Power distribution network grounding fault accurate positioning method based on traveling wave detection and artificial intelligence

Abstract

The invention relates to the field of power distribution network ground fault accurate judgment, and discloses a power distribution network ground fault accurate positioning method based on traveling wave detection and artificial intelligence, which comprises the following steps: (a) collecting a power distribution network traveling wave voltage signal through a multi-node synchronous sensor array, and generating a preprocessing signal through baseline correction and wavelet threshold denoising; (b) constructing a Gaussian random field model of line resistance, inductance and capacitance, and simulating a traveling wave propagation process in combination with a nonlinear wave equation; and (c) defining Finsler measurement on the traveling wave signal manifold, and separating the aliasing mode by minimizing path integration. According to the invention, the technical scheme of multi-node synchronous signal acquisition and self-adaptive preprocessing is adopted, so that the effect of eliminating coupling of power frequency interference and high-frequency noise is achieved. Compared with a scheme depending on a single sensor and a fixed filtering threshold in the prior art, the method solves the defect that the wave head time difference extraction is not accurate due to asynchronous sampling.

Description

Technical Field

[0001] The present invention relates to the technical field of accurate judgment of grounding faults in distribution networks, and in particular to a method for accurately locating grounding faults in distribution networks based on traveling wave detection and artificial intelligence. Background Art

[0002] The field of ground fault location in distribution networks primarily utilizes traveling wave detection, impedance analysis, and artificial intelligence (AI) technologies. Traveling wave detection relies on the propagation characteristics of transient signals generated by faults and utilizes dual- or multi-terminal synchronous acquisition to achieve preliminary location. Artificial intelligence (AI) technology incorporates machine learning models to process the nonlinear relationship between signal characteristics and line parameters, improving location accuracy. Existing methods involve key technologies such as wavelet transform extraction of waveheads, neural network pattern recognition, and distributed traveling wave sensor networking. Related algorithms include support vector machines and the application of convolutional neural networks for noise suppression and feature extraction.

[0003] Traveling wave signals are susceptible to interference from the superposition of branch line reflections in complex distribution network topologies, causing traditional threshold trigger mechanisms to misjudge the arrival time of the wave head; Gaussian models with fixed parameters find it difficult to quantify the spatial heterogeneity of propagation distortion caused by line aging, resulting in the accumulation of deviations between simulated waveforms and real signals; most artificial intelligence algorithms rely on static data sets trained offline and are unable to adapt to parameter drift under dynamic grid conditions in real time; existing quantum mapping methods have not established a dynamic calibration mechanism for potential field strength and wave head amplitude, and weak fault signals are easily submerged by noise; there is a lack of adaptive elimination methods for the phase interference between the injected signal and the reflected wave in closed-loop control, reducing positioning reliability in strong interference environments. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method for accurately locating grounding faults in distribution networks based on traveling wave detection and artificial intelligence, which solves the problems in the existing technology such as traveling wave signal reflection interference leading to wave head misjudgment, line aging distortion causing simulation deviation, static model unable to adapt to dynamic parameter drift, weak fault signal quantum mapping failure, and phase interference reducing positioning reliability.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for accurately locating ground faults in a distribution network based on traveling wave detection and artificial intelligence, comprising the following steps: (a) The traveling wave voltage signal of the distribution network is collected by a multi-node synchronous sensor array, and the preprocessed signal is generated through baseline correction and wavelet threshold denoising; A distributed sensor array synchronously captures the three-phase busbar traveling wave voltage signal, then combines power frequency baseline subtraction with adaptive wavelet denoising to achieve signal purification. The precise synchronization ensures spatiotemporal consistency for subsequent modeling. A dynamic noise suppression mechanism automatically adjusts the filter threshold based on the signal's spectral characteristics, avoiding the wavefront distortion caused by traditional fixed thresholds. The preprocessed signal is transmitted to the central processing unit via a high-speed bus, providing unbiased input for physics-driven modeling.

[0006] Synchronous acquisition eliminates phase differences in multi-node signals, ensuring accurate measurement of the arrival time of traveling waves. Baseline subtraction removes power frequency components, preventing low-frequency oscillations from masking the characteristics of the traveling wave's head. Wavelet denoising jointly suppresses random noise in the time and frequency domains, preserving the integrity of the high-frequency components of the head. Each module forms a cascaded processing chain, with the output of the previous stage serving as the input of the next stage, improving overall signal quality.

[0007] (b) Construct a Gaussian random field model of line resistance, inductance, and capacitance, and combine it with nonlinear wave equations to simulate the traveling wave propagation process; Line parameters are modeled as spatially correlated Gaussian random fields, and a wave equation is constructed incorporating nonlinear self-focusing effects. This model uses a covariance kernel function to quantify the impact of heterogeneous characteristics such as line aging and loose joints on traveling wave propagation. Current-dependent nonlinear terms are also introduced to characterize the dynamic effects of corona discharge. The simulation output includes the spatiotemporal evolution of multipath reflection waveforms, providing a foundation for modal separation.

[0008] The Gaussian random field reflects the uncertainty of parameter spatial distribution, capturing the degree of line aging through the covariance correlation length. The nonlinear term simulates the self-focusing of energy in traveling wave propagation, compensating for the high-frequency attenuation error of traditional linear models. This modeling approach integrates physical mechanisms with statistical properties to generate a traveling wave propagation model that is closer to the real power grid.

[0009] (c) Finsler metric is defined on the traveling wave signal manifold to separate aliased modes by minimizing the path integral; A Finsler metric space is constructed on the traveling wave signal manifold, and the primary modes are extracted by minimizing the path integral. This metric integrates the signal's time-frequency characteristics with the manifold's geometric properties, employing a parallel path search strategy to find the optimal solution in parameter space. Momentum gradient descent accelerates convergence while avoiding local minima. The optimized modal data preserves the spatiotemporal signature of the fault reflection head.

[0010] The Finsler metric maps signal features into a distance measure in geometric space. Path integral optimization essentially seeks the propagation path with the lowest energy on the manifold. The momentum gradient incorporates historical update information to suppress parameter oscillations. This process achieves a synergistic effect between physical constraints and data-driven approaches, effectively separating true fault reflections from aliased modes.

[0011] (d) Mapping the distribution network into a quantum spin chain, calculating the local entanglement entropy based on virtual time evolution and tensor network inversion, and locating the fault point; The distribution network topology is mapped to a spin chain model, and node impedance parameters are converted to interspin coupling strengths, significantly enhancing the potential energy field at the fault point. The ground state is searched through imaginary time evolution, and the entanglement entropy of the reduced density matrix at each node is calculated. The statistical anomaly of the entanglement entropy caused by the sudden change in the local potential field at the fault point enables sub-meter location based on threshold comparison.

[0012] The spin chain model transforms the physical structure of the power grid into the interactions of a quantum system. The enhanced potential energy at the fault point disrupts system symmetry, leading to a sudden change in the local entanglement entropy. The virtual time evolution rapidly converges to the lowest energy state through non-unitary transformations, and the tensor network compresses and stores quantum state information, efficiently extracting fault characteristics. This mapping mechanism deeply links quantum many-body effects with power grid fault characteristics.

[0013] (e) Dynamically adjust the amplitude of the applied signal according to the mutation characteristics of the entanglement entropy to optimize the detection sensitivity; Wherein, the Hamiltonian of the quantum spin chain in step (d) is: Where, is the inductance parameter, is the capacitance parameter, is the potential energy intensity of the fault point, 、 、 is the spin operator; Based on the fault location results, the amplitude-frequency characteristics of the applied excitation signal are dynamically adjusted to form a quantum-classical hybrid control loop. The injected signal energy is focused on the fault zone, and a feedback mechanism suppresses noise interference. The protection circuit monitors the line voltage in real time and shuts off signal output in abnormal conditions, ensuring system safety.

[0014] Closed-loop control feeds detection results back to the signal source in real time, forming an adaptive regulation loop. Energy focusing of the applied signal enhances the signal-to-noise ratio of the fault reflection wave, while the feedback suppression module blocks the propagation of signal distortion. This mechanism achieves a dynamic balance between detection sensitivity and system stability.

[0015] Preferably, the baseline correction eliminates the power frequency component by sliding window integration, and the window width is an integer multiple of the power frequency period.

[0016] By setting an integration window that strictly matches the power frequency period, a sliding average calculation is performed on the original traveling wave signal to dynamically extract the time-varying baseline of the power frequency component. Designing the window width to be an integer multiple of the power frequency period ensures that the integration interval fully covers the power frequency fluctuation period, avoiding residual oscillations caused by truncation. After correction, the signal is passed through a high-speed differential circuit to remove the baseline component, preserving the high-frequency characteristics of the traveling wave.

[0017] The power frequency component exhibits periodic variation in the time domain. The sliding integration window captures its complete fluctuation pattern through a time synchronization mechanism. The time domain mean of the integration result is equivalent to the instantaneous amplitude of the power frequency baseline. Low-frequency components are eliminated through subtraction operations to avoid aliasing distortion caused by their overlap with the high-frequency components of the traveling wave head in the frequency domain. This correction method overcomes the baseline residual problem of traditional fixed windows in frequency offset scenarios through adaptive matching of time windows. Data is transmitted between the preprocessing module and the subsequent modeling module through a standardized interface to ensure that the signal time domain alignment accuracy meets the modeling requirements.

[0018] Preferably, the covariance kernel function of the Gaussian random field is determined by maximum likelihood estimation of historical data, and the correlation length is set to 5 meters or 10 meters according to the degree of line aging.

[0019] By analyzing the statistical distribution characteristics of line parameters in historical fault data, a maximum likelihood estimation algorithm was used to fit the key parameters of the covariance kernel function. For aging lines with more than 10 years of service, insulation degradation causes parameter fluctuations to exhibit short-range correlations, with a correlation length of 5 meters. For newly laid lines, due to the high material uniformity, the correlation length was set to 10 meters. This parameter setting strategy is implemented through interaction between the data management module and the modeling module. The historical database updates line service life information in real time to dynamically adjust the kernel function parameters.

[0020] The covariance kernel function reflects the spatial correlation characteristics of line parameters. Short correlation lengths indicate frequent parameter changes in aging lines, while long correlation lengths reflect continuous and smooth parameters in new lines. Maximum likelihood estimation maximizes the joint probability density of historical data to infer the kernel function form that best matches actual operating conditions. The correlation between correlation length and line aging stems from the increased discretization of local parameters caused by insulation degradation. The modeling module automatically matches the preset correlation length parameters by receiving aging level labels from the data management module, ensuring that the spatial heterogeneity of the random field model is consistent with the actual grid state. Kernel function configuration parameters are transmitted between the parameter estimation module and the modeling module via a standardized protocol, supporting high-fidelity calculations for subsequent nonlinear fluctuation simulations.

[0021] Preferably, the nonlinear wave equation includes a nonlinear self-focusing coefficient dynamically calculated from the line current.

[0022] During the traveling wave propagation modeling process, the nonlinear self-focusing coefficient is dynamically adjusted by real-time monitoring of line current changes, reflecting the effect of nonlinear phenomena such as corona discharge and arcing on the energy concentration of traveling waves. Current parameters are collected in real time by a sensor array and input into the parameter calculation module. The self-focusing coefficient is automatically updated based on the rate of change of current amplitude, replacing the simplified assumption of fixed nonlinear terms in traditional models. This dynamic calculation process interacts with the wave equation solver in real time through a parameter interface, ensuring that the simulation model adapts to changing grid conditions.

[0023] The nonlinear self-focusing effect originates from the interaction between the electromagnetic field and the conductor medium during the propagation of traveling waves. When the instantaneous value of the current increases, the degree of ionization on the conductor surface intensifies, resulting in the enhanced concentration of energy toward the wavefront. The dynamic calculation mechanism uses the mapping relationship between current parameters and nonlinear coefficients to correct the nonlinear response intensity in the model in real time, so that the simulated waveform is more in line with the distortion characteristics of the actual traveling wave. The parameter calculation module obtains real-time current data from the data bus, and after normalization, it is input into the nonlinear coefficient generation unit. The generated dynamic coefficients are synchronously updated to the wave equation solving module through the high-speed cache area, forming a closed-loop parameter correction link. This dynamic coupling mechanism breaks through the limitation of traditional static models that cannot characterize transient nonlinear effects, and significantly improves the accuracy of wave head amplitude prediction under complex working conditions.

[0024] Preferably, the path integral optimization of the Finsler metric adopts a momentum gradient descent method to search at least 100 paths in parallel and screen the optimal solution.

[0025] A multidimensional optimization space is constructed on the traveling wave signal manifold. Using a momentum gradient descent algorithm, hundreds of search paths are simultaneously initiated. Each path carries historical gradient updates to accelerate convergence. The initial parameters of each path are randomly distributed within the feasible domain. Local optimal solutions are shared between computational nodes, and invalid paths that deviate from the main mode are dynamically eliminated during the iteration process. Ultimately, an energy function is used to evaluate the globally optimal path, whose corresponding separated mode retains the complete wavefront delay and amplitude characteristics.

[0026] The momentum gradient introduces an inertial effect to suppress parameter oscillations, allowing the search process to rapidly traverse local extremes on complex manifolds. Parallel search expands the exploration range through distributed computing, increasing the probability of capturing true modes. A path screening mechanism, based on the minimum characteristics of the energy functional and combined with parameter similarity between adjacent paths, performs cluster analysis to ensure that the optimal solution satisfies both physical constraints and data matching criteria. Path state information is exchanged between the optimization module and the signal processing unit via a high-speed data bus. A collaborative filtering mechanism between parallel computing nodes reduces redundant computation, resulting in an efficient multimodal separation architecture.

[0027] Preferably, the step size of the imaginary time evolution is dynamically adjusted according to the maximum coupling strength of the spin chain, satisfying: Where, is the inductance parameter, is the capacitance parameter.

[0028] During the ground-state search for a quantum spin chain, the iterative step size of the imaginary-time evolution is adaptively adjusted by real-time calculation of the extreme values ​​of the coupling strength distribution between adjacent nodes. When abnormally high coupling strength is detected in a local region, the step size is dynamically shortened to improve evolutionary stability; for regions of low coupling strength, the step size is increased to accelerate convergence. This adjustment mechanism is implemented through real-time data exchange between the parameter calculation module and the quantum inversion module, ensuring a balanced evolutionary process with both efficiency and accuracy.

[0029] The imaginary time evolution approaches the system ground state through non-unitary transformations. The setting of the step size inversely proportional to the maximum coupling strength stems from the need to suppress quantum fluctuations in the high-coupling region. The dynamic adjustment strategy balances the contradiction between evolution accuracy and speed - the high-coupling region requires a small step size to accurately capture local state changes, while the low-coupling region can quickly cross the flat energy surface with a large step size. The parameter calculation module receives real-time coupling intensity distribution data from the quantum inversion module, generates step size control instructions through the extreme value extraction algorithm, and feeds back to the evolution calculation unit to form a closed-loop regulation. The collaborative working mechanism between modules enables the system to adapt to the heterogeneous characteristics of the power grid topology while avoiding the convergence oscillation or stagnation problems caused by the traditional fixed step size.

[0030] Preferably, the potential energy intensity of the fault point It is 3 times the normal value and is linearly mapped according to the wave head amplitude.

[0031] By extracting the amplitude characteristics of the traveling wave front in real time and establishing a linear proportional relationship between amplitude and potential energy intensity, the potential energy intensity at the fault point is set to three times the baseline value in a fault-free state. The potential energy parameters for the normal section are preset using the line impedance parameter library. When a fault occurs, a surge in the wave front amplitude triggers dynamic adjustment of the potential energy intensity. Abnormal potential energy parameters are written into the spin chain model via the quantum inversion module. The parameter mapping module, linked to the anomaly detection unit, updates the potential energy parameter table for each node in real time.

[0032] The 3-fold setting of the potential energy intensity stems from the energy accumulation effect caused by the sudden change in impedance at the fault point. This multiple relationship has been statistically verified by historical fault data and can effectively amplify the quantum state disturbance caused by the fault. The linear mapping of the wave head amplitude converts the physical signal intensity into the potential field intensity scale of the quantum system, ensuring that the severity of the fault is positively correlated with the amplitude of the change in entanglement entropy. The parameter mapping module receives the pre-processed wave head feature data, generates dynamic potential energy parameters after normalization, and synchronizes them to the quantum inversion calculation core through a high-speed interface. This design breaks through the limitations of traditional methods that rely on fixed threshold criteria, allowing the positioning sensitivity to automatically adjust with the fault current, avoiding missed detection of weak faults or false alarms of strong interference. Millisecond-level parameter synchronization is achieved between modules through a ring data bus to ensure real-time consistency between the quantum model and the power grid status.

[0033] Preferably, the dynamic adjustment of the amplitude of the applied signal includes establishing a random optimal control equation of the amplitude and iteratively solving it through the Hamilton-Jacobi-Bellman equation.

[0034] Based on real-time monitoring data of ambient noise and line impedance, a stochastic optimal control model for the applied signal amplitude is constructed. The noise interference is modeled as a random process and embedded in the control equations. The control strategy generator solves the Hamilton-Jacobi-Bellman equations via reverse recursion and iteratively calculates the optimal amplitude adjustment strategy. This solution integrates the current fault location confidence with historical adjustment records to generate dynamic adjustment commands that balance transient response and long-term stability. These commands drive the signal generator to perform amplitude corrections.

[0035] Stochastic optimal control quantifies the uncertainties of electromagnetic interference and load fluctuations into a probabilistic model, achieving robust regulation by minimizing the expected cost function. The Hamilton-Jacobi-Bellman equation dynamically plans the regulation path from a global optimal perspective, overcoming the short-sighted local optimization flaw of traditional PID control. The parameter generation module inputs real-time noise spectrum characteristics and impedance distribution data to the control strategy generator. The strategy generator uses a time-decomposition algorithm to iteratively solve the problem in stages and outputs amplitude adjustment coefficients to the signal injection module for dynamic compensation. A closed-loop feedback link is formed between the modules. Through a hybrid mechanism of stochastic optimization and deterministic regulation, signal injection stability and adaptability are achieved in strong interference environments.

[0036] A device based on the above method comprises: Multi-node traveling wave sensor array with a sampling rate of no less than 10 MHz and a time synchronization error of less than 10 ns; Heterogeneous processors, including ARM architecture CPUs and FPGA chips, run information geometry optimization algorithms and quantum inversion algorithms respectively; External signal generator, output amplitude is dynamically adjustable and frequency covers 0.1Hz-10MHz; High-speed communication module, supporting IEEE 1588 protocol and power line carrier redundant communication.

[0037] The present invention provides a method for accurately locating ground faults in distribution networks based on traveling wave detection and artificial intelligence. It has the following beneficial effects: 1. This invention utilizes multi-node synchronous signal acquisition and adaptive preprocessing technology to eliminate the coupling of power frequency interference and high-frequency noise. Compared to existing solutions that rely on a single sensor and a fixed filtering threshold, it overcomes the drawback of asynchronous sampling that leads to inaccurate extraction of wavefront time differences.

[0038] 2. This invention achieves high-fidelity simulation of traveling wave propagation by integrating random field modeling with nonlinear energy focusing. Conventional techniques simplify line parameters to a uniform distribution model, which fails to characterize the spatial heterogeneity of aging lines. This invention overcomes this limitation and significantly reduces the error in reflection wave amplitude prediction.

[0039] 3. This invention, based on quantum spin chain mapping and entanglement entropy mutation detection, improves fault location accuracy to sub-meter levels. Existing traveling wave methods are limited by wave velocity calibration errors and threshold sensitivity. This solution exploits the localized perturbation properties of quantum many-body systems, achieving precise location without pre-setting wave velocity parameters.

[0040] 4. This invention combines dynamic closed-loop control with a fault-tolerant signal injection strategy to ensure robust detection under complex operating conditions. Compared to traditional open-loop signal injection techniques, it overcomes the false triggering problem caused by signal distortion. An adaptive regulation mechanism ensures stable operation of the system even under strong electromagnetic interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] Please see the attached Figure 1 The embodiment of the present invention provides a method for accurately locating ground faults in a distribution network based on traveling wave detection and artificial intelligence, comprising the following steps: Step S1: Multi-node synchronous signal acquisition and adaptive preprocessing.

[0044] This step utilizes high-precision synchronous acquisition and adaptive filtering techniques to address the phase deviation and noise interference issues associated with asynchronous multi-node acquisition in the distribution network's traveling wave signals. This provides high-quality, spatially and temporally consistent input signals for subsequent random field modeling and quantum inversion. In terms of hardware architecture, a multi-channel synchronous sampling device and dynamic compensation mechanism ensure that the original signal is fully power-frequency suppressed and noise filtered before transmission to the central processing unit.

[0045] In some embodiments, the traveling wave voltage signal acquisition module includes an AD7606-16-bit analog-to-digital converter with a sampling rate configured at 10 MHz and an analog input range set to ±10 V. The sensor array is installed on a three-phase busbar, with spacing between adjacent sensors less than 5 cm to reduce electromagnetic coupling interference. The time synchronization module utilizes a dual-mode redundant design using GPS satellite clocks and IRIG-B codes: the GPS module uses the U-blox NEO-M8N, outputting a 1PPS pulse-per-second signal; the IRIG-B code decoder utilizes the Analog Devices ADM2582E chip, achieving a time resolution accuracy of 10 ns. The two are combined and output via a logic OR gate circuit to ensure that multi-node sampling clock deviations do not exceed 20 ns.

[0046] Specifically, the signal conditioning circuit consists of a preamplifier and an anti-aliasing filter. The preamplifier uses the AD8421 instrumentation amplifier, with a gain setting of 20dB and a bandwidth covering 0.1Hz to 50MHz. The anti-aliasing filter uses a 6th-order Bessel low-pass filter with a transfer function of:

[0047] Among them, the coefficient ~ By normalizing the cutoff frequency Calculation shows that the passband ripple is less than 0.1dB and the group delay fluctuation is controlled within 5ns.

[0048] In general, baseline correction is used to eliminate the interference of power frequency components on traveling wave head detection. Set the sliding integration window width to an integer multiple of the power frequency period: In general, baseline correction is used to eliminate the interference of power frequency components on traveling wave head detection. Set the sliding integration window width to an integer multiple of the power frequency period:

[0049] in, is an integer and is usually 1, is the power frequency of the power grid (50Hz or 60Hz). The average value of the signal in the window is:

[0050] The corrected signal is obtained by subtraction:

[0051] As an option, when harmonic resonance exists in the line, an adaptive notch filter can be used instead of the sliding integration method. Its transfer function is:

[0052] in, , is the sampling period, is the damping coefficient (usually 0.95 to 0.99).

[0053] For high-frequency noise suppression, wavelet threshold denoising uses the Symlet-8 wavelet basis for 6-layer decomposition. The noise standard deviation estimation formula is:

[0054] in, is the detail coefficient of the 6th layer. The hard threshold processing formula is defined as:

[0055] Among them, the threshold , The reconstructed signal retains the rising edge steepness of the wave head greater than 95%.

[0056] Preprocessed signal The data is transmitted to the central processing unit via Gigabit Ethernet in 32-bit floating point format with a timestamp accuracy of 1 ns. To meet the initial condition requirements for random field modeling in step S2, the signal time alignment error must meet the following requirements:

[0057] in, The highest effective frequency component of the traveling wave signal. The central processing unit performs frame check on the received signal. If the packet loss rate exceeds 0.1%, the redundant channel switching mechanism is triggered.

[0058] In one possible implementation, when ambient temperature fluctuations cause sensor drift, the temperature compensation circuit is enabled and the correction formula is:

[0059] in, is the gain temperature coefficient, is the offset temperature coefficient, is the calibration temperature. After compensation, the signal temperature drift error is less than 0.05% FSR.

[0060] As an option, when a single sensor failure is detected, spatial interpolation reconstruction is performed using adjacent node signals. The interpolation formula is:

[0061] Among them, the weight coefficient , is the decay length constant. The correlation coefficient between the reconstructed signal and the original signal is greater than 0.98.

[0062] In the case of non-stationary noise, variational mode decomposition (VMD) can be used instead of the wavelet threshold method. Its optimization objective function is:

[0063] in, is the modal function, is the center frequency, is the preset modal number (usually 6 to 8).

[0064] Step S2: Random field modeling and nonlinear energy propagation simulation.

[0065] This step takes the high-precision preprocessed signal output from step S1 and constructs a physical driving model of traveling wave propagation by integrating random field theory with nonlinear wave equations. This model quantifies the spatial heterogeneity and nonlinear effects of line parameters, providing precise dynamic constraints for subsequent information geometry optimization and quantum inversion, and addressing the cumulative positioning errors caused by parameter simplification and linear assumptions in traditional methods.

[0066] In some embodiments, the spatial distribution characteristics of the line parameters are characterized by a Gaussian random field. The resistance field expression is:

[0067] in, Indicates the nominal resistance value (unit: Ω / m), is the covariance kernel function, is the correlation length (unit: m), reflecting the spatial correlation of line aging degree; It is an independent increment Wiener process, which characterizes the random fluctuations caused by environmental disturbances.

[0068] Specifically, Determined by maximum likelihood estimation of historical fault data. For lines that have been in operation for more than 10 years, the typical value is m; Newly laid lines Inductive field With capacitance field Using the same modeling method, the corresponding parameters are 、 、 and 、 、 .

[0069] In one possible implementation, the dynamics of traveling wave propagation is described by the modified Schrödinger equation:

[0070] in, is the normalized voltage traveling wave amplitude (dimensionless), whose initial condition comes from the output signal of step S1 , the time alignment error is less than 2ns; is the nonlinear coefficient (unit: m2 / V2), which is calculated in real time from the line current:

[0071] Here, is the average line current (unit: A), Corresponding to the center frequency of the traveling wave 1MHz; potential energy term Integrate random field parameters:

[0072] This potential energy term reflects both the resistive loss and the inductive-capacitive coupling effect, and its imaginary part characterizes the energy dissipation characteristics.

[0073] Typically, spatial discretization uses a spectral method, configuring 512 Fourier modes covering the 0–50 MHz frequency range. Time integration uses a fourth-order Runge-Kutta algorithm with a step size of 0.1 μs, satisfying the Nyquist sampling theorem. GPU acceleration is achieved using CUDA kernel functions, with 1024 parallel threads, reducing computation time by 18–22 times compared to CPU.

[0074] Specifically, the boundary conditions are set to non-reflective to avoid standing wave interference:

[0075] in, is the nominal wave velocity (unit: m / s), is the total length of the line.

[0076] As an option, the line joints and branch points parameter mutation areas are smoothed by window function. Define the half-width of the transition area , the window function form is:

[0077] in, is the center position of the mutation point (unit: m), Control the transition steepness. The corrected resistance field is:

[0078] Here, and Represent the resistance parameters on both sides of the mutation point respectively.

[0079] In one possible implementation, when it is detected that the line current exceeds the safety threshold When , the nonlinear coefficient is dynamically amplified:

[0080] This adjustment compensates for the additional loss caused by corona discharge and avoids abnormal amplitude attenuation.

[0081] As an extension of the static random field, stochastic differential equations can be used to describe the time-varying characteristics of parameters:

[0082] in, is the regression rate, which controls how quickly the parameters converge to the nominal values; is the fluctuation intensity, reflecting the random influence of changes in ambient temperature and humidity; is the space-time joint Wiener process increment.

[0083] Step S3: Information geometry optimized traveling wave mode separation.

[0084] This step, based on the traveling wave propagation model containing nonlinear effects and random field perturbations output from step S2, achieves precise separation of multipath reflection wave heads by constructing a high-dimensional signal manifold and optimizing the geometric path. This method combines physics-driven wave equation constraints with data-driven geometric learning to address the positioning failure problem caused by modal aliasing in traditional methods in ultra-large-scale distribution networks, providing a pure wave head signal aligned in time and space for quantum inversion in step S4.

[0085] In some embodiments, the set of traveling wave signals is defined as a differential manifold embedded in the Hilbert space The local coordinate system of the manifold consists of three parameters: Composition, of which: is the nonlinear coefficient calculated in step S2 (unit: m2 / V2), which is directly inherited from the nonlinear wave equation; is the spatial mean of the potential energy term (unit: 1 / m2), which is calculated as , the integration interval covers the entire length of the line; The candidate fault location (unit: m) is generated by sampling at 10m intervals, covering the range .

[0086] Specifically, the Riemann metric tensor is calculated by the inner product of the signal partial derivatives:

[0087] Here, the partial derivative The central difference method is used for approximation, and the step size is set to 1% of the parameter value. Set to 2 times the round trip time of the traveling wave (typical value: 1km line ).

[0088] In one possible implementation, the following Finsler norm minimization problem is constructed to extract the main modes:

[0089] in: is the sparse regularization coefficient (dimensionless), with a default value of 0.1, which is used to suppress interference from secondary reflection paths; is the Symlet-8 wavelet transform operator, and its frequency band division is consistent with the preprocessing stage in step S1; is the parameter path change rate.

[0090] The initial set of paths is generated by Latin hypercube sampling, covering the parameter range: , ,

[0091] Here, and The potential energy field in step S2 Statistically derived.

[0092] In general, the momentum gradient descent method is used to update the path parameters:

[0093] in: is the adaptive step size (dimensionless), with an initial value of 0.01 and a decay of 50% every 10 iterations; is the momentum factor, accelerating convergence and escaping from the local minimum; It is a tangent space projection operator that ensures that the path update lies on the manifold tangent plane.

[0094] When screening paths, calculate the loss function values ​​of all paths , retain satisfaction The top 5% paths, threshold The final output is the weighted average of the retained paths: ,

[0095] When multiple local extreme values ​​are detected in the parameter space (multiple fault point scenarios), the adaptive regularization mechanism is enabled:

[0096] in, is the initial coefficient, is the current gradient norm, is the sensitivity factor. This mechanism strengthens the sparse constraint in flat areas and relaxes the constraint in steep areas to avoid overfitting.

[0097] In one possible implementation, an exclusion term is introduced for redundant paths to improve search diversity:

[0098] in, is the repulsion strength coefficient, A smoothing term to prevent the denominator from being zero.

[0099] Step S4: Quantum spin chain mapping and tensor network inversion positioning.

[0100] This step, based on the spatiotemporally aligned traveling wave modal signal output from step S3, achieves sub-meter ground fault location through quantum many-body system analogy and tensor network optimization. This method maps the physical topology of the distribution network into a spin chain model and exploits the localized mutation characteristics of quantum entanglement entropy to capture microscopic perturbations at the fault point, overcoming the accuracy bottleneck of traditional traveling wave methods, which are limited by wave velocity calibration errors.

[0101] In some embodiments, the distribution network is discretized into a spin chain model, and the node spacing is fixed to Δx = 1m Δx = 1m. The Hamiltonian expression is:

[0102] in: (Unit: Hz), the inductance output by step S2 (Unit: H / m) and capacitance (Unit: F / m) Export; (Unit: Hz), where Hz / V is the amplitude-frequency conversion coefficient, is the wave front amplitude extracted in step S3; 、 is the spin-up / down operator, is the Pauli-Z operator.

[0103] Specifically, the potential energy term of the node corresponding to the fault point is It is set to 3 times the normal value. When the fault point .

[0104] In one possible implementation, the base state search uses an imaginary time evolution algorithm. Initialize the uniform matrix product state (MPS) , initial bond dimension . Evolution step Dynamic adjustment based on coupling strength:

[0105] The single-step evolution operator is:

[0106] Here, The current energy expectation value. The truncation strategy sets the maximum bond dimension. , truncation error threshold .

[0107] In general, the reduced density matrix of node i is calculated as:

[0108] The entanglement entropy is obtained by eigenvalue decomposition of the density matrix:

[0109] in, for The kth eigenvalue of is the spin dimension. The fault criterion is set as:

[0110] Here, is the mean entropy of all nodes in the network, is the standard deviation.

[0111] When the entanglement entropy spatial gradient is detected to satisfy Enable dynamic key dimensioning when:

[0112] in, is the gradient threshold, m is the node spacing.

[0113] In one possible implementation, the imaginary time step Adaptive adjustment based on energy residual:

[0114] in, is the energy difference between adjacent iterations.

[0115] Step S5: Adaptive adjustment of the closed-loop control signal and fault-tolerant injection.

[0116] This step, based on the sub-meter fault location results output from step S4, constructs a quantum-classical hybrid control loop by dynamically adjusting the amplitude-frequency characteristics and phase response of the applied excitation signal. This solves the problem of location failure caused by signal distortion and noise coupling under complex operating conditions. This method maps the quantum entanglement entropy mutation characteristics of the fault point into classical control parameters, achieving adaptive matching of signal energy and grid status, and ensuring the robustness of the detection process under strong interference.

[0117] In some embodiments, the Hamiltonian of the closed-loop control system is designed as:

[0118] in: 、 are the quantization operators of the injected signal and the feedback signal, respectively, and their eigenvalues ​​correspond to the classical signal amplitude (unit: V); (dimensionless), The voltage amplitude of the fault point located in step S4 (unit: V); , time constant (determined by the attenuation characteristics of the pre-processed signal in step S1); , SNR is the signal-to-noise ratio (dimensionless) calculated in real time in step S1.

[0119] Specifically, the operator product term Describes the coupling effect between signal and feedback, and its coefficient Inversely proportional to the signal-to-noise ratio, the coupling strength is enhanced to suppress noise at low signal-to-noise ratios.

[0120] In one possible implementation, the control parameters are updated via stochastic gradient descent:

[0121] in: is the parameter vector to be optimized (dimensionless); is the learning rate (dimensionless), which determines the parameter update step size; is the L2 regularization coefficient (dimensionless) to prevent overfitting; By Monte Carlo sampling estimation, the number of samples (Determined by the path screening result of step S3).

[0122] The algorithm is executed every 10ms, synchronized with the power grid frequency cycle, to avoid phase mismatch caused by control delay.

[0123] In general, the injection signal is output by the AD9106 waveform generator, whose digital-to-analog converter (DAC) resolution is 16 bits, the amplitude range is -10V to +10V, and the update rate is 1MHz (consistent with the sampling rate of step S1). Before output, it is processed by a second-order Butterworth filter with a cutoff frequency of MHz, group delay compensation , to ensure that the signal waveform is not distorted.

[0124] When the line voltage exceeds the safety threshold, the overvoltage protection mechanism is triggered:

[0125] Here, is the line nominal voltage (unit: V), The voltage value (unit: V) monitored in real time in step S1.

[0126] When the signal-to-noise ratio of step S1 is lower than 20dB, the noise injection enhancement mechanism is enabled:

[0127] This mechanism actively introduces controllable noise under strong interference to break the lock of local extreme values ​​on the control system.

[0128] In one possible implementation, the control weight According to the fault confidence probability output in step S4 Dynamic Adjustment:

[0129] when When , the weight is saturated to avoid overdriving; when When , the weight is reset to zero to reduce the risk of false triggering.

[0130] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for accurately locating ground faults in distribution networks based on traveling wave detection and artificial intelligence, characterized in that: The following steps are involved: (a) The traveling wave voltage signal of the distribution network is collected by a multi-node synchronous sensor array, and the preprocessed signal is generated through baseline correction and wavelet threshold denoising; (b) Construct a Gaussian random field model of line resistance, inductance, and capacitance, and combine it with nonlinear wave equations to simulate the traveling wave propagation process; (c) Finsler metric is defined on the traveling wave signal manifold to separate aliased modes by minimizing the path integral; (d) Mapping the distribution network into a quantum spin chain, calculating the local entanglement entropy based on virtual time evolution and tensor network inversion, and locating the fault point; (e) Dynamically adjust the amplitude of the applied signal according to the mutation characteristics of the entanglement entropy to optimize the detection sensitivity; Wherein, the Hamiltonian of the quantum spin chain in step (d) is: Where, is the inductance parameter, is the capacitance parameter, is the potential energy intensity of the fault point, 、 、 is the spin operator.

2. The method for accurately locating ground faults in a distribution network based on traveling wave detection and artificial intelligence according to claim 1 is characterized in that: The baseline correction eliminates the power frequency component by integrating a sliding window, and the window width is an integer multiple of the power frequency period.

3. The method for accurately locating ground faults in a distribution network based on traveling wave detection and artificial intelligence according to claim 1 is characterized in that: The covariance kernel function of the Gaussian random field is determined by maximum likelihood estimation of historical data, and the correlation length is set to 5 meters or 10 meters according to the aging degree of the line.

4. The method for accurately locating ground faults in a distribution network based on traveling wave detection and artificial intelligence according to claim 1 is characterized in that: The nonlinear wave equation contains a nonlinear self-focusing coefficient calculated from the line current dynamics.

5. The method for accurately locating ground faults in a distribution network based on traveling wave detection and artificial intelligence according to claim 1 is characterized in that: The path integral optimization of the Finsler metric adopts the momentum gradient descent method to search at least 100 paths in parallel and screen the optimal solution.

6. The method for accurately locating ground faults in a distribution network based on traveling wave detection and artificial intelligence according to claim 1 is characterized in that: The step size of the imaginary time evolution is dynamically adjusted according to the maximum coupling strength of the spin chain, satisfying: Where, is the inductance parameter, is the capacitance parameter.

7. The method for accurately locating ground faults in a distribution network based on traveling wave detection and artificial intelligence according to claim 1 is characterized in that: The potential energy intensity of the fault point It is 3 times the normal value and is linearly mapped according to the wave head amplitude.

8. The method for accurately locating ground faults in a distribution network based on traveling wave detection and artificial intelligence according to claim 1 is characterized in that: The dynamic adjustment of the amplitude of the externally applied signal includes establishing a random optimal control equation of the amplitude and iteratively solving it through the Hamilton-Jacobi-Bellman equation.

9. A device based on the method according to any one of claims 1 to 8, characterized in that: include: Multi-node traveling wave sensor array with a sampling rate of no less than 10 MHz and a time synchronization error of less than 10 ns; Heterogeneous processors, including ARM architecture CPUs and FPGA chips, run information geometry optimization algorithms and quantum inversion algorithms respectively; External signal generator, output amplitude is dynamically adjustable and frequency covers 0.1Hz-10MHz; High-speed communication module, supporting IEEE 1588 protocol and power line carrier redundant communication.

Images