Rail transit transportation delay recovery support device

Abstract

The invention relates to the technical field of rail transit operation management, in particular to a rail transit transportation delay recovery support device which comprises a shell and a processing module arranged in the shell, and the processing module at least comprises a data fusion module, a delay deduction module, an optimization scheme generation module, an evaluation and sorting module and a collaborative decision module. According to the method, rapid and scientific generation and multi-target tradeoff of the delay recovery scheme can be realized, and the operation recovery efficiency and the decision level are improved.

Description

Technical Field

[0001] This invention relates to the field of rail transit operation management technology, and in particular to a rail transit delay recovery support device. Background Technology

[0002] The execution and adjustment of train timetables are of paramount importance in operations management. Any external interference or internal malfunction can cause train operations to deviate from the plan, leading to service delays. Among these, the rapid recovery from train delays is a key technological direction in the field of rail transit operations control. This direction aims to minimize the impact on passenger travel and overall network operations when train operations are disrupted by unforeseen events (such as equipment failure, passenger surges, or severe weather) by adjusting the running times, stop times, and turnaround plans of subsequent trains.

[0003] Existing technologies often rely on dispatchers' experience for manual adjustments or use automated systems based on fixed rules for limited intervention. These methods have significant limitations when facing complex, cascading delay scenarios. They struggle to quickly assess the propagation impact of delays across the entire road network, and adjustment strategies are often limited to local lines or individual stations, lacking a global optimization perspective. Furthermore, the current integration and utilization of real-time dynamic data is insufficient, potentially leading to recovery plans that exacerbate imbalances between lines or ignore passengers' actual waiting times. Summary of the Invention

[0004] The purpose of this invention is to provide a rail transit delay recovery support device to solve the technical problems of existing train delay recovery relying on manual experience or fixed rules, lacking global optimization, having difficulty integrating real-time dynamic data, and being unable to provide alternative solutions with multi-objective trade-offs, resulting in a slow recovery process.

[0005] This invention provides a rail transit delay recovery support device, comprising a housing and a processing module deployed within the housing. The processing module includes at least a data fusion module, a delay prediction module, an optimization scheme generation module, an evaluation and ranking module, and a collaborative decision-making module. The data fusion module is used to collect and fuse heterogeneous data streams from various subsystems of the rail transit network in real time to output a globally consistent real-time snapshot of the rail transit network's status. The delay impact propagation prediction module receives the real-time status snapshot and automatically activates when it detects that the actual train arrival / departure point deviates from the planned value by more than a preset threshold. Based on discrete event simulation and graph theory models, it dynamically predicts the propagation path and impact range of the initial delay event within the entire rail transit network in future time windows, and outputs... The delay propagation prediction report is used to construct the delay recovery problem as a multi-objective combinatorial optimization problem based on the delay propagation prediction report, and solve it using an improved non-dominated sorting genetic algorithm with an elitist strategy to output a Pareto optimal solution set including multiple feasible recovery solutions. The evaluation and ranking module is used to perform microscopic simulations on each candidate recovery solution in the Pareto optimal solution set, and to perform comprehensive evaluation and ranking based on the simulation output using a multi-attribute decision model based on the approximation ideal solution ranking method to output a list of candidate solutions arranged in descending order of comprehensive utility score. The collaborative decision-making module is used to present the ranked candidate recovery solutions and their evaluation results in the form of a visual decision dashboard, and to receive interactive operations and decision feedback from the scheduler.

[0006] In some embodiments, the data fusion module includes at least a train status perception submodule, a passenger flow status perception submodule, an infrastructure status perception submodule, and an external environment perception submodule.

[0007] In some embodiments, the delay prediction module's prediction process includes: constructing a spatiotemporal state network graph with stations as nodes and section operation and station stopping operations as directed edges; using the detected initial delayed train and its affected stations as seed events, and based on the connection relationship specified in the train timetable, minimum tracking interval, turnaround operation time, train route plan, and real-time passenger flow load obtained from the data fusion module, iteratively calculating the expected arrival and departure delay time of subsequent trains at each station; wherein, the planned stopping time is dynamically adjusted according to the real-time passenger flow load, and the adjustment method is to add a compensation time based on passenger flow load to the base stopping time.

[0008] In some embodiments, the compensation time based on passenger flow load is determined by a compensation time calculation model, which divides the platform passenger density into 5 levels and assigns a baseline boarding and alighting time parameter to each level; when performing extrapolation calculations, a nonlinear compensation term obtained by fitting a second-order polynomial function is added to the baseline boarding and alighting time parameter according to the real-time perceived platform density level.

[0009] In some embodiments, the optimization scheme generation module defines and solves conflicting core optimization objectives; the core optimization objectives include at least: minimizing total delay time, minimizing total passenger waiting time, and minimizing operational adjustment costs; the improved non-dominated sorting genetic algorithm with elitist strategy merges the parent and offspring populations for non-dominated sorting after each generation of evolution, and preferentially selects all individuals in the first-ranked non-dominated layer to enter the next generation population.

[0010] In some embodiments, the train operation simulator built into the evaluation and ranking module integrates a random disturbance factor model; the random disturbance factor model includes at least: train section running time disturbances that follow a normal distribution with a mean of 0 and a standard deviation of 5 seconds, and station stopping accuracy disturbances that follow a uniform distribution within a range of ±3 seconds.

[0011] In some embodiments, the evaluation index set calculated by the evaluation and ranking module includes at least total delay time, total passenger waiting time, adjustment cost, average train travel speed, line section load imbalance coefficient, spare car utilization efficiency, and scheme robustness index; the scheme robustness index is quantified by injecting multiple sets of random small delays that conform to historical statistical patterns into the simulation and observing the degree of decay of the key performance indicators of the scheme.

[0012] In some embodiments, the collaborative decision-making module includes a scheme feedback learning unit; the scheme feedback learning unit is used to record the decision result and subsequent actual operation recovery data after the scheduler selects and executes a certain scheme, and to periodically update the weight parameters of the optimization problem in the optimization scheme generation module by comparing the scheme prediction indicators with the actual results.

[0013] In some embodiments, the model update process of the scheme feedback learning unit adopts an incremental learning approach. Every 24 hours, the scheme feedback learning unit uses the sample pairs formed by newly collected scheduler decision data and subsequent actual operation indicators to adjust the weight parameters online using the gradient descent method, so as to minimize the mean square error between the comprehensive utility score of the recommended scheme and the evaluation score of the actual operation results.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] 1. By using the data fusion module, global status perception of all elements of the rail transit network is achieved at the millisecond level, laying a precise data foundation for subsequent analysis;

[0016] 2. The delay prediction module adopts discrete event simulation based on spatiotemporal network diagrams, which can scientifically and quantitatively predict the chain reaction of delays and overcome the locality and lag of human experience judgment.

[0017] 3. The optimization solution generation module formalizes the recovery problem into a multi-objective optimization problem and uses an improved evolutionary algorithm to solve the Pareto optimal solution set, fundamentally resolving the contradiction that a single solution cannot balance multiple objectives such as recovery speed, passenger experience and operating costs.

[0018] 4. The evaluation and ranking module uses microscopic simulation and multi-attribute decision-making models to refine and rank the optimized solutions, ensuring that the recommended solutions are not only theoretically optimal but also practically feasible.

[0019] 5. The collaborative decision-making module organically combines intelligence with dispatcher experience, providing clear and intuitive decision-making basis while retaining necessary channels for manual intervention, and enabling the system to continuously self-optimize through a feedback learning mechanism;

[0020] 6. The entire device works in concert and can generate multiple well-demonstrated global optimization recovery plans within minutes, significantly improving the response speed, scientific decision-making, and recovery effectiveness of the rail transit system in the face of delays, and effectively ensuring operational efficiency and passenger service levels. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the rail transit delay recovery support device of the present invention;

[0023] Figure 2 This is the core principle diagram for the generation and simulation evaluation of optimization schemes in this invention. Detailed Implementation

[0024] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0025] Partial interpretation:

[0026] 1. Discrete event simulation and graph theory model: These are two methods used in system analysis, optimization and modeling. They are independent of each other, but can also be deeply combined through graph structures to abstractly describe discrete event systems.

[0027] 2. Non-dominated sorting genetic algorithm: This is a class of evolutionary algorithms for multi-objective optimization problems. Its core is to replace the fitness function of single-objective optimization with a "non-dominated sorting" mechanism to efficiently search for Pareto optimal solution sets. It is widely used in complex scenarios such as engineering design, resource allocation, and scheduling optimization.

[0028] 3. Pareto optimal solution set: The Pareto optimal solution set is the core solution set of multi-objective optimization problems, and it is also the search target of multi-objective optimization algorithms such as non-dominated sorting genetic algorithm;

[0029] 4. Approximation to Ideal Solution Ranking Method: This is a classic multi-objective decision-making method. Its core is to rank the alternatives by "calculating the closeness between the candidate alternatives and the ideal solution / negative ideal solution".

[0030] 5. Multi-attribute decision model: The multi-attribute decision model is a decision-making method for a limited number of alternatives and multiple evaluation indicators. Its core is to rank and select the optimal alternatives by quantifying attribute weights and alternative scores.

[0031] 6. Gradient Descent: This is a class of iterative optimization algorithms based on gradient information. Its core is to minimize or maximize the objective function by updating the parameters along the negative gradient direction of the objective function.

[0032] Example

[0033] This embodiment provides a rail transit delay recovery support device, including a housing and a processing module deployed within the housing. It is understood that the housing can be any existing enclosure, such as a small chassis or encapsulation shell. The deployment of the processing module within the housing is adapted to actual needs. The processing module includes at least a data fusion module, a delay prediction module, an optimization scheme generation module, an evaluation and ranking module, and a collaborative decision-making module. Specifically: the data fusion module is used to collect and fuse heterogeneous data streams from various subsystems of the rail transit network in real time to output a globally consistent real-time snapshot of the rail transit network status; the delay impact propagation prediction module receives the real-time status snapshot and automatically activates when it detects that the actual train arrival / departure point deviates from the planned value by more than a preset threshold. Based on discrete event simulation and graph theory models, it dynamically predicts the propagation path and impact range of the initial delay event within the entire rail transit network in future time windows to output a delay propagation prediction report; the optimization scheme generation module is used for... Based on the delay propagation prediction report, the delay recovery problem is constructed as a multi-objective combinatorial optimization problem and solved using an improved non-dominated sorting genetic algorithm with an elitist strategy, outputting a Pareto optimal solution set including multiple feasible recovery schemes. The evaluation and sorting module performs microscopic simulations on each candidate recovery scheme in the Pareto optimal solution set, and based on the simulation output, uses a multi-attribute decision model based on the approximation ideal solution sorting method to perform comprehensive evaluation and sorting, outputting a list of candidate schemes arranged in descending order of comprehensive utility score. The collaborative decision-making module presents the sorted candidate recovery schemes and their evaluation results in the form of a visual decision dashboard, and receives interactive operations and decision feedback from the scheduler.

[0034] This device integrates a complex system of data acquisition, intelligent analysis, optimization decision-making, and visualization interaction. Its core objective is to quickly generate scientific, feasible, and optimal operation recovery plans when train operations are delayed.

[0035] To better understand this invention, the details are as follows:

[0036] first,

[0037] The data fusion module continuously receives raw data streams from heterogeneous sources through multiple dedicated data interfaces in a streaming manner. The data fusion module is first divided into four sensing sub-modules according to the data domain: train status sensing sub-module, passenger flow status sensing sub-module, infrastructure status sensing sub-module, and external environment sensing sub-module. Each sub-module is responsible for the collection and preliminary analysis of specific types of data.

[0038] Furthermore, the train status awareness submodule directly accesses the data distribution service of the train automatic monitoring system and the vehicle-to-ground communication system via an industrial Ethernet network based on a security protocol. This submodule actively polls or subscribes to receive status messages from all online trains in the network at a fixed interval of 2 seconds. Each status message contains the following key fields:

[0039] 1. A unique train identifier, which is usually a combination of train number and trainset number;

[0040] 2. Longitude and latitude coordinates are provided by the train's onboard GPS receiver, with an accuracy better than 5 meters;

[0041] 3. Instantaneous speed value, in kilometers per hour;

[0042] 4. Operating direction code, indicating the up or down direction of the train in the current section;

[0043] 5. The planned arrival time at the next station;

[0044] 6. The actual estimated time stamp of arrival at the next station.

[0045] This submodule has a built-in message parsing and verification engine. The engine first verifies the digital signature and integrity of the message, and then extracts the above fields and calculates the deviation between the planned time and the actual time.

[0046] Furthermore, the passenger flow status perception submodule is responsible for capturing the distribution and movement of passengers in real time. The input sources for this submodule include two types:

[0047] The first type is video surveillance cameras deployed on station platforms, entrance and exit gates, transfer passages, and inside train carriages. These cameras output video streams at a rate of 25 frames per second.

[0048] The second type is infrared or laser passenger flow counting sensors installed at key entrances and exits.

[0049] The hardware carrier of this submodule is an edge computing device equipped with a graphics processor, which is deployed in the station equipment room. Each edge computing device runs a target detection and tracking neural network model based on deep learning. This model uses a pre-trained deep convolutional neural network as the backbone feature extractor to perform multi-scale feature map calculations on each frame of the input image.

[0050] Subsequently, the region proposal network generates hundreds of candidate region boxes on the feature map that may contain passengers. These candidate boxes are fed into the classification and regression heads. The classification head outputs the probability that the target in each box is a passenger, while the regression head refines the coordinate position of the box.

[0051] For targets identified as passengers, the model simultaneously extracts high-dimensional appearance feature embedding vectors. The tracking algorithm employs a joint detection and tracking paradigm, performing cross-frame identity association by calculating the cosine similarity of appearance features between the target detected in the current frame and the tracked target in the previous frame, as well as the cross-union ratio of motion positions predicted based on Kalman filtering, thus maintaining a temporary identity for each passenger.

[0052] Based on this, the submodule can output the real-time number of passengers, the pixel coordinates of each passenger, and the motion vector within its monitoring field of view in 1-second cycles.

[0053] Meanwhile, the central server of the passenger flow perception submodule aggregates the analysis results from all edge devices and, based on the preset geographic mapping relationship between cameras and sensors, synthesizes the raw counts into operationally meaningful indicators, including:

[0054] The passenger density of each platform waiting area, that is, the number of passengers per unit area, is expressed as people per square meter.

[0055] The passenger flow rate in each direction at each station, i.e., the number of people passing through the entrance gates per minute;

[0056] The real-time occupancy rate of each train carriage is the ratio of the current number of passengers to the rated capacity.

[0057] Furthermore, the infrastructure status awareness submodule extracts status information from the signaling system, power supply system, track structure health monitoring system, and maintenance scheduling system through standardized data interfaces. The data obtained from the signaling system includes:

[0058] The current display status of all signals along the entire line, such as green light, yellow light, red light, or off light;

[0059] The fixed / reverse position and locked state of all turnouts;

[0060] The status of all track circuits: occupied, idle, or faulty.

[0061] Data obtained from the power supply system includes: output voltage and current values ​​of each traction substation, and the status of contact wire section insulators; data obtained from the track monitoring system includes: track geometry parameters of key sections, such as changes in gauge, level, and elevation; data obtained from the maintenance and dispatch system includes: planned construction closure sections and times, and temporary equipment failure repair orders.

[0062] Furthermore, the external environment perception submodule periodically obtains refined weather forecast data for the operating area from the meteorological department's data platform, including: real-time temperature, precipitation, wind speed, visibility, and early warning signals such as rainstorms, strong winds, and lightning.

[0063] The raw data streams generated by the above four sub-modules are uniformly sent to the data cleaning and timestamp alignment engine of the data fusion module. This engine is a stream processing program running on an in-memory computing framework. Its cleaning logic includes outlier removal and missing value imputation. Specifically:

[0064] For numerical data, such as train speed, engine maintenance dynamics, historical sliding window mean and standard deviation, any value exceeding the mean by plus or minus 3 times the standard deviation is considered an outlier and will be replaced by the predicted value at that moment.

[0065] For state-based data, such as signal generator status, the engine performs continuity checks. If an impossible state transition occurs, the valid state from the previous moment is temporarily stored. Missing value imputation uses time series forecasting methods. For short-term missing values, linear interpolation is used; for missing values ​​over longer periods, patterns from the same historical period of the data source are used to fill the gaps. Timestamp alignment is a crucial step in the engine's process.

[0066] The engine runs a high-precision network time protocol client that is synchronized with the atomic clock in the operations control center with a time error of less than 1 millisecond. For each piece of input data, the engine first identifies its built-in timestamp field. If the data source itself provides a high-precision timestamp, it uses it directly; if the data source timestamp is not precise enough or is missing, it uses the local timestamp when the engine receives the data.

[0067] After being cleaned, all data is re-stamped with a unified millisecond-level timestamp based on Coordinated Universal Time. Finally, the engine generates a globally consistent real-time snapshot of the rail transit network every 5 seconds. The real-time snapshot of the rail transit network includes a complete set of train, passenger flow, infrastructure and environmental status data that is strictly synchronized in time and aggregated from all sub-modules.

[0068] then,

[0069] The delay prediction module continuously monitors the real-time status snapshots output by the data fusion module. Its activation condition is controlled by a configurable threshold trigger. The trigger continuously calculates the deviation between the planned arrival and departure time and the actual arrival and departure time of each train in the network. When it is detected that the absolute value of the arrival or departure deviation of any train at any station exceeds the preset 3-minute threshold for the first time, the trigger immediately activates the delay prediction module and marks the train and its station as the initial delay event seed.

[0070] Once the delay prediction module is activated, its first step is to construct a spatiotemporal state network graph for prediction calculations. This graph is a directed weighted graph where nodes represent the station's state at a specific time, and edges represent the train's movement or stopping process between nodes. Node attributes include station number, time slice index, and resource occupancy status. Edges are divided into two categories:

[0071] The first type is the interval running edge, with a weight equal to the planned running time plus possible perturbations;

[0072] The second category is the stop operation side, with the weight being the planned stop time.

[0073] The simulation process uses a discrete event kernel. The simulation clock advances in an event-driven manner, with the initial event being the triggered seed delay event. The kernel maintains a list of future events, each containing an event type (e.g., train arrival, train departure, turnaround start) and the event's occurrence time. The simulation starts from the current real time and progresses forward, with a default time window of 120 minutes, which can be dynamically adjusted based on the delay scale. For each processed train arrival event, the delay simulation module needs to calculate its dwell time at the station. This dwell time is not a fixed value but is dynamically adjusted based on the real-time platform passenger density provided by the data fusion and sensing module.

[0074] The delay prediction module incorporates a stop time compensation model. Based on historical data, this model divides platform density into five levels, from Level 1 (very sparse) to Level 5 (extremely crowded), with each level corresponding to a baseline boarding and alighting time parameter. During prediction calculations, the current platform density level is first obtained. Then, a compensation time is added to the baseline time. The compensation time is calculated using a fitted second-order polynomial function, which takes the density level as input and outputs the additional seconds. Specifically: Let the density level be d and the baseline time be T. base Compensation time T comp Calculated using the following formula: T comp =0.5×d+0.1×d, then the total stopping time is: T total =T base +T comp， This model ensures that the estimation of stop time is more conservative and more in line with reality under high passenger flow load.

[0075] After determining the stop time, the kernel calculates the train's planned departure time, which must satisfy two constraints:

[0076] First, not earlier than the arrival time plus the total stop time;

[0077] Second, no earlier than the time before the preceding train leaves the block section where the station is located plus the minimum safe following interval.

[0078] The minimum safe tracking interval is preset according to the line signaling system, for example, it may be 90 seconds under moving block. If the calculated departure time is later than the original plan, a departure delay will occur, which will be treated as a new event and affect the operation of the next section. When calculating the section running time, the kernel will consider the line gradient, curvature, and real-time speed limits, and use the estimated value based on the train dynamics model.

[0079] At the same time, strict handling of train routing and turnaround logic is required. If the arrival time of a train is delayed, the start time of its subsequent turnaround operation will be postponed accordingly. This may lead to two consequences:

[0080] Firstly, the train's departure time was delayed after it turned back;

[0081] Secondly, if the rolling stock is planned to be used for the opposite route, it may cause delays to the opposite trains due to the lack of rolling stock, i.e., a delay in rolling stock supply.

[0082] Finally, the delay simulation module generates a structured delay propagation prediction report, which includes: a list of all trains expected to be affected and their predicted delay times at each station; a list of affected stations; identified key bottleneck sections and stations; and a graphical representation of the delay propagation chain.

[0083] Immediately afterwards,

[0084] The optimization scheme generation module receives the prediction report output by the delay impact propagation simulation module and transforms it into a multi-objective combinatorial optimization problem. The decision space of this problem encompasses the main means of operational adjustments. Decision variables are encoded as a hybrid-coded chromosome, containing multiple gene segments, each corresponding to an adjustment decision. For example, the time adjustment gene segment is a real-number vector recording the time adjustment amount for each affected train at key stations, in seconds, ranging from -300 seconds to +600 seconds.

[0085] The skip station decision gene segment is a binary vector, where each bit indicates whether a specific station is skipped by a train, with 1 indicating skipping and 0 indicating stopping. The route change gene segment is an integer vector, indicating whether a train shortens or lengthens its route, changing to one of the predefined alternative route modes. The spare train deployment gene segment is also a binary vector, indicating whether a spare train is deployed in a specific depot or storage line to replace a severely delayed train or increase the headway.

[0086] This module defines three core objective functions that need to be minimized simultaneously:

[0087] The first objective function is to minimize total delay time. Mathematically, it is the sum of the absolute values ​​of the differences between the actual arrival and departure times of all trains at all stations and the original scheduled arrival and departure times under the adjusted new operating schedule. This objective directly measures the overall degree to which the operating schedule deviates from the original plan, aiming to restore operational order to normal as quickly as possible.

[0088] The second objective function is to minimize the total passenger waiting time. This objective focuses more on passenger experience, and its calculation requires integrating passenger flow data from delay propagation prediction. For each time slot at each station, the system dynamically simulates the arrival and departure of passengers based on the predicted train arrival and departure times and the real-time passenger flow rate. When a train is delayed or skips a station, passengers waiting on the platform accumulate, and the extra waiting time for each passenger is added up. Simultaneously, for stations that are skipped, passengers who were originally scheduled to board need to wait for subsequent trains, and their waiting time is also included. The value of this objective function is the sum of all these extra waiting times, expressed in person-hours.

[0089] The third objective function is to minimize the operational adjustment costs. This cost is a comprehensive metric that quantifies the resources consumed and indirect impacts of implementing adjustments. Its calculation method is as follows:

[0090] Each skip-station operation is assigned a fixed cost coefficient, which is set according to the passenger flow level of the station. Skipping stations at large stations has higher costs. Each route change is assigned another cost coefficient, reflecting the degree of disorder in the rolling stock utilization plan. The deployment of each spare train is associated with not only the fixed activation cost, but also the cost of its empty run to the deployment location.

[0091] To solve the aforementioned objective function optimization problem, an improved non-dominated sorting genetic algorithm with an elite strategy is adopted. During algorithm initialization, an initial population of 200 is randomly generated, with each individual representing a complete adjustment scheme chromosome. When generating random individuals, the algorithm has a built-in feasibility checker to ensure that the randomly generated gene combinations do not violate basic operational safety rules, such as: skipping stations cannot skip two or more first-level passenger flow stations consecutively; and the adjusted train schedule must meet the minimum tracking interval constraint.

[0092] The algorithm's evolutionary process is iterative, spanning 100 generations. In each generation, firstly, each individual in the population is decoded, converting its chromosomes into a specific set of adjustment instructions. Then, a fast evaluation model is used to calculate the individual's fitness value on three objective functions. The fast evaluation model is a simplified version of the inferential model, using approximate calculations to estimate the effectiveness of the adjustment scheme, thus ensuring the speed of the evolutionary process.

[0093] Subsequently, the algorithm enters the selection phase, first merging the parent population with the offspring population just generated through crossover mutation to form a merged population of size 400. Then, a non-dominated sort is performed on this merged population, which divides the individuals in the population into multiple levels:

[0094] The first layer is the non-dominated solution set, which consists of individuals in the population that are not superior to any other individual in all three objectives.

[0095] The second layer is the solution set dominated only by individuals in the first layer, and so on.

[0096] After sorting, the elite strategy begins. The algorithm prioritizes selecting all individuals from the first non-dominated layer into the next generation of the new population. If the number of individuals in the first layer is less than the population size of 200, individuals from the second layer are added. If adding an entire non-dominated layer would cause the population size to exceed 200, the entire layer is not added. Instead, individuals within that layer are selected based on their crowding distance. Crowding distance measures the density of an individual in the target space relative to its neighbors. A larger distance indicates a sparser environment around the individual, which is more valuable for maintaining population diversity. Therefore, the algorithm prioritizes retaining individuals with large crowding distances within that layer until the population size of 200 is reached.

[0097] After 100 generations of evolution, the algorithm converges to a Pareto optimal solution set, which typically contains 20 to 50 non-dominated solutions. Each solution represents a feasible recovery plan that achieves different balances among total delay time, passenger waiting time, and operational adjustment costs.

[0098] at last,

[0099] The evaluation and ranking module receives the Pareto optimal solution set from the optimization scheme generation module and performs microscopic simulation and comprehensive evaluation on each candidate scheme. The core of the evaluation and ranking module is a high-precision train operation simulator. This simulator is not a simple logical deduction, but a simulation environment that integrates multi-physics field details. The simulator models the traction and braking characteristic curves of the train and can calculate the acceleration, constant speed and deceleration process of the train in real time according to the gradient, curvature and speed limit of the line.

[0100] The simulator also integrates a stochastic disturbance factor model to simulate real-world uncertainties. This model includes two main disturbance sources:

[0101] 1. The train interval running time disturbance follows a normal distribution with a mean of 0 and a standard deviation of 5 seconds, and is sampled independently during each interval run;

[0102] 2. Station stopping accuracy disturbance, which is uniformly distributed between -3 seconds and +3 seconds, affecting the actual position of the train when it stops at the reference point each time;

[0103] For each input candidate recovery scheme, the simulator starts from the current real time and simulates the complete execution process of the scheme over the next 2 hours in a fixed step of 0.5 seconds. During the simulation, every adjustment instruction defined in the scheme is strictly followed, such as time adjustment, station skipping, and route change, and random perturbations are applied to the simulation environment.

[0104] After the simulation ends, the simulator outputs a detailed operation log, recording the position, speed, and status of each train at each moment, as well as the accumulation and dissipation of passenger flow at each station.

[0105] Based on this simulation log, the evaluation and ranking module calculates a more refined and comprehensive set of evaluation metrics than the optimization phase. This set of metrics includes six key metrics.

[0106] Indicator 1 is the total delay time, which is consistent with the optimization objective, but is calculated based on more accurate simulation results;

[0107] Indicator 2 is the total waiting time for passengers, which is also accurately accumulated based on the passenger travel chain dynamically generated in the simulation;

[0108] Indicator 3 is the operational adjustment cost, calculated based on the actual adjustment actions implemented in the plan;

[0109] Indicator 4 is the average train travel speed, which measures the overall operational efficiency of the entire train network after the implementation of the plan;

[0110] Index 5 is the load imbalance coefficient of the line section. This coefficient is obtained by calculating the ratio of peak section passenger flow to average section passenger flow, and reflects whether the scheme leads to excessive spatial concentration of passenger flow.

[0111] Indicator 6 is on standby.

[0112] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0113] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A rail transit delay recovery support device, comprising a housing, and processing modules disposed within the housing, characterized in that, The processing module includes at least a data fusion module, a delay prediction module, an optimization scheme generation module, an evaluation and ranking module, and a collaborative decision-making module, wherein: The data fusion module is used to collect and fuse heterogeneous data streams from various subsystems of the rail transit network in real time, so as to output a globally consistent real-time snapshot of the rail transit network status. The delay impact propagation simulation module is used to receive the real-time status snapshot and automatically activates when it detects that the actual departure and arrival points of the train deviate from the planned value by more than a preset threshold. Based on discrete event simulation and graph theory model, it dynamically predicts the propagation path and impact range of the initial delay event in the entire rail transit network within a future time window, and outputs a delay propagation prediction report. The optimization scheme generation module is used to construct the delay recovery problem into a multi-objective combinatorial optimization problem based on the delay propagation prediction report, and to solve it using an improved non-dominated sorting genetic algorithm with an elitist strategy, so as to output a Pareto optimal solution set including multiple feasible recovery schemes. The evaluation and ranking module is used to perform microscopic simulation on each candidate recovery scheme in the Pareto optimal solution set, and to perform comprehensive evaluation and ranking based on the simulation output using a multi-attribute decision model based on the approximation ideal solution ranking method, so as to output a list of candidate schemes arranged in descending order of comprehensive utility score. The collaborative decision-making module is used to present the sorted candidate recovery schemes and their evaluation results in the form of a visual decision dashboard, and to receive interactive operations and decision feedback from the dispatcher.

2. The apparatus according to claim 1, characterized in that, The data fusion module includes at least a train status perception submodule, a passenger flow status perception submodule, an infrastructure status perception submodule, and an external environment perception submodule.

3. The apparatus according to claim 1, characterized in that, The deduction process of the delay deduction module includes: Construct a spatiotemporal state network graph with stations as nodes and section operation and station stopping operations as directed edges; Using the detected initial delayed train and its affected stations as seed events, the estimated arrival and departure delay times of subsequent trains at each station are iteratively calculated based on the connection relationships specified in the train timetable, minimum tracking intervals, turnaround operation times, train route plans, and real-time passenger flow load obtained from the data fusion module. The planned stopping time is dynamically adjusted according to the real-time passenger flow load by adding a compensation time based on passenger flow load to the baseline stopping time.

4. The apparatus according to claim 3, characterized in that, The compensation time based on passenger flow load is determined by a compensation time calculation model. The compensation time calculation model divides the platform passenger density into 5 levels and assigns a baseline boarding and alighting time parameter to each level. When performing extrapolation calculations, a nonlinear compensation term obtained by fitting a second-order polynomial function is added to the baseline boarding and alighting time parameter based on the real-time perceived platform density level.

5. The apparatus according to claim 1, characterized in that, The optimization scheme generation module defines and solves conflicting core optimization objectives; The core optimization objectives include at least: minimizing total delay time, minimizing total passenger waiting time, and minimizing operational adjustment costs; The improved non-dominated sorting genetic algorithm with an elite strategy merges the parent and offspring populations for non-dominated sorting after each generation of evolution, and prioritizes all individuals in the first-ranked non-dominated layer to enter the next generation.

6. The apparatus according to claim 1, characterized in that, The train operation simulator built into the evaluation and ranking module integrates a random disturbance factor model; the random disturbance factor model includes at least: train section running time disturbances that follow a normal distribution with a mean of 0 and a standard deviation of 5 seconds, and station stopping accuracy disturbances that follow a uniform distribution within a range of ±3 seconds.

7. The apparatus according to claim 6, characterized in that, The evaluation index set calculated by the evaluation and ranking module includes at least total delay time, total passenger waiting time, adjustment cost, average train travel speed, line section load imbalance coefficient, spare car utilization efficiency, and scheme robustness index. The robustness index of the proposed scheme is quantified by injecting multiple sets of random micro-delays that conform to historical statistical patterns into the simulation and observing the degree of decay of the key performance indicators of the scheme.

8. The apparatus according to claim 1, characterized in that, The collaborative decision-making module includes a scheme feedback learning unit. After the dispatcher selects and executes a scheme, the scheme feedback learning unit records the decision result and subsequent actual operation recovery data, and periodically updates the weight parameters of the optimization problem in the optimization scheme generation module by comparing the scheme prediction indicators with the actual results.

9. The apparatus according to claim 8, characterized in that, The model update process of the scheme feedback learning unit adopts an incremental learning approach. Every 24 hours, the scheme feedback learning unit uses the sample pairs formed by newly collected dispatcher decision data and subsequent actual operation indicators to adjust the weight parameters online using the gradient descent method, so as to minimize the mean square error between the comprehensive utility score of the recommended scheme and the evaluation score of the actual operation results.

Images