Multi-source cooperative traction power supply system adaptive protection method and system

Abstract

The invention discloses a multi-source collaborative traction power supply system adaptive protection method and system, multi-source data of each end of a traction power supply system is synchronously collected, and the multi-source data comprises SVG data of a power supply side, PMU data of a traction network and TCMS data of a locomotive end; performing data cleaning and space-time alignment on the multi-source data to obtain multi-source heterogeneous data; performing nonlinear feature decoupling extraction on the multi-source heterogeneous data to obtain mixed features; constructing a cooperative control model for cooperative work of the LSTM-RF hybrid model and the quantum particle swarm optimization model, and inputting the hybrid features into the cooperative control model to obtain predicted fixed values corresponding to various faults; and cooperatively controlling compensation of the SVG at the power supply side, action of the circuit breaker of the traction network and locking of the locomotive according to the predicted fixed value.

Description

Technical Field

[0001] This invention relates to the field of traction power supply technology, and specifically to an adaptive protection method and system for a multi-source coordinated traction power supply system. Background Technology

[0002] As the most crucial component of electrified railways, the reliability of the traction power supply system directly impacts railway transportation efficiency and safety. However, with the increasing complexity of the power supply network, the protection system faces severe challenges. The adaptive protection mechanism of the traction power supply system maintains the safe operation of the traction power supply network. Existing adaptive protection methods are generally based on traditional particle swarm optimization (PSO). In traditional PSO protection schemes, only the effective voltage / current values ​​of the PUM on the traction network side are used. This relies solely on traction network data without integrating locomotive-side TCMS data, making it impossible to identify the reverse current characteristics under braking conditions. This leads to false tripping during regenerative braking (such as misjudging braking energy feedback as a short-circuit fault). Furthermore, traditional schemes only aim to minimize the action time, and during execution, they only control the circuit breaker to trip, resulting in poor fault suppression and a high false tripping rate. Summary of the Invention

[0003] The purpose of this invention is to provide an adaptive protection method for a multi-source collaborative traction power supply system. This method captures the dynamic coupling characteristics of electrical, mechanical, and control systems through multi-dimensional data fusion, constructs a hierarchical decision-making system by combining hybrid intelligent algorithms, and ultimately forms a protection and control closed loop with self-learning capabilities, thereby reducing the false trip rate.

[0004] To achieve the above objectives, this application provides the following solution:

[0005] On the one hand, the present invention provides an adaptive protection method for a multi-source coordinated traction power supply system, specifically including the following steps:

[0006] S1. Synchronously collect multi-source data from each end of the traction power supply system. The multi-source data includes SVG data from the power supply side, PMU data from the traction network, and TCMS data from the locomotive end.

[0007] S2. Perform data cleaning and spatiotemporal alignment on the multi-source data to obtain multi-source heterogeneous data;

[0008] S3. Nonlinear feature decoupling and extraction are performed on multi-source heterogeneous data to obtain hybrid features;

[0009] S4. Construct a collaborative control model that combines the LSTM-RF hybrid model and the quantum particle swarm optimization model, input the hybrid features into the collaborative control model, and collaboratively output the predicted setpoints corresponding to various faults.

[0010] The Quantum Particle Swarm Optimization (QPSO) model sets up a dynamic update mechanism for the global optimal solution: every 5 minutes, the current system load rate and topology state are used as the initial population, and the protection setpoint refresh instruction is generated after 20 iterations to ensure that the setpoint is updated within 1 second after the topology changes.

[0011] S5. Based on the predicted setpoint, coordinate the control of SVG compensation on the power supply side, the operation of the traction network circuit breaker, and the locomotive interlocking.

[0012] In some specific implementation schemes, multi-source data includes:

[0013] SVG data includes multiple current harmonic components captured by the high-frequency acquisition unit on the power supply side, and uses an FIR digital filter to suppress sampling aliasing;

[0014] PMU data includes voltage waveforms sampled by the traction-side GPS synchronization device at a fixed periodic sampling frequency.

[0015] TCMS data includes converter switching frequency and duty cycle characteristics extracted from the locomotive side.

[0016] In some specific implementation schemes, the data cleaning and spatiotemporal alignment in step S2 include: establishing a three-level verification mechanism to achieve CRC verification of data integrity, threshold filtering to eliminate sensor outliers, and sliding correlation analysis to detect equipment step loss; completing time-scale normalization processing of multi-source data, performing cubic spline interpolation on asynchronous data streams, and achieving data point alignment within a preset time window.

[0017] In some specific implementation schemes, the specific process of step S3 is as follows:

[0018] S31. Wavelet packet decomposition method is used to analyze the frequency domain features of multi-source heterogeneous data;

[0019] S32. Construct a third-order differential tracker to capture the temporal transients of multi-source heterogeneous data;

[0020] S33. Based on time-domain characteristics and time-domain transients, high-frequency switching transient characteristics and low-frequency oscillation characteristics are separated from multi-source heterogeneous data;

[0021] S34. Calculate multi-dimensional entropy values ​​based on high-frequency switching transient characteristics and low-frequency oscillation characteristics;

[0022] S35. Obtain system load rate and topology status, high-frequency switching transient characteristics, low-frequency oscillation characteristics, and multi-dimensional entropy values ​​as mixed features.

[0023] In some specific implementation schemes, the specific process of step S31 is as follows:

[0024] The SVG data was decomposed using an 8-layer wavelet packet decomposition method to extract the energy proportion of 32 sub-bands; based on the energy proportion, the high-frequency band abrupt change, the mid-frequency band gradient value, and the low-frequency band energy entropy were divided.

[0025] Calculate the energy entropy of the PMU data voltage waveform to obtain the low-frequency energy entropy of the PMU;

[0026] The switching frequency of the TCMS data is mapped to the intermediate frequency band to obtain the first intermediate frequency gradient value;

[0027] A frequency domain feature matrix is ​​constructed based on the high-frequency band mutation amount, mid-frequency band gradient value, first mid-frequency gradient value, PMU low-frequency band energy entropy, and low-frequency band energy entropy.

[0028] In some specific implementations, in step S32, the multi-source heterogeneous data is input into the third-order differential tracker, and the processing procedure of the third-order differential tracker is as follows:

[0029] A first-order differential tracker detects the rising edge steepness of the PMU voltage waveform;

[0030] The second-order differential tracker locates the oscillation inflection point of the SVG current waveform;

[0031] Third-order differential tracker analysis of the distortion propagation path of TCMS duty cycle.

[0032] In some specific implementations, the high-frequency switching transient characteristics include high-frequency abrupt changes and the first derivative of the PMU voltage output by the first-order differential tracker.

[0033] Low-frequency oscillation characteristics include the low-frequency energy entropy of the PMU and the second derivative of the SVG current output by the second-order differential tracker.

[0034] In some specific implementations, in step S34, the multidimensional entropy value includes electromagnetic entropy, mechanical entropy, and energy entropy;

[0035] Calculation of electromagnetic entropy based on transient characteristics of high-frequency switching;

[0036] Calculate mechanical entropy based on low-frequency oscillation characteristics;

[0037] Calculate energy entropy based on multiple current harmonic components.

[0038] In some specific implementations, the LSTM-RF hybrid model includes an LSTM time series network, a random forest classifier, and a dynamic weight assigner. The LSTM time series network is a bidirectional network consisting of 128 memory units, and the input window of the LSTM time series network covers 20 power frequency cycles.

[0039] The random forest classifier constructed 200 decision trees and used the Gini index to rank the importance of mixed features;

[0040] The dynamic weight allocator adjusts the output weight coefficients of the LSTM-RF hybrid model based on the current load rate.

[0041] In the quantum particle swarm optimization model, the execution cycle of the QPSO algorithm is set to refresh the protection settings every 5 minutes, and the optimization dimensions include action time, equipment lifespan, and selectivity.

[0042] Secondly, this application provides a multi-source coordinated traction power supply protection system, comprising:

[0043] The multi-source data acquisition module is used to synchronously acquire multi-source data from various ends of the traction power supply system. The multi-source data includes SVG data from the power supply side, PMU data from the traction network, and TCMS data from the locomotive end.

[0044] The data processing module is used to perform data cleaning and spatiotemporal alignment on multi-source data to obtain multi-source heterogeneous data.

[0045] The nonlinear feature decoupling and extraction module is used to perform nonlinear feature decoupling and extraction on multi-source heterogeneous data to obtain hybrid features;

[0046] The dynamic parameter self-tuning module is used to construct a collaborative control model that combines the LSTM-RF hybrid model and the quantum particle swarm optimization model. The hybrid features are input into the collaborative control model to obtain the predicted values ​​corresponding to various faults.

[0047] The multi-actuator coordination module is used to coordinate the control of SVG compensation on the power supply side, the operation of traction network circuit breakers, and locomotive interlocking based on predicted setpoints.

[0048] The beneficial effects of this invention are as follows:

[0049] This invention captures the dynamic coupling characteristics of electrical, mechanical, and control systems through multi-dimensional data fusion, constructs a hierarchical decision-making system using hybrid intelligent algorithms, and ultimately forms a protection and control closed loop with self-learning capabilities. By performing spatiotemporal alignment on multimodal data, it solves the data asynchrony problem caused by the sampling rate differences of heterogeneous devices such as SVG, PMU, and TCMS. By establishing a unified time scale benchmark through sliding window interpolation technology, it ensures the time synchronization accuracy of millisecond-level events. Through multi-source data fusion (TCMS identifies the reverse characteristics of braking current) and collaborative control (SVG suppresses arc reignition), the false tripping rate under regenerative braking conditions is reduced by 93.75% compared to the traditional PSO scheme (measured value 8% → 0.5%).

[0050] Breaking through the frequency domain limitations of traditional Fourier transform, a wavelet packet decomposition and differential tracking algorithm are used to jointly analyze the transient process, achieving feature separation between high-frequency switching transients and low-frequency oscillations. A collaborative working mechanism of quantum particle swarm optimization (QPSO) and LSTM-RF hybrid model is designed to adjust the protection setpoint in real time according to the system operating status, solving the adaptability defects of fixed threshold under variable operating conditions. Attached Figure Description

[0051] Figure 1 A flowchart of a multi-source coordinated traction power supply protection method provided in an embodiment of the present invention;

[0052] Figure 2 This is a flowchart of a multi-source heterogeneous data processing method provided in an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of a multi-actuator collaborative timing control process provided in an embodiment of the present invention. Detailed Implementation

[0054] 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 some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. 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.

[0055] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0056] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0057] Furthermore, for clarity and brevity, descriptions of well-known structures, functions, and configurations may have been omitted. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of this disclosure.

[0058] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0059] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0060] Example 1

[0061] like Figure 1 As shown, this embodiment provides an adaptive protection method for a multi-source coordinated traction power supply system, specifically including the following steps:

[0062] S1. Synchronously collect multi-source data from each end of the traction power supply system. The multi-source data includes SVG data from the power supply side, PMU data from the traction network, and TCMS data from the locomotive end.

[0063] Multi-source data includes:

[0064] The SVG data includes the 0-50th current harmonic components captured by the high-frequency acquisition unit on the power supply side, and uses an FIR digital filter to suppress sampling aliasing;

[0065] PMU data includes voltage waveforms sampled by the traction-side GPS synchronization device at a fixed sampling frequency (e.g., 80 points per cycle);

[0066] TCMS data includes the 1-3kHz converter switching frequency and duty cycle characteristics extracted from the locomotive side.

[0067] S2. Perform data cleaning and spatiotemporal alignment on the multi-source data to obtain multi-source heterogeneous data;

[0068] Step S2, data cleaning and spatiotemporal alignment, includes: establishing a three-level verification mechanism to achieve CRC verification of data integrity, threshold filtering to eliminate sensor outliers, and sliding correlation analysis to detect equipment synchronization failures; completing time-scale normalization of multi-source data, performing cubic spline interpolation on asynchronous data streams, and achieving data point alignment within a preset time window (50ms). Table 1 illustrates the verification method and example scenarios of the three-level verification mechanism.

[0069] Table 1 Three-level verification mechanism

[0070]

[0071]

[0072] The purpose of multimodal data spatiotemporal alignment is to solve the data asynchrony problem caused by the difference in sampling rates of heterogeneous devices such as SVG, PMU, and TCMS. A unified time scale reference is established through sliding window interpolation technology to ensure the time synchronization accuracy of millisecond-level events. When performing multi-source data synchronous acquisition, a high-frequency acquisition unit is deployed on the power supply side to capture the current harmonic components (0-50th order) of the SVG output, and an FIR digital filter is used to suppress sampling aliasing. A GPS-synchronized PMU device is configured on the traction side, sampling the voltage waveform at 80 points per cycle with a phase measurement error of <0.01°. On the locomotive side, the PWM control signal of the TCMS system is analyzed to extract the converter switching frequency (1-3kHz) and duty cycle characteristics.

[0073] Specific example: Data alignment for regenerative braking operation conditions

[0074] Table 2 shows the raw asynchronous data (time window: xxxx year-xx month-01 12:00:00.000-12:00:00.050). It includes data on regenerative braking conditions.

[0075] equipment Timestamp Data value Sampling interval SVG 12:00:00.012 Current harmonics 52.3A 0.1ms 12:00:00.025 Data missing PMU 12:00:00.020 Voltage 28.7kV 0.25ms 12:00:00.045 Voltage 29.1kV TCMS 12:00:00.000 Duty cycle 78% 0.5ms 12:00:00.050 Duty cycle 15%

[0076] Processing procedure:

[0077] Step 1: Data Cleaning

[0078] CRC check revealed that the SVG data packet at 12:00:00.025 failed to pass verification → it was discarded and a retransmission was triggered.

[0079] Threshold filtering detection PMU 12:00:00.020 voltage value (28.7kV) is within a reasonable range (22-30kV) → retain.

[0080] Calculate the SVG-PMU correlation coefficient using the sliding correlation algorithm:

[0081]

[0082] Step 2: Spatiotemporal Alignment

[0083] Establish a baseline time axis (resolution 0.1ms).

[0084] TCMS data interpolation (original data 0.5ms interval → target 0.1ms interval)

[0085] S3. Nonlinear feature decoupling and extraction are performed on multi-source heterogeneous data to obtain hybrid features;

[0086] The data processing methods for multi-source heterogeneous data are shown in Table 3:

[0087]

[0088] Nonlinear feature decoupling extraction overcomes the frequency domain limitations of traditional Fourier transform. It employs a combined wavelet packet decomposition and differential tracking algorithm to analyze transient processes, achieving feature separation between high-frequency switching transients and low-frequency oscillations. Figure 2 As shown, the specific process is as follows:

[0089] S31. Frequency Domain Feature Analysis:

[0090] Wavelet packet decomposition was used to extract the frequency domain features from SVG data in multi-source heterogeneous data.

[0091] Specifically, the 0-50th harmonic components of SVG data in multi-source heterogeneous data are input into 8-layer wavelet packet decomposition to extract the energy proportion of 32 sub-frequency bands, divide the high / medium / low frequency band features, and obtain the power supply side harmonic distortion distribution; according to the energy proportion, the high frequency band (>5kHz) abrupt change (representing the transient process of switching devices), the mid frequency band (500Hz-5kHz) gradient value (resonance risk), and the low frequency band (<500Hz) energy entropy (grid stability) are divided; the high frequency band abrupt change is used to identify transient events such as lightning strikes, and the low frequency band energy entropy is related to the grid oscillation stability.

[0092] Specifically, the energy entropy of each harmonic is divided into high-frequency band (>5kHz) abrupt change, mid-frequency band (500Hz-5kHz) gradient value, and low-frequency band (<500Hz) energy entropy according to the energy ratio; a frequency domain feature matrix is ​​constructed based on the high-frequency band abrupt change, mid-frequency band gradient value, and low-frequency band energy entropy.

[0093] The TCMS data (switching frequency (1-3kHz)) is associated with frequency domain features. Specifically, the switching frequency is mapped to the mid-frequency gradient value of the wavelet packet to obtain the first mid-frequency gradient value and identify converter switching disturbances.

[0094] The braking condition is determined by using TCMS data (switching frequency and duty cycle). The combined criterion of duty cycle change and current reversal characteristics determines whether the condition is regenerative braking or a short circuit fault.

[0095] Energy entropy calculation is performed on the PMU voltage waveform to obtain the low-frequency energy entropy of the PMU, which characterizes the voltage oscillation intensity.

[0096] A frequency domain feature matrix is ​​constructed based on the high-frequency band mutation amount, mid-frequency band gradient value, first mid-frequency gradient value, PMU low-frequency band energy entropy, and low-frequency band energy entropy.

[0097] S32, Temporal Transient Capture

[0098] A third-order differential tracker is constructed to capture the time-domain transients of PMU data in multi-source heterogeneous data. The voltage waveform is input into the third-order differential tracker; the third-order differential tracker satisfies the dynamic threshold condition.

[0099] First-order derivative (detecting the rising edge steepness of the PMU voltage waveform): dv / dt > 10kV / μs indicates a lightning strike;

[0100] Second-order differential (locating the oscillation inflection point of the SVG current waveform): zero-crossing interval <1ms is judged as arc reignition;

[0101] Third-order differential (distortion propagation path analysis based on TCMS duty cycle): Slope change rate > 5% / ms converter breakdown, triggering tracking.

[0102] S33. Based on time-domain features and time-domain transients, high-frequency switching transient features and low-frequency oscillation features are separated from multi-source heterogeneous data. The high-frequency switching transient features include the high-frequency abrupt change of SVG harmonics and the first derivative of PMU voltage. The low-frequency oscillation features include the low-frequency energy entropy of PMU and the second derivative of SVG current. The high-frequency switching transient features are used to judge fast transient events such as arc / lightning strikes. The feature dimension is 8-dimensional. The low-frequency oscillation features are used to judge the stability fluctuation of the power grid. The feature dimension is 6-dimensional.

[0103] S34. Calculate the multidimensional entropy value:

[0104] Mechanical entropy (mechanical vibration state of the overhead contact line): The mechanical entropy is calculated based on the low-frequency oscillation characteristics to obtain the joint distribution dispersion of the overhead contact line vibration acceleration and frequency;

[0105] Electromagnetic entropy (vehicle-to-net impedance matching): Electromagnetic entropy is calculated based on the transient characteristics of high-frequency switching, and the dynamic fluctuation variance of the vehicle-to-net impedance ratio (Z_train / Z_net) is analyzed.

[0106] Energy entropy (harmonic energy distribution dispersion): Shannon entropy of the active power distribution of each harmonic;

[0107] S35. The system load rate and topology status, high-frequency switching transient characteristics, low-frequency oscillation characteristics, and multi-dimensional entropy values ​​are obtained as mixed features and input into the cooperative control model.

[0108] S4. Construct a collaborative control model that combines the LSTM-RF hybrid model and the quantum particle swarm optimization model to obtain the real-time system state (system load rate and topology state, with topology state divided into dual-power / single-power modes). Input the hybrid features, multi-dimensional entropy values, and real-time system state into the collaborative control model to obtain the predicted setpoints corresponding to various faults.

[0109] Real-time parameter optimization: In the quantum particle swarm optimization model, the execution cycle of the QPSO algorithm is set to refresh the protection settings every 5 minutes. The optimization dimensions include action time, equipment lifespan, and selectivity.

[0110] S5. Based on the predicted setpoints, coordinate the control of SVG compensation on the power supply side, the operation of the traction network circuit breaker, and locomotive interlocking. Multi-actuator coordination process:

[0111] Fast-acting circuit breaker: Equipped with a magnetic latching mechanism, the operating time is reduced to 8ms;

[0112] SVG compensation control: Under fault conditions, SVG feedforward compensation is triggered (Kp = 0.8 to suppress arc reignition); the feedforward channel predicts the compensation amount, and the feedback channel implements PID regulation (Kp = 0.8, Ki = 0.05);

[0113] Locomotive interlocking: Under braking conditions, the locomotive converter is interlocked (dual channels of hard-wired signal + GOOSE message), with dual protection through hard-wired signal (response time 50μs) and GOOSE message (transmission delay <2ms).

[0114] Specifically, the process of generating control commands based on the cooperative control model is as follows:

[0115] Step 1: Input data preparation. As shown in Table 4, the data input to the collaborative model is as follows:

[0116] Table 4 Input Data and Dimensions

[0117] source Feature type Physical meaning Dimension S33 output High-frequency switching transient characteristics Rapid transient events such as electric arcs / lightning strikes 8D Low-frequency oscillation characteristics Power grid stability fluctuations 6D S34 output Mechanical entropy Mechanical vibration state of overhead contact line 1D Electromagnetic entropy Vehicle-to-grid impedance matching 1D Energy entropy Harmonic energy distribution dispersion 1D System real-time status Load factor (30-150%) System load level 1D Topology state (0 / 1) Dual power supply / single power supply mode 1D

[0118] Step 2: Real-time fault diagnosis using the LSTM-RF hybrid model

[0119] Specifically, the LSTM-RF hybrid model includes an LSTM time series network, a random forest classifier, and a dynamic weight assigner. The LSTM time series network is a bidirectional network consisting of 128 memory units, and the input window of the LSTM time series network covers 20 power frequency cycles (400ms).

[0120] 200 decision trees were constructed in the random forest classifier to rank the importance of switch features (Gini index) and improve the accuracy of braking condition identification.

[0121] Harmonic energy entropy is used as a low-frequency oscillation characteristic input to the LSTM-RF hybrid model to predict power grid stability risks.

[0122] The time-domain transient characteristics of voltage (such as arc reignition markers) are input into an LSTM timing network to predict short-circuit faults. The time-domain transient characteristics cover a 20-cycle power frequency window.

[0123] The dynamic weight allocator adjusts the output weight coefficients of the LSTM-RF hybrid model based on the current load rate (30% to 150%).

[0124] Step 3: Quantum Particle Swarm Optimization (QPSO) Parameter Tuning

[0125] Optimization is performed every 5 minutes, including initial population, fitness calculation, quantum position update, and solving for the global optimum. QPSO optimization objectives: harmonic distortion rate is incorporated into the "equipment life" optimization dimension, SVG compensation strategy is dynamically adjusted, and QPSO is dynamically refreshed: voltage distortion rate triggers protection settings to be updated in seconds (responding within 1 second after topology change).

[0126] Step 4: Based on the fault types diagnosed in real time by the LSTM-RF hybrid model, obtain the control commands for the power supply side, traction network, and locomotive end. As shown in Table 5, the mapping between fault types and control commands for each end is given:

[0127] Table 5 Fault Types and Control Command Mapping

[0128]

[0129] like Figure 3 As shown, the multi-actuator collaborative timing control process is as follows:

[0130] The Quantum Particle Swarm Optimization (QPSO) model sends optimization setpoints to the LSTM-RF hybrid model every 5 minutes. The fault diagnosis cycle of the LSTM-RF hybrid model is 10ms. Within 10ms, the LSTM-RF hybrid model executes according to the following timing sequence:

[0131] Send a compensation command to the power supply side (SVG) (0.5ms), send an action command to the circuit breaker of the traction network (1ms), send a lockout command to the locomotive end (0.05ms), then receive compensation feedback from the SVG (2ms), receive the circuit breaker action status (5ms), and receive the execution confirmation returned by the locomotive end (0.1ms).

[0132] S6, Feedback Optimization Mechanism

[0133] A three-tiered knowledge base is established, consisting of a basic database, an incremental database, and an emergency database. The basic database stores standard fault waveforms, the incremental database records the characteristics of new faults in the field, and the emergency database pre-sets contingency plans for extreme working conditions. A model update strategy is established, which performs online learning every 24 hours. When the update increment exceeds 5%, a dynamic collaborative analysis module is triggered for verification.

[0134] Example 2

[0135] This embodiment provides a multi-source coordinated traction power supply protection system, including:

[0136] The multi-source data acquisition module is used to synchronously acquire multi-source data from various ends of the traction power supply system. The multi-source data includes SVG data from the power supply side, PMU data from the traction network, and TCMS data from the locomotive end.

[0137] The data processing module is used to perform data cleaning and spatiotemporal alignment on multi-source data to obtain multi-source heterogeneous data.

[0138] The nonlinear feature decoupling and extraction module is used to perform nonlinear feature decoupling and extraction on multi-source heterogeneous data to obtain hybrid features;

[0139] The dynamic parameter self-tuning module is used to construct a collaborative control model that combines the LSTM-RF hybrid model and the quantum particle swarm optimization model. The hybrid features are input into the collaborative control model to obtain the predicted setpoints for various faults. The dynamic parameter self-tuning mechanism is designed to work in collaboration between the quantum particle swarm optimization (QPSO) and the LSTM-RF hybrid model. The protection setpoints are adjusted in real time according to the system operating status to solve the adaptability defects of fixed thresholds under varying operating conditions.

[0140] The multi-actuator coordination module is used to coordinate the control of SVG compensation on the power supply side, the operation of traction network circuit breakers, and locomotive interlocking based on predicted setpoints.

[0141] Understandably, the data input in the Particle Swarm Optimization (PSO) adaptive protection method for traction power supply systems only uses the effective values ​​of voltage / current from the traction network PMU. The adaptive logic of the traditional scheme is: based on the load factor β = P... 实际 / P 分段 Segmented adjustment of overcurrent protection settings (divided into 3 segments): The traditional PSO algorithm optimizes the protection settings every 30 minutes, with the optimization objective being to minimize the action time; the protection action only controls the circuit breaker to trip. The technical drawbacks of this traditional method are:

[0142] 1. Single optimization dimension: Only the action time is used as the optimization target, ignoring the balance of multiple objectives such as selectivity and equipment life, resulting in overly aggressive protection settings and a high false trip rate (measured >8%).

[0143] 2. Insufficient response speed: The 30-minute optimization cycle cannot adapt to dynamic changes in the topology (such as the need for a second-level response when switching between dual power supplies);

[0144] 3. Data source limitations: The lack of integration with locomotive TCMS data makes it impossible to identify the reverse current characteristics under braking conditions, leading to false activation during regenerative braking (such as misjudging braking energy feedback as a short circuit fault).

[0145] 4. Limited execution methods: It only controls circuit breakers and lacks coordination with SVG compensation and locomotive power control, resulting in poor fault suppression (such as the inability to quickly suppress arc reignition).

[0146] Comparison of the technical advantages of this application with traditional methods

[0147]

[0148]

[0149] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An adaptive protection method for a multi-source coordinated traction power supply system, characterized in that, include: S1. Synchronously collect multi-source data from each end of the traction power supply system. The multi-source data includes SVG data from the power supply side, PMU data from the traction network, and TCMS data from the locomotive end. S2. Perform data cleaning and spatiotemporal alignment on the multi-source data to obtain multi-source heterogeneous data; S3. Nonlinear feature decoupling and extraction are performed on multi-source heterogeneous data to obtain hybrid features; S4. Construct a collaborative control model that combines the LSTM-RF hybrid model and the quantum particle swarm optimization model, input the hybrid features into the collaborative control model, and collaboratively output the predicted setpoints corresponding to various faults. S5. Based on the predicted setpoint, coordinate the control of SVG compensation on the power supply side, the operation of the traction network circuit breaker, and the locomotive interlocking.

2. The adaptive protection method for a multi-source coordinated traction power supply system according to claim 1, characterized in that, Multi-source data includes: SVG data includes multiple current harmonic components captured by the high-frequency acquisition unit on the power supply side, and uses an FIR digital filter to suppress sampling aliasing; PMU data includes voltage waveforms sampled by the traction-side GPS synchronization device at a fixed periodic sampling frequency. TCMS data includes converter switching frequency and duty cycle characteristics extracted from the locomotive side.

3. The adaptive protection method for a multi-source coordinated traction power supply system according to claim 2, characterized in that, Step S2, data cleaning and spatiotemporal alignment, includes: establishing a three-level verification mechanism to achieve CRC verification of data integrity, threshold filtering to eliminate sensor outliers, and sliding correlation analysis to detect equipment synchronization failures; completing time-scale normalization processing of multi-source data, performing cubic spline interpolation on asynchronous data streams, and achieving data point alignment within a preset time window.

4. The adaptive protection method for a multi-source coordinated traction power supply system according to claim 2, characterized in that, The specific process of step S3 is as follows: S31. Wavelet packet decomposition method is used to analyze the frequency domain features of multi-source heterogeneous data; S32. Construct a third-order differential tracker to capture the temporal transients of multi-source heterogeneous data; S33. Based on time-domain characteristics and time-domain transients, high-frequency switching transient characteristics and low-frequency oscillation characteristics are separated from multi-source heterogeneous data; S34. Calculate multi-dimensional entropy values ​​based on high-frequency switching transient characteristics and low-frequency oscillation characteristics; S35. Obtain system load rate and topology status, high-frequency switching transient characteristics, low-frequency oscillation characteristics, and multi-dimensional entropy values ​​as mixed features.

5. The adaptive protection method for a multi-source coordinated traction power supply system according to claim 4, characterized in that, The specific process of step S31 is as follows: The SVG data was decomposed using an 8-layer wavelet packet decomposition method to extract the energy proportion of 32 sub-bands; based on the energy proportion, the high-frequency band abrupt change, the mid-frequency band gradient value, and the low-frequency band energy entropy were divided. Calculate the energy entropy of the PMU data voltage waveform to obtain the low-frequency energy entropy of the PMU; The switching frequency of the TCMS data is mapped to the intermediate frequency band to obtain the first intermediate frequency gradient value; A frequency domain feature matrix is ​​constructed based on the high-frequency band mutation amount, mid-frequency band gradient value, first mid-frequency gradient value, PMU low-frequency band energy entropy, and low-frequency band energy entropy.

6. The adaptive protection method for a multi-source coordinated traction power supply system according to claim 5, characterized in that, In step S32, the multi-source heterogeneous data is input into the third-order differential tracker. The processing procedure of the third-order differential tracker is as follows: A first-order differential tracker detects the rising edge steepness of the PMU voltage waveform; The second-order differential tracker locates the oscillation inflection point of the SVG current waveform; Third-order differential tracker analysis of the distortion propagation path of TCMS duty cycle.

7. The adaptive protection method for a multi-source coordinated traction power supply system according to claim 6, characterized in that, The transient characteristics of high-frequency switching include high-frequency abrupt changes and the first derivative of the PMU voltage output by the first-order differential tracker. Low-frequency oscillation characteristics include the low-frequency energy entropy of the PMU and the second derivative of the SVG current output by the second-order differential tracker.

8. The adaptive protection method for a multi-source coordinated traction power supply system according to claim 7, characterized in that, In step S34, the multidimensional entropy values ​​include electromagnetic entropy, mechanical entropy, and energy entropy; Calculation of electromagnetic entropy based on transient characteristics of high-frequency switching; Calculate mechanical entropy based on low-frequency oscillation characteristics; Calculate energy entropy based on multiple current harmonic components.

9. The adaptive protection method for a multi-source coordinated traction power supply system according to claim 1, characterized in that, The LSTM-RF hybrid model includes an LSTM time series network, a random forest classifier, and a dynamic weight assigner. The LSTM time series network is a bidirectional network consisting of 128 memory units, and the input window of the LSTM time series network covers 20 power frequency cycles. The random forest classifier constructed 200 decision trees and used the Gini index to rank the importance of mixed features; The dynamic weight allocator adjusts the output weight coefficients of the LSTM-RF hybrid model based on the current load rate.

10. A multi-source coordinated traction power supply protection system, characterized in that, include: The multi-source data acquisition module is used to synchronously acquire multi-source data from various ends of the traction power supply system. The multi-source data includes SVG data from the power supply side, PMU data from the traction network, and TCMS data from the locomotive end. The data processing module is used to perform data cleaning and spatiotemporal alignment on multi-source data to obtain multi-source heterogeneous data. The nonlinear feature decoupling and extraction module is used to perform nonlinear feature decoupling and extraction on multi-source heterogeneous data to obtain hybrid features; The dynamic parameter self-tuning module is used to construct a collaborative control model that combines the LSTM-RF hybrid model and the quantum particle swarm optimization model. The hybrid features are input into the collaborative control model to obtain the predicted values ​​corresponding to various faults. The multi-actuator coordination module is used to coordinate the control of SVG compensation on the power supply side, the operation of traction network circuit breakers, and locomotive interlocking based on predicted setpoints.

Images