A big data assisted multi-objective commuting route scheduling decision support system

Abstract

The application relates to the technical field of public transport scheduling, in particular to a big data assisted multi-target commuting route scheduling decision support system, which comprises a digital twin ontology, a passenger flow response model, a collaborative optimization engine and an abnormal response model; the big data assisted multi-target commuting route scheduling decision support system completely solves the problem that static passenger flow prediction is disconnected with scheduling by embedding passenger flow elasticity response in the digital twin ontology and forming a two-way iterative closed loop of scheme input-behavior feedback-demand correction-reassessment with the optimization engine, so that the actual passenger flow matching degree of the final scheme is greatly improved; the adaptive double-precision switching mechanism based on the Pareto frontier entropy value and uncertainty is used to exert the high-speed evaluation advantage of the surrogate model in the initial optimization stage, the solution quality is guaranteed through directional accurate simulation in the later optimization stage, the surrogate model is updated incrementally online to prevent distribution drift, the best balance between simulation precision and solving efficiency is achieved, and the solving speed is improved on the premise of ensuring that the precision loss is reduced.

Description

Technical Field

[0001] This invention relates to the technical field of public transportation scheduling, and in particular to a big data-assisted multi-objective commuter route scheduling decision support system. Background Technology

[0002] Commuter routes (such as customized buses, shuttle buses, and company shuttles) are an important part of urban public transportation, and their scheduling quality directly affects commuting efficiency, operating costs, and passenger satisfaction. Currently, the scheduling of commuter routes is mostly based on manual experience, with departure plans formulated according to fixed timetables and rough passenger flow estimates, which makes it difficult to adapt to the spatiotemporal fluctuations of commuting demand and real-time changes in traffic conditions.

[0003] In the existing technology, some studies have used bus IC card data or GPS data to statistically analyze historical passenger flow patterns, but there are still the following shortcomings: (1) The data source is single. Relying solely on card swiping data is insufficient to fully reflect potential commuting travel demand, especially the demand for non-card swiping or transfers. Furthermore, there is a lack of deep integration of multi-source heterogeneous data such as mobile phone signaling, resulting in incomplete commuting OD identification and large deviations in passenger flow prediction; (2) Scheduling optimization often adopts a single-objective model, such as minimizing operating costs or passenger waiting time, ignoring the conflicts and trade-offs among multiple stakeholders, and the generated solutions lack flexibility and interpretability; (3) The existing decision-making process generally adopts an open-loop model of "prediction-optimization-execution", without constructing a dynamic closed-loop feedback between passenger flow demand and scheduling solutions. Passengers face extended waiting times and carriages The behavior of adjusting travel choices when the service level changes, such as congestion, is completely ignored, resulting in a low matching degree of actual passenger flow in the scheduling scheme; (4) There is a lack of high-fidelity virtual-real synchronization capability. The evaluation of the scheduling scheme relies heavily on offline simulation, which is not only slow to calculate, but also cannot map the real state of the physical system in real time, making it difficult to support high-frequency dynamic decision-making; (5) In the optimization process, it is difficult to balance simulation accuracy and solution efficiency. If high-precision simulation is used for all, it will take too long. If only a simple proxy model is used, it is easy to produce unacceptable accuracy loss. There is a lack of a mechanism to adaptively balance accuracy and efficiency in the optimization process; (6) The anomaly detection and response mechanism is weak. Most of them are based on a single threshold alarm. There is a lack of reasoning on the anomaly propagation path and the scope of influence. It is easy to trigger unnecessary full-line rearrangement, resulting in high adjustment costs and poor operational stability. For example, the route deviation bus dispatching system disclosed in the invention patent with announcement number CN113450242B, although it attempts a flexible operation mode in terms of route deviation dispatching, its dispatching decision is still based on offline static optimization. It fails to build a dynamic closed-loop feedback between passenger flow demand and dispatching plan. Its "elasticity" is only reflected in the limited deviation of the physical path and still does not have the true flexible dispatching capability for real-time demand fluctuations. Another example is the commuter bus dynamic dispatching method based on employee off-duty time and drop-off point disclosed in the invention patent with announcement number CN115034522B. Although it involves the scheduling allocation of multiple vehicles and multiple trips, its essence is still a one-way downward static combination optimization based on fixed registration data. It fails to achieve unified decision-making on dynamic vehicle turnover and time-varying passenger flow feedback between trips and lacks the dynamic linkage dispatching capability for real-time operation status.

[0004] Therefore, there is an urgent need for a big data-assisted multi-objective commuter route scheduling decision support system to improve the above-mentioned problems. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention provides a big data-assisted multi-objective commuter route scheduling decision support system. This system embeds passenger flow elasticity response into a digital twin and forms a bidirectional iterative closed loop with the optimization engine, encompassing scheme input, behavioral feedback, demand correction, and re-evaluation. This completely resolves the disconnect between static passenger flow forecasting and scheduling, significantly improving the actual passenger flow matching degree of the final scheme. Utilizing an adaptive double-precision switching mechanism based on Pareto front entropy and uncertainty, the system leverages the high-speed evaluation advantage of the surrogate model in the early stages of optimization and ensures solution quality through targeted and precise simulation in the later stages. Simultaneously, online incremental updates of the surrogate model prevent distribution drift, achieving an optimal balance between simulation accuracy and solution efficiency. This improves solution speed while minimizing accuracy loss. Furthermore, the system introduces anomaly propagation inference and graded response, not only detecting deviations but also inferring propagation paths and impact levels. This avoids a one-size-fits-all full-line rearrangement, incorporates adjustment costs into decision-making, significantly reduces the number of ineffective scheduling operations, and improves operational stability.

[0006] This invention provides a big data-assisted multi-objective commuter route scheduling decision support system, comprising: Digital twin ontology: Constructs a virtual road network for commuter routes, a virtual vehicle queue, a virtual passenger flow intelligent agent, and a virtual operating environment. It maintains state synchronization with the physical commuter system through real-time data streams, providing a high-fidelity simulation environment. Passenger flow response model: Embedded in the digital twin body, it includes passenger flow prediction and passenger flow elastic response. Passenger flow prediction outputs the initial passenger flow OD based on historical and real-time commuting data and injects it into the virtual passenger flow intelligent agent. Passenger flow elastic response receives the departure interval and vehicle congestion parameters of the scheduling scheme to be evaluated, drives the virtual passenger flow intelligent agent to adjust travel choices based on the travel utility function, generates the corrected passenger flow OD and feeds it back to the collaborative optimization engine. Collaborative optimization engine: Built-in multi-objective evolutionary algorithm and hierarchical simulation evaluation. Hierarchical simulation evaluation includes surrogate model and exact simulation. It is used to dynamically switch between the two simulation methods to evaluate the target value of the scheduling scheme based on the iterative convergence state of the optimization population and the Pareto front distribution characteristics. The surrogate model is updated online incrementally based on the results of exact simulation. The multi-objective evolutionary algorithm is combined with the simulation evaluation results to iteratively solve the problem and output the Pareto optimal departure timetable scheme set. Anomaly Response Model: Constructs a spatiotemporal correlation graph using the sites and road segments of the digital twin ontology as nodes, detects the state deviation of the physical and virtual systems in real time, infers the scope and degree of anomaly propagation and classifies anomaly levels, and pushes the corresponding granularity of scheduling and adjustment schemes to the decision-making end. By embedding passenger flow elasticity response into the digital twin and forming a two-way iterative closed loop of scheme input-behavioral feedback-demand correction-re-evaluation with the optimization engine, the problem of disconnect between static passenger flow forecasting and scheduling is completely solved, significantly improving the actual passenger flow matching degree of the final scheme. Utilizing an adaptive double-precision switching mechanism based on Pareto front entropy and uncertainty, the high-speed evaluation advantage of the surrogate model is leveraged in the early stage of optimization, while targeted precise simulation ensures the quality of the solution in the later stage. Simultaneously, online incremental updates of the surrogate model prevent distribution drift, achieving the best balance between simulation accuracy and solution efficiency, improving solution speed while minimizing accuracy loss. The introduction of anomaly propagation inference and graded response not only detects deviations but also infers propagation paths and impact levels, avoiding a one-size-fits-all full-line rearrangement, incorporating adjustment costs into decision-making, significantly reducing the number of ineffective schedulings, and improving operational stability.

[0007] Preferably, the travel utility function of the virtual passenger flow agent in the passenger flow elastic response includes waiting time cost, vehicle congestion cost, and ticket cost. The utility contributions of waiting time and congestion are non-linear to reflect passengers' aversion to losses due to waiting time and congestion. When the scheduling scheme causes the agent's utility loss to exceed the individual tolerance threshold determined based on historical travel data, the agent chooses to travel during off-peak hours, take other routes, or cancel the trip with corresponding probabilities, generating a corrected passenger flow OD distribution.

[0008] Preferably, the dynamic switching rule for hierarchical simulation evaluation is as follows: In the early stage of population iteration, when the Pareto front entropy value is higher than a set threshold, all individuals are evaluated by the surrogate model. When the Pareto front entropy value is lower than the threshold and the convergence phase is entered, accurate simulation is called to calibrate individuals at the edge of the Pareto front and individuals whose surrogate model prediction uncertainty is higher than the threshold. The surrogate model is composed of a residual fully connected network, which takes departure time coding and passenger flow characteristics as input and outputs approximate values ​​of passenger travel time cost, operating cost and carbon emissions.

[0009] Preferably, the online incremental update of the surrogate model uses the scheme-target value samples obtained from accurate simulation as the incremental training set, and updates the network parameters using mini-batch gradient descent, with the update frequency synchronized with the sampling frequency of accurate simulation.

[0010] Preferably, the anomaly reasoning calculates the weight of the propagation of node anomalies to adjacent road segments through a graph attention network. Combined with the anomaly duration prediction results, the anomaly state is divided into three levels: slight fluctuation, local anomaly, and global failure. Correspondingly, three adjustment strategies are triggered: fine-tuning of departure intervals, adding trains to sections, and rearranging the timetable for the entire line. Among them, the fine-tuning of departure intervals and adding trains to sections are added with minimizing the adjustment range as one of the optimization objectives.

[0011] Preferably, the passenger flow prediction adopts a spatiotemporal graph convolutional neural network with an attention mechanism, which includes an adaptive adjacency matrix generation layer, multiple spatiotemporal convolutional blocks composed of gated dilated convolutions and graph convolutions, and a multi-head self-attention module across time steps. It outputs the passenger flow OD prediction interval for multiple future time periods through quantile regression.

[0012] Preferably, the multi-objective evolutionary algorithm is NSGA-III or MOEA / D, the departure scheme adopts integer encoding based on a fixed time step, and a heuristic vehicle depot dispatch strategy is embedded during decoding.

[0013] Preferably, it also includes data acquisition for acquiring multi-source commuting data, with a built-in deep bidirectional long short-term memory network and a network structure combining multi-head self-attention, to identify commuting origin and destination points from mobile communication signaling and card swiping data, generate a standardized commuting OD matrix, and also includes a convolutional neural network for GPS trajectory road network matching to generate dynamic inter-station travel time distribution.

[0014] Preferably, it also includes decision-making interaction, which uses a digital twin visual engine to display the three-dimensional vehicle location, in-vehicle congestion heat map and carbon emission distribution, and incorporates a preference learning network based on inverse reinforcement learning to learn the preference utility function from expert interaction trajectories and rank and recommend Pareto optimal solutions.

[0015] A preferred method for operating a big data-assisted multi-objective commuter route scheduling decision support system includes the following steps: S1. Construct a digital twin of commuter routes, synchronize the physical system status in real time, and embed flexible passenger flow response; S2. Predict the initial passenger flow OD using a spatiotemporal graph convolutional network and inject it into a virtual passenger flow agent. The collaborative optimization engine then generates a set of departure plans to be evaluated. S3. For each scheme to be evaluated, the passenger flow elastic response driving agent adjusts the travel choice according to the scheme attributes, generates a corrected passenger flow OD and feeds it back. Based on the corrected passenger flow evaluation target value, the engine iterates through a multi-objective evolutionary algorithm and finally outputs the Pareto optimal departure scheme set. S4. Select an execution plan from the Pareto scheme set with the assistance of a preference learning network; S5. Real-time monitoring of virtual and real deviations in abnormal reasoning. When the deviation reaches the corresponding level, the abnormal response model is automatically triggered to provide an adjustment plan with the goal of minimizing the adjustment range and restoring operational order.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By embedding passenger flow elasticity response into the digital twin and forming a two-way iterative closed loop of scheme input-behavioral feedback-demand correction-re-evaluation with the optimization engine, the problem of disconnect between static passenger flow forecasting and scheduling is completely solved, and the actual passenger flow matching degree of the final scheme is greatly improved. 2. By utilizing an adaptive double-precision switching mechanism based on Pareto front entropy and uncertainty, the high-speed evaluation advantage of the surrogate model is leveraged in the early stages of optimization, while the quality of the solution is ensured through directional precise simulation in the later stages of optimization. At the same time, the surrogate model is updated incrementally online to prevent distribution drift, thus achieving the best balance between simulation accuracy and solution efficiency. The solution speed is improved while ensuring that the accuracy loss is reduced. 3. By introducing anomaly propagation reasoning and graded response, not only are deviations detected, but propagation paths and impact levels are also inferred, avoiding a one-size-fits-all full-line rearrangement. Adjustment costs are incorporated into decision-making, significantly reducing the number of ineffective scheduling and improving operational stability. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. The present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0019] Example 1: As Figure 1 As shown, a big data-assisted multi-objective commuter route scheduling decision support system includes: Digital twin ontology: High-precision virtual road network and vehicle models are constructed using 3D GIS and BIM. Each operating vehicle corresponds to a digital twin instance. Virtual passenger flow consists of tens of thousands of intelligent travel agents, and its initial travel chain is calibrated by generating a model based on historical mobile phone signaling and IC card data. Physical GPS, vehicle-mounted sensors, and card swipe data are used to calibrate the virtual vehicle position and the number of people inside the vehicle in real time through an LSTM state estimator, keeping the virtual-real deviation within 2%. Passenger Flow Response Model: Passenger flow prediction employs a spatiotemporal graph convolutional network that integrates multi-head self-attention. Inputting the past hour's network-wide OD sequence, weather, and holiday features, it outputs initial passenger flow ODs and their prediction intervals at 15-minute granularity for the next hour. This OD matrix is ​​loaded into a virtual passenger flow agent as the initial travel plan.

[0020] After the collaborative optimization engine generates a batch of departure timetable schemes, the passenger flow elastic response receives the departure interval vector of each scheme and the estimated carriage occupancy rate based on the current passenger flow. Each virtual passenger flow agent maintains a travel utility function, which has the form U=w1×E[waiting time]^α+w2×E[standing density in the carriage]^β+fare cost, where E[waiting time] and E[standing density in the carriage] are the average waiting time and average standing density that the agent is expected to experience under the current scheme, respectively; the exponents α and β reflect the passenger's loss aversion to waiting time and crowding. The coefficients, all less than 1, make the utility diminish marginally with increasing waiting time or congestion, meaning that passengers are more sensitive to additional waiting or congestion. The weights w1, w2 and loss aversion coefficients α and β are determined using historical passenger travel behavior data (including actual waiting time, vehicle occupancy rate and historical records of passengers changing or abandoning their trips) for this route, through the maximum likelihood estimation method. The parameters for different routes and different time periods can be determined separately to adapt to localized travel characteristics. The individual tolerance threshold θ is taken as the 80th percentile of the historical average waiting time and average occupancy rate of passengers for this route. When a new plan causes an agent's utility loss ΔU to exceed its tolerance threshold θ, the agent decides whether to change its travel plans using a sigmoid probability function: P(change) = 1 / (1 + exp(-κ(ΔU-θ))), where κ is a sensitivity coefficient that controls the steepness of the probability change. If the decision is made to change, the possible actions include: switching to an adjacent time slot, taking a parallel route, or abandoning the trip. The specific choice is randomly determined by a multinomial logit model based on the utility value of each option. After all agents have made their decisions, they are reassembled into a corrected passenger flow OD matrix and fed back to the optimization engine in real time to recalculate the objective function value, thus forming a closed loop of prediction-optimization-behavioral feedback-demand correction-re-optimization. Collaborative Optimization Engine: The engine is based on the NSGA-III algorithm. Individual encoding is an integer string of length 144 (divided into 5-minute steps for each day's operating hours), representing the number of vehicles dispatched in each time period. During decoding, feasible scheduling is generated by combining vehicle range and depot constraints. The hierarchical simulation evaluation includes a residual fully connected surrogate model and a digital twin simulation interface. In the early optimization stage (first 30 generations), the Pareto front entropy value E = -Σ(p_i·log p_i) of the current population's non-dominated solutions is calculated, where p_i is the proportion of individuals on the front edge after dividing it into target space grids. When E > 1.5, the solution set is considered to be still relatively dispersed, and all individuals are quickly evaluated for target values ​​by the surrogate model. As iterations proceed, the front edge gradually converges, and the entropy value decreases. When E < 1.2, a double-precision strategy is activated: first, the surrogate model prediction uncertainty of each individual is calculated (estimated by the sampling variance of the last three dropout layers of the model); then, individuals located at the edge of the Pareto front (near each target extreme value) with an uncertainty higher than 0.15, as well as representative individuals from the sparsest region of the front edge, are selected and fed into the digital twin for accurate simulation. The (solution, target value) samples obtained from the accurate simulation are immediately added to the incremental training cache. When the cache is full of 16 samples, the surrogate model is updated with a small batch of gradient descent in one step to achieve online incremental learning and prevent distribution drift. Under this mechanism, only about 15% of the individuals are accurately evaluated, but the hypervolume index of the final non-dominated solution set is maintained at more than 98% compared with the full accurate simulation. A single optimization can be completed within 30 seconds. Anomaly Response Model: Using stations and road segments in the digital twin as nodes, a spatiotemporal graph based on physical connections and passenger flow correlation is constructed. Each node features include: the deviation rate between real-time boarding numbers and simulated boarding numbers, the vehicle arrival time deviation rate, and the speed deviation rate of adjacent road segments. A 3-layer GraphSAGE graph attention network infers once every 30 seconds, outputting the anomaly score and anomaly propagation weight matrix for each node. A built-in temporal convolutional predictor predicts the possible duration of anomalies based on historical deviation trends. Based on the impact range of the abnormal scores, the abnormalities are divided into three levels: Slight fluctuations: If the abnormal score of a single station is >0.7 and the affected area is <3 stations, a fine-tuning of the departure interval is triggered. Based on the current timetable, the departure time of 3 consecutive trips is locally adjusted in 5-minute increments. The optimization goal is to minimize the "sum of squared deviations from the original scheduled departure time". Local anomalies: If the anomaly score is >0.7 and the affected area is between 3 and 10 stations or the predicted duration is more than 30 minutes, additional trains will be added in the affected section. Alternative solutions for inserting extra trains will be generated in the problematic section, with the optimization goals being operating costs and adjustment range. Global failure: If the abnormal score is >0.7 and the affected area exceeds 10 stations or the predicted duration exceeds 1 hour, a full timetable reordering will be triggered. At this time, the complete collaborative optimization engine will be called again, but the optimization target will include an adjustment range to avoid excessive changes. After all adjustment plans are quickly simulated using a digital twin, a comparison window pops up on the decision-making interaction interface. After confirmation by the dispatcher, the plans are issued for execution. In historical data backtesting, this hierarchical mechanism reduced the number of invalid full-line rearrangements by 72% and shortened the time for restoring operational order by 35%. Decision Interaction: Provides a 3D visualization interface that can preview vehicle operation heatmaps and carbon emission distributions for each Pareto scheme. It has a built-in preference network based on maximum entropy inverse reinforcement learning. Dispatchers can interact by dragging the "cost-time-dispatch" weight slider. The network learns its utility function in real time and automatically highlights and recommends the scheme that best matches the preference, greatly reducing the decision-making burden.

[0021] Data Acquisition: Used to acquire multi-source commuting data, it has a built-in network structure that combines a deep bidirectional long short-term memory network with multi-head self-attention. It identifies commuting origin and destination points from mobile communication signaling and card swipe data, generates a standardized commuting OD matrix, and also includes a convolutional neural network for GPS trajectory road network matching to generate dynamic inter-station travel time distribution.

[0022] The operation method of a big data-assisted multi-objective commuter route scheduling decision support system includes the following steps: S1. Construct a digital twin of commuter routes, synchronize the physical system status in real time, and embed flexible passenger flow response; S2. Predict the initial passenger flow OD using a spatiotemporal graph convolutional network and inject it into a virtual passenger flow agent. The collaborative optimization engine then generates a set of departure plans to be evaluated. S3. For each scheme to be evaluated, the passenger flow elastic response driving agent adjusts the travel choice according to the scheme attributes, generates a corrected passenger flow OD and feeds it back. Based on the corrected passenger flow evaluation target value, the engine iterates through a multi-objective evolutionary algorithm and finally outputs the Pareto optimal departure scheme set. S4. Select an execution plan from the Pareto scheme set with the assistance of a preference learning network; S5. Real-time monitoring of virtual and real deviations in abnormal reasoning. When the deviation reaches the corresponding level, the abnormal response model is automatically triggered to provide an adjustment plan with the goal of minimizing the adjustment range and restoring operational order.

[0023] The main functions achieved by this invention are: 1. By embedding passenger flow elasticity response into the digital twin and forming a two-way iterative closed loop of scheme input-behavioral feedback-demand correction-re-evaluation with the optimization engine, the problem of disconnect between static passenger flow forecasting and scheduling is completely solved, and the actual passenger flow matching degree of the final scheme is greatly improved. 2. By utilizing an adaptive double-precision switching mechanism based on Pareto front entropy and uncertainty, the high-speed evaluation advantage of the surrogate model is leveraged in the early stages of optimization, while the quality of the solution is ensured through directional precise simulation in the later stages of optimization. At the same time, the surrogate model is updated incrementally online to prevent distribution drift, thus achieving the best balance between simulation accuracy and solution efficiency. The solution speed is improved while ensuring that the accuracy loss is reduced. 3. By introducing anomaly propagation reasoning and graded response, not only are deviations detected, but propagation paths and impact levels are also inferred, avoiding a one-size-fits-all full-line rearrangement. Adjustment costs are incorporated into decision-making, significantly reducing the number of ineffective scheduling and improving operational stability.

[0024] OD is a standard term in the field of traffic engineering, which stands for Origin-Destination.

[0025] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A big data-assisted multi-objective commuter route scheduling decision support system, characterized in that, include: Digital twin ontology: Constructs a virtual road network for commuter routes, a virtual vehicle queue, a virtual passenger flow intelligent agent, and a virtual operating environment. It maintains state synchronization with the physical commuter system through real-time data streams, providing a high-fidelity simulation environment. Passenger flow response model: Embedded in the digital twin body, it includes passenger flow prediction and passenger flow elastic response. Passenger flow prediction outputs the initial passenger flow OD based on historical and real-time commuting data and injects it into the virtual passenger flow intelligent agent. Passenger flow elastic response receives the departure interval and vehicle congestion parameters of the scheduling scheme to be evaluated, drives the virtual passenger flow intelligent agent to adjust travel choices based on the travel utility function, generates the corrected passenger flow OD and feeds it back to the collaborative optimization engine. Collaborative optimization engine: Built-in multi-objective evolutionary algorithm and hierarchical simulation evaluation. Hierarchical simulation evaluation includes surrogate model and exact simulation. It is used to dynamically switch between the two simulation methods to evaluate the target value of the scheduling scheme based on the iterative convergence state of the optimization population and the Pareto front distribution characteristics. The surrogate model is updated online incrementally based on the results of exact simulation. The multi-objective evolutionary algorithm is combined with the simulation evaluation results to iteratively solve the problem and output the Pareto optimal departure timetable scheme set. Anomaly Response Model: A spatiotemporal correlation graph is constructed using the sites and road segments of the digital twin ontology as nodes. The deviation of the physical and virtual systems is detected in real time. The scope and degree of anomaly propagation are inferred and the anomaly level is classified. After matching the corresponding granularity of the scheduling and adjustment plan, it is pushed to the decision-making end.

2. The big data-assisted multi-objective commuter route scheduling decision support system as described in claim 1, characterized in that, The travel utility function of the virtual passenger flow agent in the passenger flow elastic response includes waiting time cost, vehicle congestion cost, and ticket cost. The utility contributions of waiting time and congestion are non-linear to reflect passengers' aversion to losses due to waiting time and congestion. When the scheduling scheme causes the agent's utility loss to exceed the individual tolerance threshold determined based on historical travel data, the agent chooses to travel during off-peak hours, take other routes, or cancel the trip with corresponding probabilities, generating a corrected passenger flow OD distribution.

3. The big data-assisted multi-objective commuter route scheduling decision support system as described in claim 1, characterized in that, The dynamic switching rule for hierarchical simulation evaluation is as follows: In the early stage of population iteration, when the Pareto front entropy value is higher than the set threshold, all individuals are evaluated by the surrogate model. When the Pareto front entropy value is lower than the threshold and the convergence phase is entered, the individuals at the edge of the Pareto front and the individuals whose surrogate model prediction uncertainty is higher than the threshold are calibrated by calling the accurate simulation. The surrogate model is composed of a residual fully connected network, which takes the departure time code and passenger flow characteristics as input and outputs approximate values ​​of passenger travel time cost, operating cost and carbon emissions.

4. The big data-assisted multi-objective commuter route scheduling decision support system as described in claim 1, characterized in that, The online incremental update of the surrogate model uses the scheme-target value samples obtained from accurate simulation as the incremental training set, and adopts mini-batch gradient descent to update the network parameters. The update frequency is synchronized with the sampling frequency of accurate simulation.

5. The big data-assisted multi-objective commuter route scheduling decision support system as described in claim 1, characterized in that, Anomaly reasoning calculates the weights of node anomalies propagating to adjacent road segments using a graph attention network. Combined with the anomaly duration prediction results, the anomaly state is divided into three levels: slight fluctuation, local anomaly, and global failure. Correspondingly, three adjustment strategies are triggered: fine-tuning of departure intervals, adding trains to sections, and rearranging the timetable for the entire line. Among them, minimizing the adjustment range of the scheduling is added as one of the optimization objectives in fine-tuning of departure intervals and adding trains to sections.

6. The big data-assisted multi-objective commuter route scheduling decision support system as described in claim 1, characterized in that, Passenger flow prediction employs a spatiotemporal graph convolutional neural network with an attention fusion mechanism, which includes an adaptive adjacency matrix generation layer, multiple spatiotemporal convolutional blocks composed of gated dilated convolutions and graph convolutions, and a multi-head self-attention module spanning multiple time steps. It outputs the passenger flow OD prediction interval for multiple future time periods through quantile regression.

7. The big data-assisted multi-objective commuter route scheduling decision support system as described in claim 1, characterized in that, The multi-objective evolutionary algorithm is NSGA-III or MOEA / D. The departure scheme adopts integer encoding based on a fixed time step, and a heuristic vehicle depot dispatch strategy is embedded during decoding.

8. The big data-assisted multi-objective commuter route scheduling decision support system as described in claim 1, characterized in that, It also includes data acquisition for acquiring multi-source commuting data, with a built-in deep bidirectional long short-term memory network and a multi-head self-attention network structure to identify commuting origin and destination points from mobile communication signaling and card swipe data, generate a standardized commuting OD matrix, and also includes a convolutional neural network for GPS trajectory road network matching to generate dynamic inter-station travel time distribution.

9. The big data-assisted multi-objective commuter route scheduling decision support system as described in claim 1, characterized in that, It also includes decision-making interaction, which uses a digital twin visual engine to display the three-dimensional vehicle location, in-vehicle congestion heat map and carbon emission distribution. It has a built-in preference learning network based on inverse reinforcement learning to learn the preference utility function from expert interaction trajectories and rank and recommend Pareto optimal solutions.

10. A big data-assisted multi-objective commuter route scheduling decision support system as described in any one of claims 1-9, characterized in that, The operation method of a big data-assisted multi-objective commuter route scheduling decision support system includes the following steps: S1. Construct a digital twin of commuter routes, synchronize the physical system status in real time, and embed flexible passenger flow response; S2. Predict the initial passenger flow OD using a spatiotemporal graph convolutional network and inject it into a virtual passenger flow agent. The collaborative optimization engine then generates a set of departure plans to be evaluated. S3. For each scheme to be evaluated, the passenger flow elastic response driving agent adjusts the travel choice according to the scheme attributes, generates a corrected passenger flow OD and feeds it back. Based on the corrected passenger flow evaluation target value, the engine iterates through a multi-objective evolutionary algorithm and finally outputs the Pareto optimal departure scheme set. S4. Select an execution plan from the Pareto scheme set with the assistance of a preference learning network; S5. Real-time monitoring of virtual and real deviations in abnormal reasoning. When the deviation reaches the corresponding level, the abnormal response model is automatically triggered to provide an adjustment plan with the goal of minimizing the adjustment range and restoring operational order.

Images