Rail transit vehicle multi-target scheduling method based on V2X architecture

Abstract

The invention discloses a rail transit vehicle multi-target scheduling method based on a V2X architecture. The method comprises the following steps: S1, obtaining a vehicle-road-cloud cooperative state; s2, scene passenger flow prediction based on multi-source data; s3, global multi-target scheduling optimization facing the prediction demand; s4, grouping and ungrouping scheduling based on prediction and real-time passenger flow; s5, operation feedback and adaptive parameter adjustment are carried out; based on a self-adaptive adjustment mechanism of an operation evaluation index, a prediction model structure or hyper-parameters in S2, a target weight in S3 and a passenger flow threshold value, a congestion degree threshold value and a safe vehicle distance coefficient in S4 are adjusted according to different deviation modes, so that all links of prediction-scheduling-marshalling are cooperatively converged in the long-term operation process, and the prediction-scheduling-marshalling efficiency is improved. And systematic performance degradation of a traditional fixed parameter model after the passenger flow structure changes is avoided. According to the method, the scheduling refinement level is improved, scene modeling of passenger flow can be realized, and the passenger flow prediction accuracy in a complex scene is improved; the method has robustness and adaptability in long-term operation.

Description

Technical Field

[0001] This invention relates to the field of rail transit technology, specifically to a multi-objective scheduling method for rail transit vehicles based on a V2X (Vehicle-to-Everything) architecture, applicable to new medium- and low-capacity urban rail transit lines with independent right-of-way, supporting driverless operation and variable formation operation. Background Technology

[0002] Medium- and low-capacity rail transit systems have lower construction costs and more flexible route layouts, making them an important component for alleviating urban congestion and improving public transportation networks. Traditional systems, which use trains as the basic operating unit and operate on fixed schedules, are gradually revealing the following problems:

[0003] 1. Insufficient accuracy of state perception

[0004] In existing systems, a significant portion of the lines still rely primarily on trackside signaling equipment and track circuit zoning for train detection and control. These detection sections are typically tens to hundreds of meters long, meaning the dispatch center can only obtain discrete location information such as whether a section is occupied, making it difficult to continuously characterize train position, speed, and acceleration. Under conditions of high-density operation, temporary train additions, or significant passenger flow fluctuations, this coarse-grained perception is insufficient to support refined scheduling of train spacing, speed curves, and train formation behavior. Safety redundancy often relies on "increasing intervals," resulting in low capacity utilization.

[0005] 2. Insufficient data utilization and limited passenger flow forecasting capabilities.

[0006] Passenger flow forecasting mainly relies on historical averages and empirical rules, lacking systematic modeling and dynamic forecasting mechanisms for multiple factors such as station level, date type, time period, weather, and large-scale events. In scenarios such as surges in passenger flow during holidays, extreme weather, or temporary events, it is difficult to reflect changes in passenger flow in a timely manner, which can easily lead to lagging capacity allocation during peak hours and low vehicle utilization during off-peak hours.

[0007] 3. The overall scheduling target is singular.

[0008] Traditional train timetables and scheduling schemes primarily rely on time-related indicators such as punctuality and minimum following intervals, with departure times, following intervals, and turnaround capacity as the main constraints. They lack a mechanism to integrate passenger travel time, energy consumption, crowding levels, and operating costs into a unified optimization framework for balancing these factors. When passenger flow fluctuates significantly or capacity is strained, it becomes difficult to achieve a dynamic balance between punctuality, energy consumption, passenger experience, and operating costs.

[0009] 4. Limited multi-mode operation and grouping scheduling capabilities.

[0010] When it is necessary to alleviate local congestion or empty runs through train formation or decoupling, the existing system mostly completes the process manually according to pre-planned schedules at the depot or turnaround station, and the train formation schemes and timetables are relatively static. The lack of a section-level train formation and decoupling mechanism that is automatically triggered based on real-time operating status and predicted passenger flow makes it difficult to achieve adaptive switching between single-car and train formation operations, and the advantages of variable train formation are not fully utilized.

[0011] 5. Lack of operational feedback and adaptive scheduling capabilities

[0012] Existing passenger flow forecasting and scheduling methods mostly employ offline modeling and fixed parameter configuration. Real-time operational data is primarily used for post-event statistical evaluation, lacking online feedback and parameter correction mechanisms. This can easily lead to model aging, mismatch between scheduling strategies and the actual operating environment, and problems such as decreased scheduling performance and difficulty in continuously optimizing energy consumption in complex scenarios.

[0013] In summary, existing technologies have not yet organically integrated high-precision state perception, scenario-based passenger flow prediction, multi-objective scheduling optimization, variable formation scheduling based on safety constraints, and operational feedback adaptive mechanisms under a unified V2X architecture to construct a closed-loop multi-objective scheduling system of "prediction-optimization-execution-feedback-re-optimization". This makes it difficult to fully unleash the comprehensive potential of new medium- and low-capacity rail transit systems in terms of safety, capacity utilization, energy consumption control, and passenger experience. Summary of the Invention

[0014] To achieve the above objectives, this invention proposes a multi-objective scheduling method for rail transit vehicles based on a V2X architecture. The steps S1 to S5 described below together constitute a closed-loop scheduling system of "state acquisition – passenger flow prediction – global optimization – local execution – feedback reconstruction". There are clear data and decision coupling relationships between each step, rather than independent functional modules.

[0015] S1 vehicle-road-cloud collaborative status acquisition;

[0016] Based on the V2X architecture, GPS / BDS antennas, RTK antennas, 5G antennas, V2V antennas, onboard readers, environmental perception sensors, inertial measurement units (IMUs), onboard units, and data processing units are installed on the vehicle side; track electronic tags, RTK base stations, and 5G base stations are deployed on the roadside. The vehicle's data processing unit fuses satellite positioning, RTK differential, IMU, and trackside tag information to periodically estimate the vehicle's high-precision position, speed, acceleration, attitude, vehicle spacing, and track occupancy status. The onboard unit transmits the above operating status to the cloud dispatching platform via 5G and other communication networks according to a preset reporting cycle. The cloud dispatching platform simultaneously accesses external environmental data (weather, temperature, line conditions, etc.) and operational data (passenger flow entering and leaving stations, passenger numbers, energy consumption, and timetable execution status, etc.) to construct a continuous, high-precision operating status map and data warehouse for the entire line, providing a unified data foundation for subsequent passenger flow forecasting and dispatching optimization.

[0017] S2 is a scenario-based passenger flow prediction system based on multi-source data;

[0018] Building upon the data warehouse constructed in S1, the cloud-based scheduling platform integrates external factors, historical operational data, and real-time operational data. It categorizes typical scenarios using combinations of "site level – date type – time period – weather / special events" and extracts corresponding passenger flow time series from the data warehouse to form scenario-based passenger flow time series. Empirical Mode Decomposition (EMD) is performed on the passenger flow time series for each typical scenario to obtain intrinsic modal components at different time scales. Then, K-Means clustering, improved using a genetic algorithm, is used to cluster the EMD features, automatically extracting passenger flow pattern labels such as "regular morning peak," "bi-peak," and "sudden event." Using the passenger flow time series, passenger flow pattern labels, and auxiliary features such as date type and weather type as input, an LSTM prediction model is constructed, outputting predicted passenger flow for each site or segment within multiple future time windows. The output of S2, the predicted passenger flow with structured scenario information such as passenger flow pattern labels, is directly input into the multi-objective scheduling model in S3 to adjust weights and constraints. When the prediction error in a typical scenario continues to exceed a preset threshold over several evaluation periods, the cloud triggers retraining or hyperparameter adjustment of the model in that scenario to ensure the effectiveness of demand characterization.

[0019] In a preferred embodiment of the present invention, the cloud scheduling platform configures the weights of the multi-objective cost function in S3 and the passenger flow threshold and congestion threshold in S4 according to the passenger flow pattern label output by S2 in a scenario-based manner: for example, under the "extreme weather mode" or "large event mode", the weights of the time target and congestion target are increased, and the weights of the energy consumption target are reduced accordingly. If necessary, the passenger flow threshold of the high-load section is reduced to increase the grouping trigger probability of the section; under the "nighttime off-peak mode", the weight of the energy consumption target is increased and the passenger flow threshold is appropriately increased to reduce unnecessary grouping behavior and empty runs.

[0020] S3 Global multi-objective scheduling optimization for predictive demand;

[0021] Based on the future passenger flow forecasts for each station or section obtained from S2, the cloud-based dispatching platform, taking into account constraints such as line topology, number of vehicles and marshalling capacity, platform and turnaround capacity, and safety intervals, needs to provide a unified evaluation standard for different possible dispatching schemes within the same forecast time domain. This standard is used to compare the merits of different schemes and thus output the optimal dispatching scheme.

[0022] Therefore, a multi-objective cost function J is constructed, considering the following types of indices: Time-related indicators such as average passenger waiting time, travel time, and delays; Energy consumption indicators such as traction energy consumption and braking energy recovery; The average level, peak level, and duration of crowding in the carriages; Cost indicators such as operating cost per passenger kilometer and vehicle turnover efficiency.

[0023] The overall objective can be written as:

[0024]

[0025] in, These are adjustable weights used to reflect which aspects of the current passenger flow pattern are prioritized: punctuality and efficiency, energy consumption control, passenger comfort, or operating costs. The initial weight values ​​are set by the operational strategy and can be adaptively adjusted in S5 based on operational feedback.

[0026] Under the aforementioned cost function and related constraints, using the spatiotemporal predicted passenger flow given by S2 as input, optimization algorithms such as ant colony optimization are employed to search for decision variables such as departure frequency, minimum following interval, vehicle route, and stop schedule within the future prediction time domain. During the iteration process, the algorithm continuously generates candidate scheduling schemes, evaluates each scheme using the cost function J, and gradually retains the scheme with better overall performance, ultimately obtaining a set of scheduling results that achieve a good trade-off between "time, energy consumption, congestion, and cost." This scheduling result constitutes a global baseline operation map and vehicle utilization scheme oriented towards predicted demand, specifically including: departure frequency and minimum following interval for each route and section within different time windows; route, departure time, and stop schedule for each vehicle; and the stations or sections and time windows within which grouping or ungrouping can occur, i.e., candidate grouping / ungrouping windows and their priorities.

[0027] S4 is a grouping and degrouping scheduling system based on predicted and real-time passenger flow.

[0028] In the baseline operation plan provided by S3, candidate stations and time windows for performing train formation / decoupling have been identified. This step involves performing a local adaptive assessment of each candidate window during actual train operation to determine whether to actually execute the formation or decoupling operation.

[0029] Within each candidate grouping / disgrouping window, the cloud-based scheduling platform comprehensively considers the following conditions:

[0030] Predicted passenger flow conditions: Based on the prediction results of S2 for the current and several subsequent time windows, determine whether the passenger flow in this section exceeds the preset passenger flow threshold (corresponding to grouping) or remains below the threshold for a period of time (corresponding to ungrouping).

[0031] Real-time congestion condition: Based on the real-time passenger information of the vehicle, determine whether the current and short-term congestion of the candidate vehicle is close to or exceeds (or significantly lower than) the preset congestion threshold.

[0032] Safety conditions: Calculate the safe following distance based on the vehicle speed and following distance provided by S1. (For example (Only when the actual vehicle spacing is not less than) Furthermore, if the signal system allows the relevant operations to be performed, the safety conditions are considered met.

[0033] When all the above conditions are met, the system confirms the execution of grouping or ungrouping operations within the candidate window, and issues control commands to the relevant vehicles via V2X to achieve speed synchronization between vehicles, coupler engagement or unlocking, and operation mode switching; if the key conditions are not met, the vehicles continue to execute according to the single-vehicle operation plan / existing grouping plan given by S3.

[0034] Therefore, within the global baseline operation diagram framework provided by S3, S4 utilizes three types of information—"predicted passenger flow, real-time congestion, and speed-related safe train spacing"—to make local decisions on whether to execute each candidate train formation / ungrouping window, thereby achieving adaptive adjustment of transport capacity at the section level without disrupting the overall structure of the global operation diagram.

[0035] S5 runtime feedback and adaptive parameter adjustment;

[0036] After executing the scheduling schemes determined in S3 and S4, the cloud-based scheduling platform statistically analyzes operational indicators such as average waiting time, segment congestion, energy consumption per passenger kilometer, punctuality rate, number of grouping / degrouping operations, and failure rate according to a preset evaluation cycle, and compares these indicators with their respective target ranges. When an indicator deviates continuously from the target range for multiple consecutive evaluation cycles, the system adjusts it according to preset rules.

[0037] 1. Weights of the multi-objective cost function in S3 ;

[0038] 2. Passenger flow threshold, congestion threshold, and safe vehicle spacing coefficient in S4 ;

[0039] 3. Some structural or hyperparameter aspects of the passenger flow forecasting model in S2;

[0040] Through the aforementioned adaptive adjustment mechanism based on operational evaluation indicators, this invention does not merely modify parameters for a single indicator. Instead, it adjusts the prediction model structure or hyperparameters of S2, the target weights of S3, and the passenger flow threshold, congestion threshold, and safe train spacing coefficient of S4 according to different deviation patterns. This ensures that each link in the "prediction-scheduling-grouping" process converges collaboratively during long-term operation, avoiding the systematic performance degradation of traditional fixed-parameter models after changes in passenger flow structure. This forms a closed-loop chain of "collection (S1) – prediction (S2) – global scheduling (S3) – local adaptive execution (S4) – operational feedback and reconstruction (S5)," constituting the holistic technical solution of this invention that distinguishes it from existing technologies.

[0041] Compared with the prior art, the present invention has at least the following beneficial effects:

[0042] 1. Achieve centimeter-level, high-frequency train operation status perception, improve the level of dispatching precision, and enhance the response capability to local congestion and speed fluctuations;

[0043] 2. By constructing passenger flow pattern labels for passenger flow at different time scales, we can achieve scenario-based modeling of passenger flow and improve the accuracy of passenger flow prediction in complex scenarios;

[0044] 3. Construct a multi-objective cost function that simultaneously incorporates passenger time, energy consumption, congestion, and operating costs, and improve the overall performance of the system scheduling strategy under a unified optimization framework;

[0045] 4. Based on V2X status information, passenger flow, and safety constraints, construct a safe train formation and ungrouping trigger mechanism to improve the capacity utilization rate of a unit section and reduce empty runs while ensuring driving safety;

[0046] 5. A closed-loop mechanism of operational feedback and adaptive parameter adjustment is introduced to continuously correct the passenger flow prediction model and scheduling parameter settings, thereby improving the robustness and adaptability of the method in long-term operation. Attached Figure Description

[0047] Figure 1 It is a new V2X architecture for rail transit systems.

[0048] The diagram is labeled as follows: 1-GPS / BDS antenna, 2-5G antenna, 3-V2V antenna, 4-RTK antenna, 5-Vehicle reader, 6-Inertial measurement unit, 7-Vehicle unit, 8-Data processing unit, 9-Ultrasonic radar, 10-Vision sensor, 11-LiDAR, 12-Millimeter-wave radar, 13-Rail electronic tag, 14-Maintenance personnel electronic tag, 15-GPS / BDS satellite, 16-5G base station, 17-RTK base station, 18-Third-party RTK supplier cloud, 19-Coupled, 20-Rail. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0050] Example: A novel multi-objective scheduling method for rail transit vehicles based on V2X architecture

[0051] S1: V2X architecture and operational status acquisition;

[0052] like Figure 1As shown, various sensing and communication devices are installed on the vehicle side, and state fusion and control execution are achieved through a data processing unit and an on-board unit. Specifically, a GPS / BDS antenna 1 is installed on the vehicle for communicating with the satellite system to obtain the vehicle's absolute position information; an RTK antenna 4 is installed to obtain centimeter-level high-precision positioning information via a trackside RTK base station 17 and a third-party RTK supplier's cloud platform 18; a 5G antenna 2 is installed to communicate with a 5G base station 16 to achieve bidirectional high-speed data transmission between the vehicle and the cloud dispatching platform; and a V2V antenna 3 is installed for direct communication with neighboring vehicles to exchange status information such as relative position, speed, and operating mode. The antennas are preferably arranged in areas with minimal signal obstruction on the vehicle's outer surface, such as the roof or the area near the roof, but the specific installation location can be adjusted according to the vehicle's structural layout and electromagnetic compatibility requirements, and this invention does not limit this. The vehicle is also equipped with an onboard reader 5 to read the track electronic tags 13 laid along the track and the maintenance personnel electronic tags 14 carried by maintenance personnel to obtain section numbers, station information, speed limit information, and maintenance status; an inertial measurement unit 6 is installed to collect information such as vehicle acceleration and angular velocity, and fuse it with satellite positioning information to improve the accuracy of attitude and position estimation; environmental perception sensors, including a visual sensor 10, a lidar 11, a millimeter-wave radar 12, and an ultrasonic radar 9, are installed to perceive obstacles and environmental conditions in front and around the vehicle; a data processing unit 8 is installed to receive and fuse data from the GPS / BDS antenna 1, RTK antenna 4, inertial measurement unit 6, onboard reader 5, V2V antenna 3, and environmental perception sensors, and output state quantities such as vehicle position, speed, attitude, vehicle spacing, and track occupancy in the track coordinate system; and an onboard unit 7 is installed to receive cloud scheduling instructions and perform control on vehicle driving, braking, steering, and grouping and ungrouping according to the scheduling results.

[0053] Track electronic tags 13 are installed along the trackside to record track number, section attributes, gradient curve, speed limit information, etc.; RTK base stations 17 are set up to provide high-precision positioning references, and 5G base stations 16 are set up to provide mobile communication coverage, providing basic support for the above-mentioned vehicle-side communication and positioning equipment.

[0054] During operation, the data processing unit 8 collects raw data from various onboard sensors and communication modules at a preset sampling period. The sampling period for position and attitude information is preferably set to 50–100 ms. At typical operating speeds (e.g., 40–80 km / h), the vehicle displacement between adjacent sampling points is approximately 0.5–2 m, which can accurately depict speed fluctuations of the train when approaching a platform, implementing temporary speed limits, and operating in sections with other trains. The sampling period for environmental perception sensors is preferably set to 50–200 ms, ensuring timely obstacle detection and environmental perception while avoiding excessive load on onboard processing capabilities and communication links. After filtering, coordinate transformation, and multi-source fusion, the data processing unit 8 generates a vehicle state vector (including vehicle identification, timestamp, position, speed, acceleration, attitude, train spacing, and track occupancy information), and uploads it to the cloud scheduling platform via the 5G antenna 2 at a relatively long interval. The state reporting period is preferably approximately 1 second, ensuring near real-time monitoring of the entire line's operating status on the cloud and supporting multi-objective scheduling optimization while keeping uplink communication bandwidth usage and cloud computing overhead within an acceptable range. The dispatching platform aggregates the status of all vehicles and trackside equipment along the entire line to form a global operational status, providing fundamental data for subsequent passenger flow forecasting, multi-objective scheduling optimization, and train formation and decoupling decisions. For different line lengths, vehicle densities, and network conditions, the aforementioned sampling and communication cycles can be appropriately adjusted within the specified range, but should adhere to the design principle that "the sampling interval is sufficient to characterize changes in operational status while the communication load does not exceed the system's processing capacity."

[0055] S2: Passenger flow prediction based on multi-source data;

[0056] Based on the V2X architecture described in Example 1, this embodiment constructs a passenger flow prediction model based on multi-source data to predict passenger flow demand at each station or section within a given time window.

[0057] The multi-source data includes:

[0058] 1. External factor data: date type (weekday, weekend, holiday), specific date, weather type (sunny, rainy, snowy, etc.), temperature range, and information on major events.

[0059] 2. Historical operational data: Passenger flow at each station, actual passenger capacity of vehicles, departure times, segment travel time, segment energy consumption, etc.

[0060] 3. Real-time operational data: Current passenger flow entering and exiting each station, real-time passenger load of vehicles, current departure intervals and operating status, etc.

[0061] First, the cloud platform categorizes stations into hub stations, large stations, and ordinary stations based on their size, function, and historical passenger flow. Then, it divides the multi-source data into several typical scenarios according to a combination of "station level—date type—time period—weather / special events," such as "hub station—weekday—morning rush hour—music festival" and "ordinary station—weekend—off-peak—rainy day." For each typical scenario, a passenger flow time series dataset is constructed, and missing value imputation, outlier removal, and normalization are performed to obtain high-quality modeling samples.

[0062] Preferably, this embodiment uses a combined model of Empirical Mode Decomposition-Genetic Algorithm Improved K-Means-Long Short-Term Memory Network (EMD-GA-KMeans-LSTM) to achieve passenger flow prediction. The specific steps include:

[0063] 1. Empirical Mode Decomposition (EMD): EMD decomposition is performed on passenger flow time series under various typical scenarios to break down the original passenger flow series into fluctuation components and long-term trends at different time scales, so as to reduce the impact of series non-stationarity on subsequent model training.

[0064] 2. Improved K-Means Clustering using Genetic Algorithm (GA-KMeans): This method clusters passenger flow features or components after EMD decomposition to obtain different passenger flow pattern labels, which are used to identify different types of passenger flow patterns such as "normal workday pattern," "extreme weather pattern," and "large-scale event pattern." These passenger flow pattern labels are input as additional features into the subsequent LSTM prediction model, enabling the model to distinguish different types of morning peak patterns when making short-term predictions, thereby improving prediction stability under abnormal conditions.

[0065] In a typical implementation, the number of clusters k is searched within the range of 3 to 10, which balances the ability to distinguish different passenger flow patterns with the statistical stability of the number of samples per cluster. The population size and the number of iterations are selected within the ranges of 20 to 200 and 50 to 300, respectively, to achieve a trade-off between search sufficiency and training time. Simulations show that when specific parameters are selected within the above range, the changes in indicators such as MAE and RMSE tend to level off, indicating that further increasing the parameters brings limited benefits.

[0066] 3. LSTM Sequence Prediction: Using passenger flow time series, passenger flow pattern labels, and auxiliary features such as date type and weather type as input, an LSTM prediction model is constructed to predict passenger flow within future time windows. Specifically, based on a comprehensive analysis of "recent passenger flow trends," "current passenger flow patterns," and external conditions, the model outputs predicted passenger flow for each future time period.

[0067] In a typical configuration, the historical time step of the LSTM input can be set to 10–60 time steps to strike a balance between "avoiding insufficient trend characterization due to too short historical information" and "increasing computational overhead and introducing too much long-term noise due to too long historical information." The prediction time domain can cover the next 30–120 minutes to accommodate short-term scheduling response needs and control prediction uncertainty. The prediction results are output in 5–15 minute time windows for predicted passenger flow at each station or section, matching the time granularity with the actual departure intervals and capacity adjustment granularity. To evaluate prediction performance, metrics such as mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) can be compared with actual passenger flow. When the prediction error continuously exceeds a preset threshold within a certain period of time in a given scenario, the cloud platform can automatically trigger model retraining or parameter adjustment to improve prediction accuracy and robustness.

[0068] The above-described model structure, parameter range, and evaluation indicators are merely optional implementations of this invention. Those skilled in the art can select other prediction models or adjust parameter configurations based on factors such as line scale, data volume, and computing resources.

[0069] S3: Multi-objective scheduling optimization;

[0070] This embodiment, based on Embodiments 1 and 2, incorporates passenger flow forecasting results to achieve multi-objective scheduling optimization of the system throughout the year. A multi-objective scheduling cost function is constructed, for example:

[0071]

[0072] in:

[0073] Total passenger travel time or total vehicle travel time;

[0074] Total energy consumption of vehicles;

[0075] Average congestion index for a line or section;

[0076] Operating costs;

[0077] The weight coefficients corresponding to each objective satisfy the following conditions:

[0078]

[0079] The weighting coefficients can be flexibly configured under different operational strategies. For example:

[0080] During the morning rush hour, time can be prioritized. Reduce cost weight This allows scheduling to prioritize shortening passenger travel time; and under conditions of energy conservation or energy shortage, the energy consumption weight can be increased. This guides scheduling results to avoid high-energy-consuming operating modes; under conditions where comfort is prioritized, the weight of congestion can be appropriately increased. Restrict high-congestion operation.

[0081] In one specific embodiment, the weighting coefficient can be set as follows, for example: Within the range of 0.3 to 0.6, Within the range of 0.1 to 0.4, Within the range of 0.1 to 0.3, The weight range is between 0.1 and 0.3. The principle for setting this weight range is as follows: In low- to medium-capacity urban rail transit scenarios, passenger travel time is usually considered the primary indicator. Therefore, its upper limit is set higher than other objectives to ensure that the time objective maintains a certain dominant position in comprehensive optimization, while reserving sufficient range for energy-saving and comfort-priority operating conditions. The weights for energy consumption, congestion, and cost are set at 0.1–0.4 or 0.1–0.3 to avoid completely ignoring any non-time objective, and to prevent excessive weights from causing individual objectives to suppress others and leading to an imbalance in the scheduling scheme. The specific values ​​can be set by the operation management department based on actual needs and operational strategies.

[0082] To solve the aforementioned multi-objective scheduling problem, this embodiment preferably employs swarm intelligence optimization algorithms such as ant colony optimization. The algorithm constructs a heuristic function that comprehensively considers running time, energy consumption, congestion, and cost, and searches for optimal running paths and departure plans for each vehicle on the track topology map using a pheromone update mechanism.

[0083] In a typical configuration, the pheromone evaporation coefficient of the ant colony algorithm can be set in the range of 0.1 to 0.9. A smaller evaporation coefficient is beneficial for preserving historically good paths, but too small a coefficient can lead the algorithm into local optima; a larger evaporation coefficient is beneficial for exploring new paths, but too large a coefficient can make the search lack memory. The above range can cover various configurations from "strong memory" to "strong exploration". The heuristic factor can be set in the range of 1 to 5 to control the influence of heuristic information (such as running time and energy consumption), preventing its weight from being too low, leading to an approximately random search, or too high, leading to premature convergence. The number of ants can be selected from 10 to 200, and the maximum number of iterations can be selected from 50 to 500. Based on the experimental phenomenon that "the solution accuracy gradually decreases with increasing scale", the above range can achieve a relatively balanced solution time and scheduling quality for most line scales. The algorithm solution results include the running path to be adopted by each vehicle in the prediction time domain, the departure time, and the speed control strategy in key sections.

[0084] S4: Grouping and ungrouping scheduling based on prediction and real-time passenger flow;

[0085] After obtaining the multi-objective scheduling results, this embodiment further combines passenger flow prediction and safety constraints to realize vehicle grouping and ungrouping scheduling.

[0086] For a certain section, during the predicted period, when the predicted passenger flow exceeds the preset passenger flow threshold... When the predicted passenger flow is consistently lower than the target, the section is considered a high-load section; when the predicted passenger flow is consistently lower than the target, the section is considered a high-load section. At that time, the section is considered a low-load section. Passenger flow threshold. It can be set according to the line design capacity, train capacity and service level requirements.

[0087] Based on the real-time passenger load of the vehicle and the vehicle's maximum passenger capacity Calculate vehicle congestion And set a congestion threshold. This serves as another condition for grouping and ungrouping decisions. For example, it represents the maximum acceptable level of crowding for passengers. Taking into account vehicle capacity, passenger comfort, and safe evacuation capacity, the range is determined to be 0.6 to 0.9. When the value is below 0.6, there is obvious redundancy in the vehicle, and the marginal benefit of continuing to form trains is small; when the value is above 0.9, the vehicle is too crowded, which is not conducive to improving passenger experience and emergency evacuation capacity.

[0088] To ensure the safety of train formation operation, this embodiment adopts a speed-based safe train spacing criterion. The safe train spacing can be expressed as:

[0089]

[0090] Where v is the vehicle's current speed. The safety distance coefficient reflects the requirements for safe distance based on vehicle braking performance, track conditions, and perceived distance. In a typical configuration, The time interval can be selected within the range of 1.0 to 4.0 seconds, taking into account the braking performance and sensing distance of typical rail vehicles: in straight sections with good visibility, a time interval of 1.0 to 2.0 seconds is sufficient to ensure the safe stopping of the vehicle in emergency braking situations; while in curved sections, sections with large gradients or limited visibility, selecting a larger value of 2.5 to 4.0 seconds can improve the tolerance for braking distance estimation errors and sensing lag.

[0091] The cloud-based scheduling platform generates a grouping scheduling instruction when the following conditions are met simultaneously:

[0092] 1. The predicted passenger flow of the section to be traversed during the corresponding time period is greater than the passenger flow threshold. ;

[0093] 2. The congestion level of candidate vehicles or existing train sets is close to or exceeds the congestion threshold. ;

[0094] 3. The real-time vehicle spacing between candidate vehicles is greater than the safe vehicle spacing at the corresponding speed. It also has the capability to perform marshalling operations at designated stations or sections.

[0095] By combining three types of information—predicted passenger flow, real-time congestion level, and speed-related safe train spacing—as the trigger conditions for train formation and disbanding, this embodiment can achieve increased capacity in high-load sections and energy consumption control in low-load sections while ensuring driving safety. Simulation results show that, compared with a fixed train formation scheme that does not differentiate between section load levels, this scheme can reduce the average passenger waiting time during peak hours and reduce energy consumption per passenger kilometer during off-peak hours.

[0096] Formation commands are issued to relevant vehicles via V2X communication. Vehicles physically form formations at designated stations or sections, and then operate in formations in high-load sections to improve the capacity utilization of a single trip. When predicted passenger flow decreases and congestion gradually drops below the congestion threshold, the dispatching platform generates ungrouping commands, arranging for vehicles to ungroup and disband at suitable stations or sections, resuming single-vehicle operation and reducing empty mileage and energy consumption. During formation and ungrouping, V2V communication is used for state synchronization and coordinated control between formed vehicles; V2I communication is used for information exchange and safety verification between vehicles and trackside equipment, ensuring that formation actions are performed while meeting signal conditions and safe clearance requirements.

[0097] S5: Operational feedback and adaptive parameter adjustment;

[0098] This embodiment provides a specific implementation method for adaptively adjusting the passenger flow prediction model, scheduling optimization parameters, and grouping and ungrouping trigger thresholds based on operational feedback, which is used to further illustrate the closed-loop scheduling mechanism of the present invention.

[0099] Based on Examples 1 to 4, the cloud-based scheduling platform sets an operation evaluation cycle, for example, 15 minutes, 30 minutes or 1 hour as an evaluation cycle. The operation data within each evaluation cycle is summarized, and the following operation evaluation indicators are calculated: 1. Average passenger waiting time, which can be obtained by statistical analysis based on the deviation between the actual departure time and the planned departure time of each station, as well as the passenger flow in and out of the station.

[0100] 2. Section congestion can be calculated as the ratio of the number of passengers in the carriages to the maximum designed passenger capacity, and the average value or 95th percentile of each section within the evaluation period can be statistically analyzed; 3. Energy consumption per passenger kilometer can be calculated as the ratio of the traction energy consumption of the section to the passenger turnover (passenger kilometers) during the same period; 4. On-time departure rate is the proportion of the number of trains that depart on schedule within the allowable error range (e.g., ±1 min or ±2 min) to the total number of trains; 5. Number of formation and de-formation events can be statistically analyzed as the total number of formation and de-formation events occurring in each formation section within the evaluation period, as well as the corresponding failure rate.

[0101] The target range can be set by the operator based on service level and safety standards, such as an average waiting time of 5 to 8 minutes, a congestion level of 0.6 to 0.85, and an on-time rate of no less than 95%.

[0102] When a certain operational evaluation indicator is detected to deviate continuously from the corresponding target range for M consecutive evaluation periods (e.g., M = 3 to 6), the cloud-based scheduling platform will trigger corresponding adaptive adjustment strategies: 1. When the average waiting time or on-time rate continues to deteriorate and the segment congestion does not increase significantly, the weight coefficient of the passenger travel time objective in the multi-objective scheduling cost function will be increased, and the weight coefficient of the energy consumption or cost objective will be appropriately reduced if necessary, so as to encourage subsequent scheduling solutions to be more aggressive in the time dimension; 2. When the energy consumption per passenger kilometer is continuously higher than the target range while the passenger waiting time and congestion are within an acceptable range, the weight coefficient of the energy consumption objective will be appropriately increased, or the passenger flow threshold and congestion threshold of some segments will be adjusted upwards, in order to reduce unnecessary train formation behavior and empty runs;

[0103] 3. When the congestion level of a section continues to approach or even exceed the upper limit in multiple evaluation periods, the passenger flow threshold or congestion threshold of the corresponding section can be appropriately reduced under the premise of maintaining safety constraints, so as to increase the grouping probability of high-load sections, or the passenger flow prediction model can be retrained to improve the prediction accuracy of peak scenarios.

[0104] 4. When the failure rate of grouping and ungrouping or the number of abnormal events increases, the safety spacing coefficient τ can be appropriately increased or the restrictions on speed and position deviations in the grouping trigger logic can be adjusted to improve the safety margin of the grouping and ungrouping process.

[0105] During the aforementioned adaptive adjustment process, the cloud-based scheduling platform can employ a combination of batch offline updates and online fine-tuning: for the passenger flow prediction model, offline retraining can be performed once after the end of each day's operations based on the accumulated data of that day; for operational parameters such as multi-objective weight coefficients, passenger flow thresholds, and congestion thresholds, minor online adjustments can be made at the end of each evaluation cycle based on the degree of indicator deviation. The adjusted model and parameters take effect in the next evaluation cycle to achieve continuous optimization of operational performance.

Claims

1. A multi-objective scheduling method for rail transit vehicles based on V2X architecture, characterized in that, The following arrangements are included: S1 vehicle-road-cloud collaborative status acquisition; Based on the V2X architecture, GPS / BDS antennas, RTK antennas, 5G antennas, V2V antennas, on-board readers, environmental perception sensors, inertial measurement units (IMUs), on-board units, and data processing units are installed on the vehicle side; track electronic tags, RTK base stations, and 5G base station equipment are deployed on the roadside; the vehicle's data processing unit fuses satellite positioning, RTK differential, IMU, and trackside tag information to periodically estimate the vehicle's high-precision position, speed, acceleration, attitude, vehicle spacing, and track occupancy status; the on-board unit transmits the above operating status to the cloud dispatching platform via the 5G communication network according to a preset reporting cycle; the cloud dispatching platform simultaneously accesses external environmental data and operational data to construct a continuous, high-precision operating status map and data warehouse for the entire line, providing a unified data foundation for subsequent passenger flow prediction and dispatch optimization; S2 is a scenario-based passenger flow prediction system based on multi-source data; Based on the data warehouse built in S1, the cloud-based scheduling platform integrates external factors, historical operational data, and real-time operational data. It categorizes typical scenarios by combination of station level, date type, time period, and weather / special events, and extracts corresponding passenger flow time series from the data warehouse to form scenario-based passenger flow time series. Empirical Mode Decomposition (EMD) is performed on the passenger flow time series under each typical scenario to obtain intrinsic modal components at different time scales. Then, K-Means clustering, improved using a genetic algorithm, is used to cluster the EMD features and automatically extract passenger flow pattern labels. The passenger flow time series, passenger flow pattern labels, and auxiliary features such as date type and weather type are used as input to construct an LSTM prediction model, outputting predicted passenger flow for each station or segment within multiple future time windows. The output of S2 is the predicted passenger flow with structured scenario information containing passenger flow pattern labels, which is directly input into the multi-objective scheduling model in S3 to adjust weights and constraints. When the prediction error under a certain typical scenario continuously exceeds a preset threshold within several evaluation periods, the cloud triggers retraining or hyperparameter adjustment of the model under that scenario to ensure the effectiveness of demand characterization. S3 Global multi-objective scheduling optimization for predictive demand; Based on the future passenger flow forecasts for each station or section obtained from S2, the cloud-based dispatching platform combines the line topology, number of vehicles and marshalling capacity, platform and turnaround capacity, and safety interval constraints. For different dispatching schemes that may exist within the same forecast time domain, a unified evaluation standard is first given to compare the merits of the schemes, thereby outputting the best dispatching scheme. S4 is a grouping and degrouping scheduling system based on predicted and real-time passenger flow. In the baseline operation plan given by S3, candidate stations and time windows for performing train formation / decoupling have been marked. In this step, during the actual operation of the train, a local adaptive judgment is made for each candidate window to determine whether to actually perform the formation or decoupling operation. S5 Operation Feedback and Adaptive Parameter Adjustment; After executing the scheduling schemes determined by S3 and S4, the cloud-based scheduling platform will statistically analyze the average waiting time, segment congestion, energy consumption per passenger kilometer, punctuality rate, number of grouping / degrouping operations, and failure rate according to the preset evaluation cycle, and compare them with their respective target intervals. When a certain indicator deviates from the target range continuously over multiple consecutive evaluation periods, adjustments are made according to preset rules. Based on the passenger flow pattern labels output by S2, the cloud-based scheduling platform configures the weights of the multi-objective cost function in S3 and the passenger flow thresholds and congestion thresholds in S4 in a scenario-based manner. In the nighttime off-peak mode, the weight of the energy consumption target is increased and the passenger flow threshold is appropriately increased to reduce unnecessary train formation behavior and empty runs.

2. The multi-objective scheduling method for rail transit vehicles based on V2X architecture according to claim 1, characterized in that, In S3, a multi-objective cost function J is constructed, considering the following types of metrics: Time-related indicators such as average passenger waiting time, travel time, and delay status; Indicators related to traction energy consumption and braking energy recovery; The average level, peak level, and duration of crowding in the carriages; Operating cost per passenger kilometer and vehicle turnover efficiency cost indicators; The overall objective is written as: ； in, Adjustable weights; Under the aforementioned cost function and related constraints, using the spatiotemporal predicted passenger flow given by S2 as input, optimization algorithms such as ant colony optimization are employed to search for decision variables related to departure frequency, minimum following interval, vehicle operating path, and stop plan within the future prediction time domain. During the iteration process, candidate scheduling schemes are continuously generated, and each scheme is evaluated using the cost function J. Schemes with superior overall performance are gradually retained, ultimately yielding a set of scheduling results that balance time, energy consumption, congestion, and cost. This scheduling result constitutes a global baseline operating map and vehicle utilization scheme oriented towards predicted demand, including: departure frequency and minimum following interval for each line and section within different time windows; operating path, departure time, and stop plan for each vehicle; and the stations or sections and time windows in which grouping or ungrouping is performed, i.e., candidate grouping / ungrouping windows and their priorities.

3. The multi-objective scheduling method for rail transit vehicles based on V2X architecture according to claim 1, characterized in that, In S4, within each candidate grouping / disgrouping window, the cloud scheduling platform comprehensively considers the following conditions: Predicted passenger flow conditions: Based on the prediction results of S2 for the current and several subsequent time windows, determine whether the passenger flow in this section exceeds the preset passenger flow threshold or remains below the threshold for a period of time. Real-time congestion conditions: Based on real-time passenger information of vehicles, determine whether the current and short-term congestion of candidate vehicles is close to or exceeds the preset congestion threshold. Safety conditions: Calculate the safe following distance based on the vehicle speed and following distance provided by S1. When the actual vehicle spacing is not less than Furthermore, if the signal system allows the relevant operations to be performed, the safety conditions are considered met; When all the above conditions are met, the system confirms the execution of the grouping or ungrouping operation within the candidate window, and issues control commands to the relevant vehicles through V2X to achieve speed synchronization between vehicles, coupler engagement or unlocking, and operation mode switching; if the key conditions are not met, the vehicles continue to execute the single-vehicle operation plan / existing grouping plan given by S3.

4. A multi-objective scheduling method for rail transit vehicles based on a V2X architecture according to claim 3, characterized in that, Within the global baseline operation diagram framework provided by S3, S4 utilizes three types of information—predicted passenger flow, real-time congestion, and speed-related safe train spacing—to make local decisions on whether to execute each candidate train formation / ungrouping window. This achieves adaptive adjustment of transport capacity at the section level without disrupting the overall structure of the global operation diagram.

5. A multi-objective scheduling method for rail transit vehicles based on V2X architecture according to claim 1, characterized in that, The rule pairs described in S5 include: 1) Weights of the multi-objective cost function in S3 ; 2) Passenger flow threshold, congestion threshold, and safe vehicle spacing coefficient in S4 ; 3) Some structures or hyperparameters of the passenger flow prediction model in S2.

Images