Rail transit catenary power supply load prediction and self-healing control method

Abstract

The present application relates to a rail transit catenary power supply load prediction and self-healing control method, and relates to the technical field of rail transit power supply prediction. Unfiltered high-frequency voltage signals, environmental stability indicators at corresponding time points and train position information are collected to screen stable sample data. The unfiltered high-frequency voltage signals are subjected to spatial mapping to generate fixed spatial point labels. Multi-scale time-frequency decomposition is performed to extract high-frequency structure feature vectors. A long-term feature baseline model is constructed, and a feature distance change rate is calculated to form a risk trend indicator. A spatial risk weight correction prediction load value is generated, and a self-healing control instruction is generated based on the corrected prediction load value. The dynamic coupling of structure state evolution and power supply load prediction is realized, the risk identification accuracy is improved, and the self-healing control capability of the power supply system is enhanced.

Description

Technical Field

[0001] This invention relates to the field of methods for predicting and self-healing the power supply load of rail transit overhead contact lines, specifically a method for predicting and self-healing the power supply load of rail transit overhead contact lines. Background Technology

[0002] As a critical infrastructure for train traction power supply, the operating status of the rail transit overhead contact line power supply system directly affects train operation safety and power supply stability. During high-speed train operation, the pantograph and the overhead contact line are in continuous dynamic contact. Affected by mechanical vibration, changes in structural stiffness, and environmental disturbances, a large number of high-frequency micro-fluctuations are inevitably superimposed on the power supply voltage signal.

[0003] Existing technologies typically treat high-frequency voltage fluctuations collected during train operation as random interference signals caused by wind speed changes, vehicle vibration, or instantaneous contact jitter. These fluctuations are generally eliminated using signal processing methods such as filtering and smoothing, with only low-frequency or average data retained for power supply load prediction or anomaly detection.

[0004] The inventors of this application have discovered that the above-mentioned technology has at least the following technical problems: First, the existing technology uniformly filters and smooths the high-frequency micro-fluctuations of voltage generated during train operation. Although this can improve signal stability, it also weakens or even masks the subtle electrical disturbances caused by changes in the stiffness of the contact network structure, fluctuations in contact pressure, or local mechanical loosening, making it difficult to identify early structural degradation signals. Second, the existing technology mainly analyzes voltage signals based on overall changes in the time dimension, without establishing a stable correspondence between high-frequency micro-fluctuations of voltage and fixed spatial points. It is difficult to repeatedly compare the characteristic changes of the same physical structure location at different time periods, and it is impossible to form a spatially directional trend judgment. Third, the existing load forecasting and self-healing control strategies mainly adjust based on the operation plan or real-time load data, failing to incorporate the gradual changes in the state of the contact network structure into the prediction and correction logic. Therefore, when the structure is in a slow deterioration stage, it is difficult to form a feedforward adjustment mechanism in a timely manner, resulting in technical defects of prediction lag and delayed control response. Summary of the Invention

[0005] This application provides a method for predicting and self-healing the power supply load of rail transit catenary, which transforms the "background noise" in the unfiltered high-frequency voltage signal in the rail transit catenary power supply system into effective features, predicts the degradation trend of the catenary physical structure in advance, and thus realizes self-healing control and maintenance optimization based on risk trends.

[0006] To achieve the above objectives, the embodiments of this application disclose the following technical solutions:

[0007] This solution discloses a method for predicting and self-healing the power supply load of rail transit overhead contact lines, including the following steps:

[0008] Step S1: The processor collects unfiltered high-frequency voltage signals, corresponding environmental stability indicators, and train position information. Based on the environmental stability indicators, stable sample data that meets the preset stability interval is obtained. The technical motivation of this step is to solve the problem that the contact network voltage fluctuation signal is easily interfered with by unstructured external variables such as sudden weather changes and severe power grid fluctuations in the complex electromagnetic environment of rail transit. If environmental stability indicators are not introduced for interval screening, the "pseudo-features" caused by strong winds or sudden large loads in the original data will seriously interfere with the judgment of the physical structure degradation trend. By screening stable sample data that meets the preset stability interval, the variable control principle in statistics is essentially used to strictly limit the causal chain of subsequent analysis to "physical structure degradation" and "micro-voltage fluctuations", providing high-purity feature input for subsequent steps.

[0009] Step S2: Spatially map the unfiltered high-frequency voltage signal in the stable sample data with the train position information to generate fixed spatial point labels and form a point-bound voltage sequence. Since the train's speed, acceleration, and operating conditions are constantly changing during operation, traditional time series analysis cannot continuously track the fixed position status of the overhead contact line. This application does not simply record time-point data, but establishes a decoupled model of "time sampling - spatial node" through spatial mapping. The necessity of this design lies in using the train pantograph as a "moving probe" to project the dynamically changing electrical signal onto a fixed spatial dimension, enabling the system to perform long-term "fixed-point auscultation" of specific supports, droppers, and other physical nodes on the overhead contact line, regardless of vehicle type, speed, or shift.

[0010] Step S3: Perform multi-scale time-frequency decomposition on the point-bound voltage sequence to obtain high-frequency structural disturbance components and extract high-frequency structural feature vectors. To extract weak precursors of structural failure from a composite signal containing fundamental, harmonic, and random noise, this application utilizes the multi-resolution analysis characteristics of non-stationary signals. By performing multi-scale time-frequency decomposition, high-frequency structural disturbance components strongly correlated with the mechanical vibration and micro-arc discharge frequency of the contact network can be accurately extracted. The correlation effect of extracting high-frequency structural feature vectors lies in compressing high-dimensional, massive amounts of raw waveform data into key quantitative indicators characterizing structural stiffness and contact pressure, greatly enhancing the sensitivity of identifying early sub-health conditions of the contact network.

[0011] Step S4: Based on the high-frequency structural feature vector, construct a long-term feature baseline model corresponding to fixed spatial point labels, and calculate the feature distance change rate between the current high-frequency structural feature vector and the long-term feature baseline model to form a risk trend indicator. The mechanism of constructing the long-term feature baseline model lies in establishing a set of digital companion criteria that can be dynamically adjusted with the natural aging of physical assets. This design avoids the technical pain point that the traditional fixed threshold judgment method cannot adapt to the long-term performance evolution of equipment. The necessity of calculating the feature distance change rate instead of a simple distance value lies in the fact that by introducing the concept of the first derivative (change rate), the system can keenly capture the "critical inflection point" of risk transformation from linear accumulation to nonlinear outbreak, thereby realizing the logical sublimation from "post-event alarm" to "feedback-based evolution prediction".

[0012] Step S5: Generate spatial risk weights based on the risk trend indicators, and input these spatial risk weights into the power supply load prediction model. Correct the predicted load value output by the power supply load prediction model, and generate self-healing control commands based on the corrected predicted load value. This step achieves a deep closed loop between physical health risk and power dispatch logic. Its core mechanism lies in transforming the uncertainty of physical structure into a quantifiable constraint on power supply capacity. Inputting spatial risk weights into the power supply load prediction model essentially embeds a "dynamic adjustment factor for safety margin" into the traditional power consumption plan. This allows the generated self-healing control commands to perform refined power reduction or path reconfiguration based on the real-time carrying capacity of the overhead contact line, avoiding large-scale power outages caused by overload impacts at physical hazard points.

[0013] Furthermore, the mechanism for calculating the frequency band energy density, kurtosis value, and peak interval period lies in performing a "three-dimensional imaging" of the contact network condition from three dimensions: energy distribution, statistical impact characteristics, and signal regularity. The kurtosis value can identify instantaneous physical impacts, while the peak interval period can characterize the frequency regularity of micro-arc discharges. This synergistic effect of the multi-dimensional frequency band feature set enables the high-frequency structural feature vector to possess the analytical ability to distinguish between "environmental random noise" and "deterministic physical fatigue."

[0014] Furthermore, the recursive update model incorporating time decay characteristics is designed to endow the algorithm with the ability to "selectively forget." This is necessary because the electrical characteristics of the overhead contact line are subject to reasonable, slow drift due to temperature variations and normal wear. Without a recursive update mechanism, outdated historical data would render the baseline invalid. Through this decay-based update method, the system can maintain a precise definition of the "current normal state," thus ensuring that risk trend assessments are based on real-time and reliable physical evolution.

[0015] Furthermore, this step utilizes a sliding time window for trend analysis to eliminate random errors introduced by monitoring at a single time point. The mechanism involves extracting trend changes reflecting the rate of equipment performance degradation through multi-point regression analysis within the window. This design effectively filters out "pulse-like false alarms" caused by momentary pantograph contact failures, allowing the risk trend indicators to more accurately reflect the mechanical fatigue or creep process of the overhead contact line's physical structure.

[0016] Furthermore, the motivation for establishing a spatial topology graph structure is to transform discrete monitoring points into a physically interconnected system network. In this structure, node voltage time series are no longer isolated numerical values, but rather a physical flow carrying information about the overhead contact system topology constraints. This topology mapping method provides underlying logical support for subsequent spatial consistency verification, enabling the system to leverage the correlation between neighboring nodes to enhance the determinism of anomaly detection.

[0017] Furthermore, spatial consistency verification incorporates the technical mechanism of "swarm intelligence verification," aiming to distinguish whether anomalies originate from the "mobile end (vehicle)" or the "fixed end (network)." The logical necessity lies in this: if the feature vectors of multiple adjacent nodes within a certain area exhibit consistent behavior, while only the target node shows a continuous shift, then it is highly probable that the anomaly is located in the catenary structure at that specific point. This judgment method based on spatial neighborhood differences completely solves the problem of interference from fluctuations in the pantograph condition of a single train on catenary status identification, greatly improving the robustness of diagnostic conclusions.

[0018] Furthermore, by weighting spatial risk weights with the basic predicted load value, a mapping operator is essentially established from the "physical risk space" to the "electricity resource demand space." Through this dynamic weighting, the system can proactively reduce the predicted power load in a segment when the physical structure enters a sub-healthy state. This synergistic effect reserves sufficient buffer space for grid dispatch, ensuring the scientific and predictive nature of self-healing control commands.

[0019] Furthermore, this feature provides a variety of specific, physically applicable self-healing response methods. Power limiting control commands can directly reduce the instantaneous stress and arcing impact on the overhead contact line, while peak-shaving control commands reduce peak loads through time-based scheduling, and backup path switching avoids high-risk points at the system topology level. The flexible combination of these three features achieves comprehensive risk mitigation for the power supply network, significantly reducing the macro-level operational risks caused by micro-degradation of the overhead contact line.

[0020] Furthermore, the necessity of range-based judgments for wind speed, temperature, and load levels stems from the fact that these environmental dimensions have significant nonlinear effects on the tension of the overhead contact line, the amplitude of the pantograph, and the foundation voltage. This application constructs a multi-dimensional "stationary observation window" to eliminate "invalid feature noise" under non-stationary operating conditions, thereby ensuring that the subsequently extracted structural feature vectors have an extremely high signal-to-noise ratio.

[0021] Furthermore, the risk trend smoothing step utilizes the low-pass filtering principle in the time domain. Its technical motivation lies in eliminating spike disturbances during risk identification and preventing frequent and ineffective actions by the self-healing control execution unit. Through smoothing, the system can output more robust and sustainable spatial risk weights, forming a stable closed loop in the entire self-healing control process and significantly improving the overall operational reliability of the rail transit power supply system.

[0022] Compared to existing rail transit power supply systems that rely solely on timetables or low-frequency voltage signals for load estimation, this invention identifies long-term trends in the physical structure of the overhead contact system by collecting unfiltered high-frequency voltage fluctuations and combining them with environmental stability indicators. Traditional methods cannot reflect gradual structural degradation; this invention, by constructing a long-term characteristic baseline model and risk trend indicators, binds minute disturbances to spatial locations, enabling feedforward prediction correction and self-healing control, thereby improving power supply safety and prediction accuracy.

[0023] Compared to existing rail transit power supply systems that rely solely on timetables or weather forecasts for load estimation, which struggle to reflect the degradation of the overhead contact system's physical structure and are prone to overload or abnormal discharge when the structure enters a sub-healthy phase, this invention addresses these issues. By collecting minute fluctuations and noise from unfiltered high-frequency voltage signals and combining this with environmental stability indicators to screen stable sample data, this invention spatially maps these voltage fluctuations (considered "noise" by traditional algorithms) to fixed spatial location labels. It extracts high-frequency structural feature vectors, constructs long-term characteristic baseline models for corresponding locations, and generates risk trend indicators based on the characteristic distance change rate. Furthermore, it generates spatial risk weights to correct predicted load values, achieving a dynamic correlation between voltage fluctuations, overhead contact system structural evolution, and power supply load prediction. This solution can identify "deteriorating" trends before any apparent faults appear in the overhead contact system. Through a self-healing control decision module, it generates power-limiting control commands or train start-up peak-shaving control commands in advance, reducing the probability of power supply capacity thresholds being triggered. This significantly improves prediction accuracy and control foresight, enhancing the safety and stability of the power supply system in complex operating environments. Attached Figure Description

[0024] Figure 1 This is an overall flowchart of an embodiment of the present invention;

[0025] Figure 2This is a flowchart of the multi-scale time-frequency decomposition of an embodiment of the present invention;

[0026] Figure 3 This is a flowchart illustrating the risk trend calculation method according to an embodiment of the present invention.

[0027] Figure 4 This is a topology diagram of the overhead contact line according to an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In the following description, numerous specific details are set forth to provide a comprehensive understanding of the present invention. The present invention may be practiced without some or all of these specific details. In other instances, well-known processes have not been described in detail to avoid unnecessarily obscuring the present invention.

[0029] When used in conjunction with the terms "comprising," "method comprising," or similar language in this specification and appended claims, the singular forms "a," "some," and "the" include plural references unless the context clearly indicates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0030] Terminology definition:

[0031] In this embodiment, "unfiltered high-frequency voltage signal" refers to the raw voltage waveform data directly collected by the voltage sampling device in the contact power supply circuit without low-pass filtering or smoothing. Its sampling frequency is several orders of magnitude higher than the fundamental frequency of traction power supply, and is used to retain information on minute disturbances. This type of signal contains a large amount of transient information reflecting physical contact quality, micro-arc discharge and mechanical vibration characteristics.

[0032] "Environmental stability indicators" include wind speed, temperature, and load levels, used to characterize whether external disturbance conditions are in a comparable state. They are a set of physical quantities reflecting irregular interferences from the external environment on the power supply system, including at least instantaneous wind speed, ambient temperature, and the overall load level of the regional power grid.

[0033] "Stable sample data" refers to the set of unfiltered high-frequency voltage signals collected when the environmental stability index meets the preset stability range.

[0034] "Fixed spatial location tags" are index information that digitally identifies the physical location of the overhead contact line, used to map time series data to spatial nodes.

[0035] "Point-bound voltage sequence" refers to a set of voltage time series data belonging to the same fixed spatial point label.

[0036] The "high-frequency structural feature vector" is a set of multi-dimensional statistical features extracted by multi-scale time-frequency decomposition of the point-bound voltage sequence, which is used to characterize the perturbation features of the catenary structure.

[0037] The "long-term characteristic baseline model" is a characteristic reference model formed for a fixed spatial location under healthy conditions. It is a dynamic digital benchmark that characterizes the electrical response of a specific point in the overhead contact system under healthy conditions. By accumulating historical stable samples, it forms a tolerance threshold for normal physical aging at that point.

[0038] The "Risk Trend Indicator" is a quantitative indicator that reflects the direction and rate of change of characteristic distance.

[0039] "Spatial risk weight" is a factor that maps risk trend indicators to a correction factor for the power supply load forecasting model.

[0040] The "self-healing control command" is a control command generated by the self-healing control decision module and sent to the power supply control execution unit.

[0041] Application Overview:

[0042] This scheme utilizes micro-fluctuations in unfiltered high-frequency voltage signals under largely controllable environmental conditions as a sensitive variable for structural state. It constructs a structural risk trend model through spatial mapping, multi-scale decomposition, long-term characteristic baseline models, and trend evolution analysis. Furthermore, it embeds risk trend indicators into the power supply load prediction model to form a closed-loop self-healing control process, thereby achieving a shift from passive fault detection to proactive risk prediction and control. Mature technologies typically focus on instantaneous anomaly detection or low-frequency voltage fluctuation analysis, while this scheme focuses on the long-term repetitive changes and spatially consistent evolution of high-frequency micro-fluctuations; the two approaches differ fundamentally in their technical approaches.

[0043] This embodiment discloses a method for predicting and self-healing the power supply load of rail transit overhead contact lines, including the following steps:

[0044] Step 1: The processor collects unfiltered high-frequency voltage signals, environmental stability indicators, and train position information at the corresponding time. Based on the environmental stability indicators, stable sample data that meet the preset stable range are obtained. The processor performs range judgment on wind speed, temperature, and load level in the environmental stability indicators. When each indicator is simultaneously within the corresponding preset range, the unfiltered high-frequency voltage signal of the corresponding time period is marked as stable sample data and stored in the stable sample data cache for subsequent processing.

[0045] In this embodiment, the processor is electrically connected to the overhead contact line voltage sampling sensor via a high-speed data acquisition module to collect unfiltered high-frequency voltage signals in real time. Simultaneously, it acquires wind speed and temperature data through a meteorological acquisition interface and load level and train location information through a dispatching system interface. The processor performs interval judgment on the wind speed, temperature, and load level in the environmental stability indicators. Only when the environmental stability indicators are within a preset stable interval is the unfiltered high-frequency voltage signal for the corresponding time period marked as stable sample data and stored in the stable sample data cache. The specific value of the preset stable interval is not limited to a fixed value but is determined by statistically analyzing the distribution of historical environmental indicators under normal operating conditions, combined with the experience of technical experts or system tolerance. For different lines, seasons, and operating conditions, the stable interval can be dynamically adjusted to ensure that the interval accurately reflects the stable state of the environment, thereby ensuring the reliability of the stable sample data. To ensure complete acquisition of high-frequency micro-fluctuation characteristics, the sampling frequency of the unfiltered high-frequency voltage signal is set to be greater than 10kHz and includes frequency components in the range of 1kHz to 50kHz.

[0046] To ensure the longitudinal comparability of subsequently extracted micro-features, unstructured noise caused by drastic environmental fluctuations must be removed. This embodiment constructs a Mahalanobis distance discriminant space based on a multi-dimensional environmental vector, defines the deviation of the current environmental state vector from the ideal baseline environmental distribution, and thus derives the environmental stability discriminant equation:

[0047] ;

[0048] Among them, when the calculated stability coefficient Greater than or equal to the preset stability threshold When this occurs, the current operating state is determined to be within a statistically stable range;

[0049] Environmental stability coefficient This is the final output of the formula, a scalar value within the interval (0,1), which intuitively reflects the confidence level of the current environmental sampling point in the multidimensional probability space. When this coefficient approaches 1, it means that the current wind speed, temperature, and load fluctuations highly coincide with historical benchmarks, and the system is in an extremely stable observation window; conversely, if the coefficient is below the preset threshold, it indicates that the environmental noise is too high, and the currently collected high-frequency voltage signal may contain a large amount of unstructured interference, which should not be included in the long-term trend analysis. Stability Threshold The stability coefficient can be determined by statistical distribution analysis of historical stability coefficients, combined with system tolerance or prediction model error level. During system operation, the threshold can be dynamically adjusted based on real-time environmental fluctuations and historical data trends to ensure the reliability and sensitivity of the stability interval determination.

[0050] Environmental stability index vector It is a dynamic input array captured in real time by the processor, which vectorizes and maps multiple external variables affecting the electrical characteristics of the overhead contact system. In actual acquisition, wind speed and ambient temperature are usually transmitted back in real time by sensors from meteorological stations located in substations or key geographical nodes, while load levels are extracted by the SCADA interface of the power grid dispatching system. This vector captures a "snapshot" of the environment at a specific moment and is the cornerstone of all subsequent calculations.

[0051] Preset environmental baseline mean vector This represents the "golden environment center" that the system considers most favorable for observing the physical degradation of the overhead contact system. It is typically derived by the processor after long-term statistical analysis of historical operational data for that section. During the algorithm initialization phase, technical experts select a large number of samples with stable train operation, no external faults, and favorable weather conditions. The expected values ​​of the environmental indicators for these samples are calculated to determine $\mu_e$. This serves as a static reference point, determining the geometric center of the stability assessment space.

[0052] The inverse matrix of the environmental covariance matrix This is the most crucial weighting mechanism in the formula, its physical meaning lying in defining the correlation between different environmental dimensions and their sensitivity to stability. This matrix doesn't simply assign equal weights to each indicator; instead, it eliminates dimensional differences between variables through covariance calculations. For example, if a small fluctuation in wind speed has a much greater impact on high-frequency voltage than temperature, then the corresponding eigenvalue in the matrix will be set larger, thus playing a dominant role in Mahalanobis distance calculations. This parameter is typically obtained offline by inverting the covariance matrix of massive historical operating condition data.

[0053] Stability threshold This is a logic gate set by the system, and its value directly determines the stringency of the stable sample selection. This threshold is usually preset by the R&D architect based on the electromagnetic environment complexity of the circuit and the fault tolerance of the prediction model. For example, setting it to 0.9 means that only extremely stable samples within the 90% confidence interval are accepted. This parameter is defined at the software configuration layer and is called as a constant when the processor executes the selection logic, thereby achieving physical isolation from "noisy data".

[0054] The technical principle of this step is to eliminate the disturbance of high-frequency micro-fluctuations in voltage caused by extreme weather or external power grid fluctuations, so that the subsequent characteristic offsets only reflect changes in the physical structure of the contact network itself, thereby improving the causal purity of the risk trend indicators from the source.

[0055] Step 2: Spatial mapping is performed between the unfiltered high-frequency voltage signal in the stable sample data and the train position information to generate fixed spatial point labels and form point-bound voltage sequences; the processor establishes a spatial topology graph structure based on the train position information and the catenary topology information, where nodes represent monitoring points and edges represent physical adjacency relationships; the unfiltered high-frequency voltage signal is mapped to the corresponding nodes according to the time period of the train passage to generate node voltage time series, and output to the multi-scale time-frequency decomposition step.

[0056] In this embodiment, the processor establishes a spatial topology map structure based on train location information and catenary topology information, where nodes represent monitoring points corresponding to the physical support structure of the catenary, and edges represent the physical connection relationships between conductors. The processor spatially segments the unfiltered high-frequency voltage signal according to the time window of the train passing through each node, projects the continuous time series onto discrete nodes corresponding to the track kilometer markers, thereby generating node voltage time series and forming fixed spatial point labels.

[0057] Considering the dynamic slippage between traction calculations and actual physical positions during train operation, this embodiment introduces a compensation coefficient to correct the kilometer markers in real time. Based on the real-time train speed and wheel wear compensation, time sampling points are constructed. With orbital coordinate nodes Nonlinear mapping equation between:

[0058] ;

[0059] This leads to the spatially bound voltage sequence. The sampling time is coming soon The voltage amplitude is projected onto its corresponding physical node coordinates. .

[0060] in, represent The precise track position corresponding to a given moment is the target spatial coordinate value that the formula ultimately solves for, usually expressed in kilometer markers. At the processing layer, the processor uses this coordinate to "suspend" the high-frequency voltage amplitude collected at the current moment onto a specific node of the digital twin map, thereby ensuring that the microscopic fluctuation data generated by different trains and different schedules when passing the same physical support can be compared longitudinally in the same spatial dimension.

[0061] This represents the starting kilometer marker position of the train and serves as the reference zero point for this positioning calculation. This parameter is typically read in real-time by the onboard train control system through the dispatching system interface when the train starts or passes a specific transponder (balise). It marks the physical starting point of this integration calculation, and its accuracy is ensured by fixed-position reference calibration equipment installed beside the track.

[0062] Representing the real-time operating speed of the train, this is an integrand that changes continuously with time. In actual acquisition, this parameter is collected in real time by speed transmission sensors or Doppler radar installed at the train axle ends, reflecting the instantaneous displacement rate of the train within a microscopic time slice. This speed value is then analyzed over the sampling time interval. By performing integral calculations, the system can calculate in real time the theoretical trajectory length traveled by the train since its starting point.

[0063] The wheel diameter wear compensation coefficient is a key dimensionless parameter used to correct errors in mechanical systems. Since trains calculate speed by measuring wheel circumference rotation, as mileage increases, wheel diameters decrease slightly due to mechanical wear, leading to an overestimation of the uncorrected speed value. This coefficient is typically preset by the onboard maintenance management system based on measured wheel diameter data from the most recent overhaul or scheduled maintenance, or automatically generated online through differential fusion of GNSS speed and axle end speed.

[0064] This represents the spatial offset caused by sensor transmission delay, and is a hard compensation term for time-space synchronization errors in high-speed operating environments. In high-speed rail transit operations, even a few milliseconds of data transmission and processing delay can lead to spatial deviations of tens of centimeters. This parameter is determined by the processor based on the system's inherent communication link delays (such as CAN bus delay and A / D conversion delay) and the train's current instantaneous speed. The product is calculated, and by subtracting this offset from the integral result, the coordinate shift caused by signal transmission lag can be compensated, ensuring that the voltage spike is accurately aligned with the drop string or support that produced it.

[0065] This step transforms the time signal, which originally changed with the train's movement, into a spatial sequence bound to the physical location, enabling the system to track long-term trends of specific structural locations.

[0066] Step 3: Perform multi-scale time-frequency decomposition on the point-bound voltage sequence to obtain high-frequency structural perturbation components and extract high-frequency structural feature vectors; perform multi-level wavelet decomposition on the point-bound voltage sequence by the processor to obtain multiple frequency band components; calculate the frequency band energy density, kurtosis value and peak interval period for each frequency band component to form a frequency band feature set; perform combination operations on the frequency band feature set to obtain the high-frequency structural feature vector; and output the high-frequency structural feature vector to the long-term feature baseline model construction step.

[0067] In this embodiment, the processor performs multi-level wavelet decomposition on the point-bound voltage sequence, decoupling the composite voltage fluctuations into different frequency band components. Preferably, the mid-to-high frequency band is selected as the structure-sensitive frequency band. The frequency band energy density, kurtosis value, and peak interval period are calculated for each frequency band component to form a frequency band feature set, and a high-frequency structure feature vector is generated through nonlinear combination operations.

[0068] To eliminate the quantitative differences in absolute voltage values ​​under different operating conditions, feature fusion is required for the extracted frequency band statistics. Based on the non-uniformity of frequency band energy distribution, a high-frequency structural feature vector reflecting the stress state of the structure is constructed. :

[0069] ;

[0070] This vector, through nonlinear coupling of energy density, signal kurtosis, and peak behavior, forms a multidimensional characterization of the health of the overhead contact system structure.

[0071] The core logic of this high-frequency structural feature vector construction formula lies in compressing complex time-frequency domain signals into digital fingerprints that can accurately characterize the physical health status of the contact network by nonlinearly combining three complementary dimensionless indices, thereby achieving feature enhancement of weak mechanical degradation signals.

[0072] in, Representing the high-frequency structural feature vector, this is the final output vector generated by the formula. As a multi-dimensional structural health characterization, it is stored in the system memory for subsequent baseline comparison and risk assessment. Each dimension of this vector locks into one aspect of the catenary's physical properties. Through this combination, the system can simultaneously identify anomalies in energy distribution, abrupt changes in impact characteristics, and deterministic patterns in random noise, thereby outlining a sub-healthy profile of the physical structure amidst complex background interference.

[0073] Representing the target frequency band energy ratio, it is the first dimension of the eigenvector, and its physical meaning lies in quantifying the energy weight of a specific structure-sensitive frequency band in the entire spectrum. During the calculation, the processor first performs a square integral on the signal of the specific frequency band (e.g., 300Hz-800Hz) after wavelet decomposition to obtain... This value is then divided by the sum of the energy in all decomposed frequency bands to obtain a relative value reflecting the degree of energy concentration. This index can sensitively detect the increase in resonant energy at a specific frequency caused by loose contact wires or excessive geometric parameters, effectively eliminating the problem of inconsistent quantitative benchmarks caused by overall voltage fluctuations in the power grid.

[0074] The nonlinear impact stability factor constitutes the second dimension of the eigenvector, which aims to identify micro-arc discharges or mechanical impacts through the statistical distribution characteristics and temporal regularity of the coupled signal. The kurtosis value of the signal is calculated by the processor using the fourth-order cumulant of the sampling points relative to the mean, and is specifically used to detect instantaneous spikes in the signal that deviate from the normal distribution; while This represents the time interval period between these impact peaks. By logarithmically compressing the periodic term and using it as the denominator, this index can produce a nonlinear numerical amplification effect when the impact intensity increases and the regularity strengthens (i.e., the period decreases), thus emitting a strong characteristic signal when the structure undergoes periodic mechanical wear or repetitive arcing.

[0075] Representing the signal-to-noise ratio gain characteristic, it is the third dimension of the feature vector. Its core function is to measure the prominence of abnormal pulses within a spatial location relative to the background environment. The processor extracts the maximum value of the voltage fluctuation amplitude within the current spatial mapping window. Compared with the average The difference between the two is calculated to remove the DC offset, and finally divided by the preset system background noise standard deviation. This calculation step is equivalent to implementing a dynamic gain regulator at the software level, ensuring that only structural disturbances that significantly exceed the random noise floor can be effectively converted into feature weights under different electromagnetic environments, greatly improving the system's identification sensitivity and anti-interference capability in complex circuit environments.

[0076] This step involves multi-scale decomposition to remove the fundamental component and random noise of the traction motor, thereby enhancing the weak features related to changes in structural stiffness and contact pressure, and thus improving the sensitivity to identify early structural degradation.

[0077] Step 4: Construct a long-term feature baseline model for corresponding fixed spatial point labels based on high-frequency structural feature vectors, and calculate the feature distance change rate between the current high-frequency structural feature vector and the long-term feature baseline model to form a risk trend indicator; The processor establishes a recursive update model with time decay characteristics to continuously update the historical high-frequency structural feature vectors of fixed spatial points; An initial baseline feature is generated based on the historical high-frequency structural feature vectors within a preset time window, and the updated baseline feature is used as the benchmark for calculating the feature distance at the next moment, and output to the risk trend indicator calculation step; The processor calculates the distance between the current high-frequency structural feature vector and the long-term feature baseline model to obtain the feature distance value; Trend analysis is performed on the continuous feature distance values ​​within a sliding time window to obtain the trend change amount; The trend change amount is output as a risk trend indicator to the spatial risk weight generation step;

[0078] To ensure the feasibility of this scheme, the feature distance can be calculated by the difference between the current high-frequency structural feature vector and the long-term feature baseline model, commonly using Euclidean or Mahalanobis distance. Subsequently, trend analysis is performed on continuous feature distance values ​​within a sliding time window to generate a risk trend indicator. For example, it can be represented as:

[0079] ;

[0080] in, Let be the high-frequency structural feature vector of the nth spatial node at time y. This represents the long-term feature baseline vector of the corresponding node; the change in feature distance over time can be used to generate risk trend indicators.

[0081] ;

[0082] This represents the mean characteristic distance within the sliding time window T. This method ensures that the risk trend indicator can sensitively reflect the evolution trend of node structural perturbations.

[0083] The preset time window can take characteristic data from k consecutive train passes through the node or characteristic data from T consecutive seconds / minutes / hours. k or T can be preset or dynamically adjusted based on the line operating frequency, environmental stability, and structural aging rate. Within the window, the processor performs a weighted average or exponential smoothing on the historical high-frequency structural feature vector to generate an initial baseline feature vector, ensuring that the initial baseline balances occasional disturbances and long-term trends. The length of the preset time window can be determined based on data from k consecutive train passes or T consecutive seconds / minutes / hours, and its setting is based on the characteristic change rate of historical operating data and environmental fluctuations, while also considering the balance between system smoothness and sensitivity. The window length can also be dynamically adjusted during system operation to adapt to changes in line operating frequency, environmental factors, or structural aging rate, thereby ensuring that the characteristic data fully reflects the operating status of the target node.

[0084] The processor calculates the difference between the high-frequency structural feature vector of the target node and the average high-frequency structural feature vector of its neighboring nodes; it counts the number of consecutive deviations of this difference within a preset time window; when the number of consecutive deviations exceeds a preset threshold, the spatial risk weight of the target node is increased and output to the power supply load prediction model; the preset threshold can be set to a number of consecutive deviations exceeding k times or a cumulative deviation exceeding a predetermined percentage, where k can be preset or dynamically adjusted based on line operating frequency, environmental fluctuations, and structural health status to ensure that the spatial risk weight adjustment can effectively reflect structural anomalies. The preset threshold can be a risk trend threshold or a spatial risk weight threshold, set based on historical deviation data statistics and system tolerance, for example, by analyzing the distribution of historical deviation counts and amplitudes to determine an appropriate threshold, balancing the sensitivity and false alarm rate of risk detection. During system operation, the threshold can be dynamically adjusted based on real-time data to adapt to changes in line operating frequency, environmental fluctuations, and structural health status.

[0085] The processor performs a weighted average calculation on risk trend indicators across multiple consecutive time windows to obtain a smoothed risk trend value. This smoothed risk trend value is used as the basis for calculating the spatial risk weight and output to the power supply load prediction model to form a closed-loop prediction and control process. In this embodiment, the power supply load prediction model is a time-series prediction model based on historical power supply data and spatial risk weights. Its inputs include the spatial risk weights of the target node and its neighboring nodes, the smoothed risk trend value, and historical power supply load data. Based on the inputs, the processor performs predictions using weighted linear regression or a Long Short-Term Memory (LSTM) network to obtain the power supply load values ​​for each node at future times. The output power supply load results are used in the closed-loop control module to trigger self-healing control strategies, such as power limiting, peak-shifting operation, or current distribution adjustment, thereby achieving a prediction and control closed loop. The model can be updated in real time through a sliding time window, allowing the prediction process to take into account both historical trends and current structural anomalies, improving the reliability and adaptability of the system operation.

[0086] In this embodiment, the processor establishes a recursive update model with time decay characteristics to continuously update the historical high-frequency structural feature vectors of fixed spatial points, thereby forming a long-term feature baseline model.

[0087] Long-term baselines need to balance the slow aging of physical structures with the smoothing of sudden fluctuations. This embodiment introduces a forgetting factor to enable the model to adapt to minor environmental changes. Based on the historical accumulation of the high-frequency structural feature vectors, a recursive update equation with time decay characteristics is constructed:

[0088]

[0089] in, This indicates the order in which the train passed this point, adjusted by the forgetting factor. This enables the dynamic fusion of historical features into the current baseline.

[0090] In the long-term characteristic baseline recursive update equation, the core logic is to use the mathematical form of exponentially weighted moving average (EWMA) to deeply integrate the historical operating patterns of specific spatial points of the overhead contact line with the current measured characteristics, thereby constructing a dynamic digital benchmark that can both reflect the slow aging process of the physical structure and effectively filter out single occasional disturbances.

[0091] in, represent The long-term feature baseline vector under the sequence is the final output variable generated by the formula. As a multi-dimensional feature reference model stored in the database, it represents the cumulative health status of the nth physical node after undergoing k train pantograph sliding "check-ups". At the system operation level, this vector is not a static dead value, but a digital "companion" that changes over time and is constantly self-correcting. It provides a unique logical benchmark for subsequently judging whether the current feature deviates from the normal trajectory.

[0092] The historical baseline representing this spatial location is the stable state value reached by the model in the previous monitoring cycle (i.e., when the (k-1)th train passed). Before executing the current update calculation, the processor needs to retrieve the historical archive data of this location from local memory or a cloud database, using it as the starting point for this recursive calculation. The existence of this parameter ensures that the system has "long-term memory" capabilities, allowing the small, trend-based physical evolutions of the overhead contact line caused by seasonal temperature differences, mechanical fatigue, etc., to be continuously recorded, rather than being treated as isolated data points.

[0093] The structural feature vector extracted in real time is the instantaneous observation calculated by wavelet decomposition and feature fusion within the current monitoring period using Formula 3. It captures the microscopic electrical characteristics of the pantograph and overhead contact line at the moment of dynamic contact when the k-th train passes the n-th node. Since a single observation is easily affected by the pantograph's wear condition or instantaneous arc fluctuations, this vector often contains a certain random component. Therefore, it cannot be directly used as a baseline, but its "contribution" must be reasonably incorporated into the long-term model through the recursive logic of this formula.

[0094] The forgetting factor, representing the exponential forgetting curve, is the key parameter in the equation responsible for adjusting the weighting of old and new information. It is typically preset within a small range of [0.01, 0.1]. This parameter determines the system's sensitivity to new data: a smaller value indicates a higher sensitivity. The value implies that the system is more inclined to trust long-term historical experience, thus greatly smoothing out occasional sensor noise or unstructured jitter; while when When moderately increased, the system can more sensitively detect abrupt changes in the physical structure (such as a component suddenly becoming loose). In practical applications, this parameter is pre-set by the algorithm architect at the software configuration layer based on the maintenance level of the circuit and the stability of the environment, thereby achieving a delicate balance between the robustness and sensitivity of the model.

[0095] Subsequently, the characteristic distance between the current high-frequency structural feature vector and the long-term characteristic baseline model is calculated, and the trend change is obtained by performing trend analysis on the continuous characteristic distance values ​​within the sliding time window.

[0096] Risk trends depend not only on the absolute degree of deviation from the baseline, but also on the acceleration of the deviation rate. Based on Mahalanobis distance and the first-order difference principle, a risk trend index is constructed. The calculation function:

[0097] ;

[0098] This indicator quantifies the deteriorating trend of the overhead contact system structure by combining the instantaneous deviation with the rate of change within a time window.

[0099] The core logic of the risk trend indicator calculation function lies in integrating the two dimensions of "instantaneous deviation state" and "time evolution rate" to improve the health assessment of the contact network physical structure from simple fault alarm to in-depth evolution trend prediction. This dual-driven assessment mode can ensure that the system not only focuses on the current performance deviation, but also has the ability to identify potential failure "acceleration".

[0100] Among them, risk trend indicators This is the final quantitative result of the formula. As a dimensionless comprehensive evaluation value, it intuitively reflects the current sub-health severity or failure risk level of a specific point in the overhead contact system. In the system execution flow, this indicator is generated in real time by the processor and output to the self-healing control module as the basis for determining whether to trigger power limiting or peak-shaving operation commands. Its numerical fluctuations directly correspond to the complete life cycle evolution of the physical structure from normal, sub-healthy to critical failure.

[0101] Current feature vector With long-term baseline vector The vector difference, i.e. This constitutes the first part of the input to the formula, and its physical meaning lies in capturing the absolute deviation displacement between the current pantograph contact mass and the historical ideal state of the area. The processor retrieves the real-time calculation results generated in step three and the baseline model generated recursively in step four from the cache, and performs a multi-dimensional subtraction operation on the two to lock the initial amplitude of the anomaly in the feature space.

[0102] Feature weight diagonal matrix This is the core parameter in the formula used to adjust the sensitivity of different feature dimensions, and it is usually stored as a preset parameter in the configuration table of the software system. Each diagonal element of this matrix corresponds to the energy ratio, kurtosis factor, and signal-to-noise ratio in the eigenvector, respectively. Its function is to differentiate the amplification of the electrical signal based on the contribution of different physical defects. For example, if experiments show that the kurtosis feature is more sensitive to the loosening of the catenary suspension wire, then... This dimension is assigned a higher weight coefficient in the matrix. These parameters are typically determined by the algorithm architect during the offline phase, after performing correlation analysis on a large number of historical failure samples.

[0103] Rate of change of feature distance within the sliding window This constitutes the second driving term in the formula, and its physical meaning lies in quantifying the "evolution speed" of structural degradation, i.e., the derivative of risk. The processor establishes a sliding time window of length m in memory, records the feature distance values ​​within the most recent m detection cycles (i.e., the most recent m trains passing the point), and calculates the slope of the distance value as a function of the train's past sequence k using least-squares fitting or a difference operator. The introduction of this parameter gives the system "foresight"; even if the current absolute deviation has not yet reached the alarm threshold, if the evolution speed shows a significant acceleration, the system can still identify the danger of rapid structural deterioration.

[0104] Static offset weighting coefficient Weighted coefficient with dynamic trend These are the decision weights responsible for balancing the importance of "current severity" and "risk of future deterioration." As global preset parameters of the system, they determine the sensitivity style of the self-healing control logic. The settings typically focus on capturing sudden physical damage, while This focuses on capturing those hidden, long-term fatigue degradations. By fine-tuning the ratio of these two coefficients during the system deployment phase, architects can ensure that the power supply system finds the optimal trigger point between ensuring operational capacity and preventing accidents, thereby achieving truly intelligent self-healing control.

[0105] Simultaneously, the feature residuals between the target node and its neighboring nodes are calculated, and the number of persistent offsets is counted for spatial consistency verification. A weighted average of the risk trend indicators over multiple consecutive time windows is performed to obtain a smoothed risk trend value.

[0106] This step, through a dual mechanism of time regression and spatial comparison, can not only identify existing structural anomalies but also perceive the trend of "deterioration," thus providing feedforward evidence for self-healing control.

[0107] Step 5: Generate spatial risk weights based on risk trend indicators and input them into a preset power supply load prediction model. Correct the predicted load value output by the power supply load prediction model and generate a self-healing control command based on the corrected predicted load value. The processor performs a weighted calculation on the basic predicted load value and the spatial risk weights to generate a corrected predicted load value. The corrected predicted load value is output to the self-healing control decision module. The processor compares the corrected predicted load value with the power supply capacity threshold. When the corrected predicted load value reaches or exceeds the power supply capacity threshold, at least one of the following is generated: a power limiting control command, a train start-up peak-shaving control command, or a backup power supply path switching command, and outputs it to the power supply control execution unit. The self-healing control decision module includes a processing unit for receiving the corrected predicted load value, a control unit for calculating the optimal control action, and an interface unit for outputting execution commands. The control unit compares the corrected predicted load value with the power supply capacity threshold through the processor. When the corrected predicted load value reaches or exceeds the power supply capacity threshold, it generates at least one of the following: a power limiting control command, a train start-up peak-shaving control command, or a backup power supply path switching command, and outputs it to the power supply control execution unit through the interface unit, thereby realizing closed-loop self-healing control of the power supply system.

[0108] To ensure operability, spatial risk weights can be calculated using risk trend indicators and spatial consistency offset analysis. For example, for the nth node:

[0109] ;

[0110] in, This is a risk trend indicator for node n at time t. This represents the number of consecutive offsets of a node within a preset time window. For the preset threshold, the function It can be implemented using weighted mapping or linear functions.

[0111] When the number of consecutive feature offsets exceeds the threshold At this time, the system automatically increases the spatial risk weight of the corresponding node to enhance the predictive impact on abnormal locations. The stable interval can be determined through historical data statistics and the experience of technical experts to ensure that the selected stable sample data truly reflects the controllable state of the environment.

[0112] In this embodiment, the processor embeds spatial risk weights as gain factors into the power supply load prediction model to perform weighted correction on the basic predicted load value.

[0113] To directly translate structural risks into constraints on power dispatch, a nonlinear mapping between physical degradation and electrical capacity needs to be established. Based on the smoothed risk trend value, a mapping equation for the predicted load correction coefficient $W_{adj}$ is constructed:

[0114] ;

[0115] Based on the corrected predicted load value The system can predict the electrical load-bearing limit of the overhead contact line at physical risk points and then generate feedforward control commands.

[0116] The core logic of the load correction weighting function is to establish a nonlinear mapping operator from "physical degradation risk" to "power dispatch constraints". By dynamically reducing the basic predicted load based on the degree of risk, it ensures that when the overhead contact line enters a sub-healthy state, it can extend the remaining life of the equipment and prevent the occurrence of discharge accidents by actively reducing the electrical load impact.

[0117] Corrected forecast load values This is the final decision variable in the formula, and it is directly passed as input data to the self-healing control decision module to determine whether the current power supply section needs to perform protection actions such as power limiting or peak-shaving start. At the system architecture level, this parameter represents the "actual maximum available power prediction value" after considering the physical safety boundary of the overhead contact line. By quantifying physical risks into electrical constraints, it realizes a fundamental shift in power supply dispatching from "blindly operating at full load" to "state-based autonomous operation".

[0118] Basic forecast load value This represents the initial power demand calculated by the system based on the current train timetable, train formation length, and preset traction characteristic curve, assuming the overhead contact system is in perfect health. This parameter is typically provided in real-time by an existing energy management system (EMS) or dispatch automation platform, reflecting the theoretical electrical energy required to maintain normal train operation. As the multiplication basis of the formula, it provides the original dimensional reference for load correction, enabling the correction logic to adaptively adjust for lines with different load intensities.

[0119] Smoothed risk trend value This is the core independent variable driving load correction. It is a physical structure risk index output from step four, processed through multi-window weighted smoothing. This parameter integrates the characteristic deviations and evolution slopes of fixed spatial locations, eliminating the interference of instantaneous sampling jitter, and providing a stable and reliable physical degradation feedback. When this value increases, it indicates that the mechanical performance of a specific point in the overhead contact line is deteriorating. Based on this, the system senses the decrease in physical load-bearing capacity, thereby triggering the subsequent logarithmic reduction logic.

[0120] Load reduction sensitivity coefficient With risk scaling factor These are two key preset tuning parameters that together determine the slope and saturation of the load correction curve. They are typically set by the power system architect at the software configuration layer based on the power supply system's safety and fault tolerance margins. This is used to adjust the sensitivity range of the risk indicator to the logarithmic function, ensuring that minor structural disturbances do not cause drastic load changes; and This defines an upper limit for the compression reduction intensity to prevent irrational excessive drops in the corrected load value. This utilizes a logarithmic function. The mapping design cleverly achieves a nonlinear control strategy of "slight correction when the risk is low and accelerated reduction when the risk is high", which achieves a precise logical balance between ensuring driving efficiency and maintaining power supply safety.

[0121] When the corrected predicted load value reaches or exceeds the power supply capacity threshold, the self-healing control decision module generates a power limiting control command, a train start-up peak-shaving control command, or a backup power supply path switching command, and sends them to the power supply control execution unit.

[0122] This closed-loop structure achieves dynamic coupling between structural risk trends and load regulation, enabling the system to adjust its operating strategy in advance when the physical structure enters a sub-healthy state, thereby reducing the risk of overload and discharge.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation methods of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for predicting and self-healing the power supply load of rail transit overhead contact lines, characterized in that, Includes the following steps: Step S1: The processor collects unfiltered high-frequency voltage signals, environmental stability indicators and train position information at the corresponding time, and filters out stable sample data that meet the preset stability range based on the environmental stability indicators. Step S2: Spatial mapping is performed between the unfiltered high-frequency voltage signal in the stable sample data and the train position information to generate fixed spatial point labels and form point-bound voltage sequences; Step S3: Perform multi-scale time-frequency decomposition on the point-bound voltage sequence to obtain high-frequency structural disturbance components and extract high-frequency structural feature vectors; Step S4: Construct a long-term feature baseline model with corresponding fixed spatial location labels based on the high-frequency structural feature vector, and calculate the feature distance change rate between the current high-frequency structural feature vector and the long-term feature baseline model to form a risk trend indicator. Step S5: Generate spatial risk weights based on risk trend indicators, input the spatial risk weights into the preset power supply load prediction model, correct the predicted load value output by the power supply load prediction model, and generate self-healing control commands based on the corrected predicted load value.

2. The method for predicting and self-healing the power supply load of rail transit overhead contact lines according to claim 1, characterized in that, The multi-scale time-frequency decomposition in step S3 includes: the processor performing multi-level wavelet decomposition on the point-bound voltage sequence to obtain multiple frequency band components; calculating the frequency band energy density, kurtosis value and peak interval period for each frequency band component to form a frequency band feature set; performing combination operations on the frequency band feature set to obtain the high-frequency structure feature vector; and outputting the high-frequency structure feature vector to the long-term feature baseline model construction step.

3. The method for predicting and self-healing the power supply load of rail transit overhead contact lines according to claim 1, characterized in that, The construction of the long-term feature baseline model in step S4 includes: establishing a recursive update model with time decay characteristics by the processor to continuously update the historical high-frequency structural feature vectors of fixed spatial points; generating initial baseline features based on the historical high-frequency structural feature vectors within a preset time window; using the updated baseline features as the benchmark for calculating the feature distance at the next moment; and outputting them to the risk trend indicator calculation step.

4. The method for predicting and self-healing the power supply load of rail transit overhead contact lines according to claim 1, characterized in that, The calculation of the risk trend index in step S4 includes: the processor calculating the distance between the current high-frequency structural feature vector and the long-term feature baseline model to obtain the feature distance value; performing trend analysis on the continuous feature distance values ​​within the sliding time window to obtain the trend change amount; and outputting the trend change amount as the risk trend index to the spatial risk weight generation step.

5. The method for predicting and self-healing the power supply load of rail transit overhead contact lines according to claim 1, characterized in that, The spatial mapping in step S2 includes: the processor establishing a spatial topology graph structure based on the train location information and the pre-stored catenary topology information, where nodes represent monitoring points and edges represent physical adjacency relationships; mapping the unfiltered high-frequency voltage signal to the corresponding nodes according to the time period of the train's passage, generating a node voltage time series, and outputting it to the multi-scale time-frequency decomposition step.

6. The method for predicting and self-healing the power supply load of rail transit catenary according to claim 1, characterized in that, Step S4 further includes a spatial consistency verification step: the processor calculates the difference between the high-frequency structural feature vector of the target node and the average high-frequency structural feature vector of its neighboring nodes; the number of continuous offsets of the difference is counted within a preset time window; when the number of continuous offsets exceeds a preset threshold, the spatial risk weight of the target node is increased and output to the power supply load prediction model.

7. The method for predicting and self-healing the power supply load of rail transit overhead contact lines according to claim 1, characterized in that, The step S5 of inputting spatial risk weights into the power supply load prediction model includes: the processor performing weighted calculations on the basic predicted load value and the spatial risk weights to generate a corrected predicted load value; and outputting the corrected predicted load value to a preset self-healing control decision module.

8. The method for predicting and self-healing the power supply load of rail transit overhead contact lines according to claim 7, characterized in that, The self-healing control decision module includes: a processor comparing the corrected predicted load value with the power supply capacity threshold; when the corrected predicted load value reaches or exceeds the power supply capacity threshold, generating at least one of the following: a power limiting control command, a train start-up peak-shaving control command, or a backup power supply path switching command, and outputting it to the power supply control execution unit.

9. The method for predicting and self-healing the power supply load of rail transit overhead contact lines according to claim 1, characterized in that, The step S1 of screening stable sample data includes: the processor makes interval judgments on wind speed, temperature and load level in the environmental stability indicators; when each indicator is simultaneously within the corresponding preset interval, the unfiltered high-frequency voltage signal of the corresponding time period is marked as stable sample data and stored in the stable sample data cache for subsequent processing.

10. The method for predicting and self-healing the power supply load of rail transit overhead contact lines according to claim 1, characterized in that, Between step S4 and step S5, there is also a risk trend smoothing step: the processor performs a weighted average calculation on the risk trend indicators of multiple consecutive time windows to obtain a smoothed risk trend value. The smoothed risk trend value is used as the basis for calculating the spatial risk weight and output to the power supply load prediction model.

Images