled within Within milliseconds, it meets real-time requirements.

[0059] Phase angle of the voltage output in step 1 Instead of performing time-frequency conversion in this module, it is directly bypassed and passed to step 3 as one of the monitoring signals for electrical quantity reconfiguration.

[0060] The output of this module The data is sent to the continuous-time evolution network inference module via a high-speed PCIe bus, and the module will simultaneously... Historical values ​​and current window kurtosis It is kept in local memory for use by the prediction network in the next frame.

[0061] In step 3, the global time-frequency feature vector and the grid topology admittance parameters are input into the physical sensing graph neural differential equation network. The latent states of neighboring nodes are aggregated through a graph attention mechanism, and the continuous-time evolution of the latent states is described using ordinary differential equations. The reconstructed time-frequency features, reconstructed voltage amplitude, and reconstructed phase angle are output through a dual-branch decoder. Based on the reconstructed voltage amplitude, reconstructed phase angle, and admittance matrix, the theoretical injected power is calculated. Specifically: This step is executed by the continuous-time evolution network inference module. This invention proposes a Physics-Aware Graph Neural ODE, which organically integrates power grid topology, electromechanical transient physical equations, and continuous-time deep learning. Its core lies in: embedding the node admittance matrix as an edge feature into the vector field function of the graph attention neural ODE, so that the continuous evolution of the latent state is simultaneously constrained by data-driven processes and prior physical structure; simultaneously, introducing AC power flow equations as explicit physical constraints at the decoding end, ensuring that the reconstructed electrical quantities strictly obey the power grid power balance law; ultimately achieving end-to-end collaborative optimization of adaptive time-frequency transformation, graph dynamics evolution, and physical constraints. The following section will provide a detailed explanation of the content involved in step 3: 1. Multi-scale time-series graph encoder Step 2 outputs the global feature vector Dimensions This module first reconstructs it into a node-level feature matrix: ; That is, each row corresponds to one node. Dimensional time-frequency characteristics.

[0062] To capture dynamic patterns at different time scales, a multi-scale temporal convolutional encoder is designed. , node feature matrix Mapped to node-level initial potential state The encoder consists of three parallel dilated convolution branches with dilation rates of [missing information]. , , Each branch contains two layers of one-dimensional causal convolutions (with a kernel size of [missing information]). The number of output channels is The outputs of each branch are spliced ​​together and then linearly projected to obtain the result. The latent state is defined. The encoder operates independently on the feature sequence of each node (along the time axis) in a node-shared manner, and is only used at system startup. Potential state initialization: ; in The number of context frames required by the encoder (corresponding to) (ms history), all nodes share the same encoder parameters. Latent state dimension .

[0063] Meanwhile, the power grid topology is determined by the node admittance matrix. Given, among which For nodes and Mutual conductance between them Mutual susceptance ( (Non-zero elements indicate the presence of transmission lines). For nodes Self-conductance (earth admittance). For nodes The self-susceptance. Based on this, construct the standard edge eigenvectors: for And there is a line connection: For self-loop : 2. Node-level continuous-time latent state dynamics For each node Introduction 3D potential state vector The potential states of all nodes are denoted as a matrix. .node The latent state evolution is described by ordinary differential equations: ; The meanings of each item are as follows: The parameter set is The neural vector field function outputs Rate of change of state in dimensional state; The message vector is aggregated from neighboring nodes via a graph attention mechanism. ; The external driving context vector is generated by the current time node. Time-frequency characteristics Obtained through linear transformation, : ; in For context projection matrix, This is the context bias vector.

[0064] Message aggregation employs an edge-feature-weighted graph attention mechanism to explicitly inject physical admittance information into the information transmission process. For node pairs connected by edges... (including self-loops) ), define the attention coefficient: ; ; in For learnable attention vectors, The node state projection matrix, The edge feature projection matrix, This represents a vector concatenation operation. For nodes The neighbor set (including self-loops) itself), negative slope Edge features The introduction of this feature allows the model to adaptively adjust information weights based on the strength of the line admittance during aggregation. Information transmitted by high admittance lines (strong electrical coupling) is given higher attention, while information transmitted by low admittance lines (weak coupling) is suppressed, resulting in a clear physical meaning.

[0065] Aggregated messages The sum of attention-weighted feature transformations of neighbor states: ; in The projection matrix is ​​a value.

[0066] Vector field function The complete input is a concatenated vector. (dimension) Internally, it is implemented as a three-layer fully connected network with a Swish activation function: First layer: Swish activation; Second layer: Swish activation; Third layer: (Right now No activation (output change rate).

[0067] All learnable parameters Set 3. Numerical solution of continuous-time state trajectory Given an initial potential state and intervals Input trajectory within The result is obtained by integration using a differentiable numerical ODE solver. : ; The solver uses the standard fourth-order Runge-Kutta method (RK4) with a fixed step size. ms. During online inference, each time a new frame is received... In the interval Internal hypothesis Keep constant as (i.e., zero-order preservation), from integral One RK4 substep to get .

[0068] The initialization strategy is as described above: when the system starts up (or is reset online), the feature maps are obtained from the initial few frames using a multi-scale temporal convolutional encoder.

[0069] 4. Dual-branch decoder and multi-target reconstruction To simultaneously satisfy both feature reconstruction and physical consistency constraints, the decoder starts from the latent state. Output two branches: Time-frequency feature decoding branch: reconstruction node Time-frequency eigenvectors It is implemented by two layers of MLP: ; in , , , .

[0070] Electrical quantity decoding branch: Reconstruction node voltage amplitude With phase angle It is implemented by two layers of MLP, where the voltage amplitude output is processed by the Softplus function. Ensure positive values: ; in , , , , . Output range is In subsequent power flow calculations, the modulus will be taken to... The interval conforms to the physical definition.

[0071] The global reconstructed features are obtained by vertically concatenating the time-frequency feature reconstruction results of all nodes: Decoder parameters , With encoder parameters Vector field parameters and the fractional-order prediction network parameters in step 2 The parameter space that constitutes the entire end-to-end joint training is denoted as... .

[0072] 5. Loss Function and End-to-End Training The training data consists of continuous time series data from the power grid's historical normal operating conditions, and does not include any fault samples. For each time step... Obtained from steps 1 and 2 , , , , The training objective is to minimize the following composite loss function: Time-frequency feature reconstruction loss: ; in This represents the total number of time steps in a training batch.

[0073] Electrical quantity reconfiguration loss is used to provide a direct monitoring signal to stabilize physical quantity estimation. ; The phase loss is implemented using a dual-component form of sine and cosine to avoid phase angle fluctuations. The instability of training caused by the discontinuity in the vicinity.

[0074] Physical consistency loss: Based on the reconstructed voltage magnitude, phase angle, and admittance matrix, the node injected power is calculated using the AC power flow equations and compared with the measured power. ; ; in For nodes and The reconstructed phase angle difference between them. The physical consistency loss is defined as: ; Latent State Smoothing Regularization Loss: To encourage the smoothness of the latent state trajectory, an L2 penalty term is added to the rate of state change. ; The total loss is the weighted sum: ; in For the weight decay regularization term, the empirical values ​​for each balance coefficient are: (Electrical quantity reconfiguration weights). (Physical consistency weight, with the highest value to highlight the core position of physical constraints). (Smoothing regularization weights) (Weight decay coefficient).

[0075] 6. Gradient Flow and End-to-End Optimization The entire computational chain, from the fractional S-transform in step 2 (based on FFT and complex number operations, fully differentiable), to the fractional prediction network... Multi-scale temporal convolutional encoders, graph attention aggregation, ODE numerical solution (RK4 differentiable), dual-branch decoders, and loss calculation can all be implemented using automatic differentiation frameworks (such as PyTorch). Total loss For all parameters The gradient is backpropagated through the numerical ODE solver using the adjoint method. The specific optimization algorithm used is AdamW, with an initial learning rate of... Weight decay coefficient Cosine annealing learning rate scheduling is adopted (period is 1). (epochs), batch size is A series of consecutive time segments (each segment contains) (time step), training iterations approximately epochs until the validation set loss converges.

[0076] After training is complete, only the encoder and ODE vector field function are retained. The dual-branch decoder and the fixed forward inference weights of the fractional-order prediction network in step 2 form the online inference model. During online inference, according to... Step receiver Obtained by encoder (Calculated only initially, followed by continuous integration using the RK4 solver), Output , , and the calculations based thereon , No backpropagation of gradients is required throughout the process, and the inference time per frame is controlled within [a certain timeframe]. Within ms.

[0077] In step 4, based on the reconstructed time-frequency characteristics, reconstructed voltage amplitude, reconstructed phase angle, theoretical injected power, and measurement data, node-level residuals in multiple dimensions are calculated, fused to obtain a global fault indication score, and compared with an adaptive threshold to determine the fault. Specifically: This step is executed by the fault detection and location module, deployed in the embedded multi-core CPU processor of the online monitoring device. It receives the online inference output from the continuous-time evolution network in step 3 and the synchronous measurement data from step 1, calculates the fault indication score in real time, and determines the fault location. The module internally includes: a residual calculation unit, an EWMA statistical unit, a fusion decision unit, and a graph structure location analysis unit. Step 4 is described in detail below: 1. Online calculation of multidimensional reconstruction residuals and physical residuals At any moment The module has obtained the following data: Step 1 measurement values: (node (Voltage amplitude measurement value) (Voltage phase angle measurement value) (Active power measurement value) (Reactive power measurement value) ; Step 2 Extracted Time-Frequency Feature Vector ; Step 3: Reconstructed time-frequency features from inference output Reconstructed voltage amplitude Reconstructing phase angle And the theoretical injected power calculated from the power flow equations based on the reconfigured electrical quantities. , .

[0078] First, define the node-level residuals in three dimensions: Time-frequency feature reconstruction residual: for nodes Take its reconstructed time-frequency vector With the true time-frequency vector (from Normalized Euclidean distance (obtained by splitting by node index): ; in The L2 norm of a vector. This represents the number of frequency sampling points.

[0079] Electrical quantity reconstruction residuals: Normalized weighted deviation of voltage amplitude and phase angle. in and They are nodes The standard deviation of the time series of voltage amplitude and phase angle during historical normal operation (training dataset) is used for dimensionless normalization; and The measured phase angle and the reconstructed phase angle represent the distances from the time of measurement and the time of reconstruction, respectively. The values ​​are wrapped within the interval and aligned according to the principle of minimum angle difference to avoid residual anomalies caused by phase angle periodicity.

[0080] Physical consistency residuals: The normalized deviation of injected power directly reflects the degree of violation of the AC power flow equations. ; in , They are nodes The time series standard deviation of active and reactive power during normal operation. This residual is particularly sensitive during faults, as faults disrupt the normal power flow distribution, leading to a significant deviation between measured power and theoretical power calculated based on reconfigured electrical quantities.

[0081] 2. Construction of Global Fault Indication Score To suppress sporadic measurement noise and integrate anomaly information from different dimensions, a local anomaly index is first calculated for each node: ; in (Time-frequency feature residual weights) (Electrical quantity residual weight) (Physical consistency residual weight) is a weighting coefficient that satisfies Typical values ​​are as follows , , This highlights the core role of physical consistency residuals in fault identification.

[0082] Then, the maximum local anomaly index of all nodes is taken as the global fault indication score of the system: ; Using the maximum value strategy can ensure that the fault signal is not diluted by a large number of normal nodes, and has higher detection sensitivity for local faults.

[0083] 3. Adaptive dynamic threshold and fault diagnosis To avoid frequent false alarms due to changes in operating conditions caused by fixed thresholds, an adaptive threshold mechanism based on Exponentially Weighted Moving Average (EWMA) is adopted. Statistical analysis is performed on an offline validation set containing only normal operating conditions. mean and standard deviation When running online, offline statistics are used as the initial values ​​for EWMA: , , .

[0084] The online frame-by-frame updates are as follows: ; ; ; in It is a fractional smoothing factor. The mean learning rate, Let be the variance learning rate. The adaptive threshold at the current time step is: ; in The sensitivity coefficient is taken as... (Corresponding to the approximation) (Confidence interval, with an extremely low false alarm rate under the normality assumption).

[0085] when And this condition is continuous 1 frame (corresponding to) If the condition is met continuously within ms, a power grid fault is determined, and a detection flag is output. And latch the fault detection time. The actual start time of the fault was traced back to... (i.e., the moment when the first frame exceeds the threshold).

[0086] To prevent threshold values ​​from being contaminated by outliers during a fault, from Start pause and The online updates maintain the statistical values ​​from the last frame in a normal state before the failure.

[0087] The condition for clearing the fault is: under the fault state, when continuous The frame fell back to If the threshold latched before the fault occurs is below the threshold, the fault is considered cleared, EWMA update is resumed, and output is performed. .

[0088] 4. Fault localization based on physical residual gradient and topology weighting once The location unit is immediately activated, utilizing the spatial distribution characteristics of physical consistency residuals and the topological connection relationship of the power grid to achieve refined fault location.

[0089] First, starting from the moment the fault begins... Begin by extracting the node physical residuals of each frame. Define nodes Deadline during the fault Cumulative physical residuals: ; in To determine the length of the positioning window, take... Frame (corresponding) (ms) to achieve a balance between detection speed and cumulative signal-to-noise ratio.

[0090] Due to the connectivity of the power grid, the physical residuals of adjacent nodes at the fault point also increase due to power flow redistribution, but the residual amplitude of the fault source node is the largest and its rising edge is the earliest. To further suppress the diffusion effect caused by the coupling between adjacent nodes, a topology-weighted fault centrality index is proposed: For each node Construct its fault centrality : ; The meanings of each parameter are as follows: For the line The admittance magnitude characterizes the electrical coupling strength between nodes; Represents a node The neighbor set (excluding self-loops), if node If a node is isolated (has no neighbors), then the summation term is: ; To prevent small constants with a denominator of zero; Contribute attenuation coefficient to neighbors, value This is used to suppress pure diffusion effects and ensure that the centrality of the true fault source is significantly higher than that of the nodes that are only affected.

[0091] The multiplication structure of this formula ensures that the node itself accumulates residuals. The dominant term is the neighbor residual, which is weighted by admittance and used only as an auxiliary correction term (multiplied by). (by suppressing it), thus realizing the physical rational positioning logic of prioritizing its own residuals and assisting with neighbor information.

[0092] End the positioning window (i.e.) At any given time, the node with the highest centrality is marked as the faulty node. ; Output its number and the corresponding bus name. If the highest centrality... With the second highest centrality The ratio is less than (Right now This indicates that the fault source is not unique or the diffusion effect is too strong, so the location range should be expanded to the centrality range. Nodes ( Take the satisfaction The system constructs a connected subgraph (with the smallest integer value of the system) and issues a regional fault alarm, indicating that there may be a bus fault or multiple anomalies caused by protection actions.

[0093] 5. Output and Connection The module will display fault detection flags. Fault indication score Adaptive threshold The fault node number or area information is packaged into real-time data packets and sent to the SCADA alarm system and maintenance workstation interface via Gigabit Ethernet in IEC61850-8-1 MMS message format, while also being stored in the local historical event logger. If The system automatically triggers the fault recording function and saves the data. front and back sampling points (i.e.) s, with The original PMU measurement waveforms, time-frequency characteristics, and reconstructed output data are obtained for post-event analysis.

[0094] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0098] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A fault monitoring method for a smart grid, characterized in that, include: Step 1: Synchronously collect measurement data from each node of the power grid and preprocess it to obtain a spatiotemporal data matrix and electrical quantity vectors at corresponding times; Step 2: Based on the spatiotemporal data matrix, extract time-frequency features node by node through adaptive fractional S-transform to obtain the global time-frequency feature vector; Step 3: Input the global time-frequency feature vector and the power grid topology admittance parameters into the physical sensing graph neural differential equation network. Aggregate the potential states of neighboring nodes through the graph attention mechanism and use ordinary differential equations to describe the continuous time evolution of the potential states. The reconstructed time-frequency features, reconstructed voltage amplitude, and reconstructed phase angle are output through the dual-branch decoder. The theoretical injected power is calculated based on the reconstructed voltage amplitude, reconstructed phase angle, and admittance matrix. Step 4: Based on the reconstructed time-frequency characteristics, reconstructed voltage amplitude, reconstructed phase angle, theoretical injected power, and measurement data, calculate the node-level residuals in multiple dimensions, fuse them to obtain the global fault indication score, and compare it with the adaptive threshold to determine the fault.

2. The fault monitoring method for a smart grid according to claim 1, characterized in that, In step 1, measurement data from each node of the power grid are collected synchronously and preprocessed to obtain a spatiotemporal data matrix and electrical quantity vectors at corresponding times, specifically: Measurement vectors containing voltage amplitude, voltage phase angle, active power and reactive power are obtained from the synchronous phasor measurement devices at each node of the power grid at a fixed sampling frequency. Zero-phase filtering is performed on the measurement vectors within each time window, and the validity of the filtered data is verified. The verified data is organized into a spatiotemporal data matrix, and the measurement values ​​at the end of the time window are extracted as electrical quantity vectors.

3. The fault monitoring method for a smart grid according to claim 2, characterized in that, The validity verification includes amplitude over-limit detection and phase angle jump detection.

4. The fault monitoring method for a smart grid according to claim 3, characterized in that, In step 2, the adaptive fractional S-transform uses a fractional Gaussian window function, and the fractional parameters of the adaptive fractional S-transform are adjusted online by a differentiable fractional parameter prediction network based on the local statistical characteristics of the signal.

5. The fault monitoring method for a smart grid according to claim 4, characterized in that, In step 3, the physical perception graph neural differential equation network includes: A multi-scale temporal convolutional encoder is used to generate the initial latent state of a node based on the global time-frequency feature vectors of multiple frames; The graph attention message passing module uses the real and imaginary parts of the node admittance as edge features to calculate the attention weights between neighboring nodes and aggregate messages. The vector field function takes the latent state of a node, the aggregated message, and the time-frequency feature context as input, and outputs the rate of change of the latent state. Numerical ordinary differential equation solver that performs continuous-time integration of the latent state based on the vector field function; The dual-branch decoder decodes the reconstructed time-frequency characteristics, reconstructed voltage amplitude, and reconstructed phase angle from the latent state, respectively.

6. The fault monitoring method for a smart grid according to claim 5, characterized in that, The training loss function of the physical perception graph neural differential equation network includes: time-frequency feature reconstruction loss, electrical quantity reconstruction loss, physical consistency loss, latent state smoothing regularization loss, and weight decay term; wherein, the physical consistency loss is constructed based on the deviation between the theoretical injected power and the measured power calculated by the reconstructed voltage amplitude and phase angle according to the AC power flow equation.

7. A fault monitoring method for a smart grid according to claim 6, characterized in that, In step 4, the node-level residuals in multiple dimensions include: time-frequency feature reconstruction residuals, electrical quantity reconstruction residuals, and physical consistency residuals; the physical consistency residuals are calculated based on the normalized deviation between the measured active power, reactive power, and theoretical injected power. The global fault indication score is the maximum value of the local anomaly index of each node, and the local anomaly index is the weighted sum of the three-dimensional residuals; The adaptive threshold is dynamically updated based on the mean and variance of the exponentially weighted moving average of the global fault indication score.

8. The fault monitoring method for a smart grid according to claim 7, characterized in that, Step 4 also includes: After identifying the fault, the location of the fault is determined by the spatial distribution of the physical consistency residuals and the grid topology connection relationship.

9. A fault monitoring device for a smart grid, applied to a fault monitoring method for a smart grid as described in any one of claims 1-8, characterized in that, include: The data acquisition and preprocessing module is used to synchronously acquire measurement data from the synchronous phasor measurement devices at each node, perform zero-phase filtering and validity verification, and output spatiotemporal data matrix and electrical quantity vector. An adaptive fractional S-transform feature extraction module, connected to the data acquisition and preprocessing module, is used to perform the adaptive fractional S-transform online and generate a global time-frequency feature vector. The physical perception graph neural differential equation inference module is connected to the feature extraction module and is used to receive the global time-frequency feature vector and use the physical perception graph neural differential equation network to output the reconstructed time-frequency features, reconstructed voltage amplitude, reconstructed phase angle and theoretical injected power. The residual fault detection and location module is integrated and connected to the inference module and the data acquisition and preprocessing module. It is used to calculate multi-dimensional node-level residuals, generate global fault indication scores, perform adaptive threshold comparisons to realize fault determination, and perform fault location based on physical consistency residuals and power grid topology when a fault occurs. The auxiliary support unit includes at least a storage array, a time synchronization unit, and a management interface.

Images