A composite route dynamic design method and system based on passenger flow prediction

Abstract

A dynamic design method and system for composite routes based on passenger flow forecasting is disclosed, relating to the fields of intelligent transportation and urban rail transit operation optimization. This method addresses the problems of low passenger flow forecasting accuracy, rigid and singular route schemes, delayed adjustment response, and incomplete evaluation objectives in existing urban rail transit route design technologies. The method includes: a station-level real-time cross-sectional passenger flow forecasting step, employing an improved LSTM combined with an Attention model to forecast real-time cross-sectional passenger flow at each station of the rail transit line; a composite route scheme generation step, dynamically generating a set of composite route schemes based on the passenger flow forecasting results and considering the line topology and operational constraints; using an improved NSGA-II multi-objective optimization algorithm to comprehensively evaluate and Pareto optimally select the generated route schemes; and issuing route scheme instructions through the ATS system, collecting actual operational data and feeding it back to the forecasting model for continuous optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and urban rail transit operation optimization technology, specifically to a composite route dynamic design method and system based on passenger flow prediction. Background Technology

[0002] Urban rail transit, as the backbone of public transportation systems in large cities, shoulders the ever-increasing passenger demand. Train route planning refers to the overall planning of train routes, starting and ending stations, departure frequencies, and stop patterns on the line, directly impacting the line's operational efficiency, service level, and economic benefits. A well-designed route can effectively reduce operating costs and improve resource utilization efficiency while meeting passenger flow demands.

[0003] However, existing urban rail transit route design technologies have the following technical shortcomings: (1) The accuracy of passenger flow forecasting is low. Current route design relies heavily on passenger flow forecasting methods that employ statistically based ARIMA (Autoregressive Integrated Moving Average) models or traditional BP (Back Propagation) neural network models. These models struggle to effectively capture long-term time-series dependencies and nonlinear fluctuations in passenger flow data, especially during peak hours, holidays, and unforeseen events. Their forecasting accuracy typically falls below 85%, failing to meet the precision requirements of refined route design.

[0004] (2) The route scheme is of a single type. Current technologies generally employ a fixed, single-route scheme, where all trains run back and forth between the origin and destination stations. This design cannot adapt to the uneven spatial and temporal distribution of passenger flow, leading to overcapacity in some sections and undercapacity in others, resulting in wasted train capacity and a decline in passenger service quality.

[0005] (3) Delayed response to route adjustment Adjustments to existing route plans rely on human experience for decision-making. The time from the discovery of changes in passenger flow to the effective implementation of route plan adjustments is usually no less than one hour, making it impossible to achieve real-time matching between route plans and changes in passenger flow. In scenarios such as peak-hour transitions and emergencies, delayed adjustment responses often lead to large-scale passenger congestion or idle service resources.

[0006] (4) The assessment objectives are not comprehensive. Current route planning evaluations primarily focus on a single objective (such as minimizing operating costs or passenger waiting time), lacking a comprehensive consideration of the conflicting relationships between multiple objectives. Single-objective optimization schemes have significant limitations in practical applications, with multi-objective conflict rates typically exceeding 20%, making it difficult to achieve a comprehensive balance between operational efficiency and service quality.

[0007] (5) Poor passenger travel experience Due to the combined effects of the aforementioned technical defects, under the current route design, the average waiting time for passengers during peak hours is usually no less than 5 minutes, the peak congestion in the carriages is relatively high, the transfer waiting time is long, and the overall travel experience urgently needs to be improved.

[0008] (6) Poor operating economics Due to the low degree of matching between the route plan and passenger flow demand, the empty running rate of existing trains is usually not less than 15%, resulting in high energy consumption and labor costs, and low operational economic efficiency.

[0009] In summary, existing urban rail transit route design technologies have significant shortcomings in terms of passenger flow forecasting, scheme diversity, dynamic response, and comprehensive evaluation. There is an urgent need for a new technology solution that can perform dynamic design of composite routes based on high-precision passenger flow forecasting. Summary of the Invention

[0010] This invention aims to overcome the technical problems in existing urban rail transit route design technologies, such as low accuracy of passenger flow forecasting, rigid and singular route schemes, delayed adjustment response, and incomplete evaluation objectives. To address these issues, this invention proposes a composite route dynamic design method and system based on passenger flow forecasting. This achieves high-precision matching between route schemes and real-time passenger flow demand, improving the operational efficiency and passenger service level of urban rail transit. The invention solves the aforementioned technical problems through the following technical solutions: Option 1: This invention proposes a dynamic design method for composite routes based on passenger flow forecasting, the method comprising the following steps: Step 1: Collect historical passenger flow data, real-time passenger flow data, time feature data and external factor data of each station on the rail transit line, construct a sliding window input sequence, and use an improved LSTM+Attention model to predict real-time cross-sectional passenger flow at the station level. Step 2: Based on the station-level cross-section passenger flow prediction results obtained in Step 1, and combined with the line topology, station connection relationships and operational constraints, dynamically generate composite route schemes. The composite route schemes include four basic types: single route, long and short route, Y-shaped route and loop route. The optimal route combination mode is adaptively selected according to the spatiotemporal distribution characteristics of the predicted passenger flow. Step 3: Use the improved NSGA-II multi-objective genetic algorithm to comprehensively evaluate and select the Pareto optimal solution set generated in Step 2; Step 4: Monitor the change rate of passenger flow density at the predicted cross-section and the signal of sudden events in real time. When the preset triggering conditions are met, automatically start the dynamic adjustment process of the route plan, re-execute steps 1 to 3, generate and switch to the new optimal route plan. Step 5: The optimal route plan obtained after evaluation and screening is sent to the train control center for execution through the ATS system. At the same time, actual train operation data, passenger flow data and plan execution effect data are collected. The feedback data is input into the improved LSTM+Attention model in Step 1, and the parameters of the improved LSTM+Attention model are updated and continuously optimized online.

[0011] Furthermore, a preferred embodiment is provided, wherein the improved LSTM+Attention model described in step 1 includes: The input layer is used to receive a multi-dimensional feature sequence within a sliding window. The multi-dimensional features include: historical cross-sectional passenger flow of each station, real-time passenger flow in and out of the station, hourly features, weekday features, holiday indicators, weather conditions, and special event indicators for the line. The LSTM encoding layer, containing multiple layers of LSTM units, is used to capture the temporal dependencies and long-term trend features of passenger flow data, outputting the hidden state sequence at each time step. ; The Attention layer is used to calculate the importance weights of the hidden states at each historical time step to the current prediction time step, and generate a context vector by weighted summation. The output layer is used to map the context vectors to station-level cross-sectional passenger flow prediction results.

[0012] Furthermore, a preferred embodiment is provided in which the Attention layer adopts a multi-head attention mechanism, which captures the multivariate dependencies of passenger flow features from different representation subspaces through multiple attention heads in parallel. The outputs of each attention head are concatenated and linearly transformed to obtain the final context vector.

[0013] Furthermore, a preferred embodiment is provided, wherein the constraints for generating the composite route scheme in step 2 include: Maximum passenger capacity constraint for trains: The number of passengers carried by a train in each route plan shall not exceed the maximum passenger capacity of the train. ; Minimum departure interval constraint: The departure interval between two adjacent trains shall not be less than the minimum departure interval allowed by the line. ; Maximum train formation constraint: The total number of train formations operating simultaneously on the line shall not exceed the maximum available train formation number. ; Line capacity constraints: Train traffic volume at each section shall not exceed the line's designed capacity. ; Connection constraints: The time coordination of each route at the connecting station meets the passenger transfer needs.

[0014] Furthermore, a preferred embodiment is provided, in step 2, the selection strategy for the four basic types of composite route schemes—single route, long-short route, Y-shaped route, and loop route—is as follows: When the predicted passenger flow is evenly distributed across all sections of the route and the total passenger flow is below a preset threshold, a single route is selected. When the predicted passenger flow shows a significant uneven distribution in different sections of the line, long and short routes are selected, with the long route covering the entire line and the short route covering the high passenger flow sections. When the line has a Y-shaped bifurcation topology and the passenger flow of each branch is highly independent, a Y-shaped route should be selected. When the route is a closed loop topology and the passenger flow has a circumferential flow characteristic, a loop route should be selected.

[0015] Furthermore, a preferred embodiment is provided, wherein the gene position includes route type, origin and destination stations, departure frequency and number of trains; An initial population generation strategy based on passenger flow distribution characteristics is introduced to determine the quality and diversity of the initial solution; Adaptive crossover and mutation probabilities are adopted and dynamically adjusted based on the population evolution generation and individual fitness to balance global and local search capabilities. An elite preservation strategy is introduced to determine the traces of Pareto optimal frontier solutions in the evolutionary process.

[0016] Furthermore, a preferred embodiment is provided, wherein the triggering conditions for the dynamic adjustment of the route plan in step 4 include: Condition for predicting the rate of change of passenger flow density at a cross-section: When the rate of change of passenger flow density at a predicted cross-section in any interval... Adjustments are triggered at specific times, among which For the newly predicted cross-sectional passenger flow density, The current plan is based on the cross-sectional passenger flow density; Emergency conditions: Adjustments will be triggered immediately in the event of train malfunction, station emergencies, extreme weather, or the end of a large event; Periodic adjustment condition: every preset time interval The system proactively detects and adjusts needs every minute.

[0017] Furthermore, a preferred embodiment is provided, wherein the feedback data in step 5 includes the actual passenger flow entering and exiting each station, the actual passenger flow of each section, the actual passenger load factor of the train, the train punctuality rate, the average waiting time of passengers, and passenger satisfaction survey data; the improved LSTM+Attention model is updated online using an incremental learning strategy, and the model is fine-tuned once after each preset batch of new data is accumulated.

[0018] Option 2: A composite route dynamic design system based on passenger flow forecasting, the system comprising: The station-level real-time cross-sectional passenger flow prediction module is used to collect historical passenger flow data, real-time passenger flow data, time feature data and external factor data of each station of the rail transit line, construct a sliding window input sequence, and use an improved LSTM+Attention model to predict station-level real-time cross-sectional passenger flow. The composite route scheme generation module is used to dynamically generate composite route schemes based on the station-level real-time cross-sectional passenger flow prediction results obtained by the station-level real-time cross-sectional passenger flow prediction module, combined with the line topology, station connection relationships and operational constraints. The composite route schemes include four basic types: single route, long and short route, Y-shaped route and loop route. The optimal route combination mode is adaptively selected according to the spatiotemporal distribution characteristics of the predicted passenger flow. The route selection module is used to comprehensively evaluate and select the Pareto optimal solution from the composite route selection module using the improved NSGA-II multi-objective genetic algorithm. The route planning dynamic adjustment module is used to monitor the change rate of passenger flow density at the predicted cross-section and emergency signals in real time. When the preset triggering conditions are met, the route planning dynamic adjustment process is automatically started, and the station-level real-time cross-section passenger flow prediction module is re-executed to the route planning evaluation module to generate and switch to a new optimal route plan. The route plan execution and feedback module is used to send the optimal route plan obtained after evaluation and screening to the train control center for execution through the ATS system. At the same time, it collects actual train operation data, passenger flow data and plan execution effect data, and inputs the feedback data into the improved LSTM+Attention model in the station-level real-time cross-sectional passenger flow prediction module to update and continuously optimize the parameters of the improved LSTM+Attention model online.

[0019] Option 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 1.

[0020] Option 4: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 1.

[0021] The advantages of this invention are: The present invention discloses a dynamic route design method and system based on passenger flow prediction. This method employs an improved LSTM+Attention deep learning model, capturing long-term dependencies in passenger flow time series through a long short-term memory network and adaptively focusing on key historical information through an attention mechanism, effectively improving prediction accuracy in abnormal scenarios. Verification using actual route data shows a prediction accuracy exceeding 95%, approximately 10 percentage points higher than existing traditional methods such as ARIMA.

[0022] This invention provides four basic route types: single route, long and short route, Y-shaped route, and loop route. It also supports adaptive selection of the optimal route combination mode based on the spatiotemporal distribution characteristics of passenger flow, which can flexibly meet the needs of different line topologies and passenger flow patterns. Compared with the existing single route scheme, it significantly improves the flexibility of capacity configuration.

[0023] This invention establishes a three-level triggering mechanism (emergency triggering, change triggering, and periodic triggering), enabling rapid matching of route plans with real-time passenger flow changes. The system adjustment cycle is no more than 5 minutes, significantly shortening the response time compared to the existing technology's adjustment cycle of no less than 1 hour.

[0024] This invention employs an improved NSGA-II multi-objective optimization algorithm, which simultaneously considers four optimization objectives: passenger waiting time, train empty running rate, capacity matching degree, and operating cost. It achieves a comprehensive balance among the multiple objectives through a normalized weighted fitness function, effectively reducing the conflict rate among the multiple objectives.

[0025] This invention, through the synergistic optimization of high-precision passenger flow prediction and dynamic design of composite routes, can effectively reduce passenger waiting time on platforms and congestion in train carriages. Practical application verification shows that the average passenger waiting time during peak hours does not exceed 3 minutes, a reduction of approximately 40% compared to existing technologies.

[0026] This invention effectively reduces train empty running rate and energy consumption by improving the matching degree between transport capacity and passenger flow. Practical application verification shows that the train empty running rate does not exceed 5%, a reduction of approximately 67% compared to existing technologies, and operating costs are reduced by approximately 15%. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the dynamic design method for composite routes based on passenger flow prediction as described in Implementation Method 1.

[0028] Figure 2 This is a schematic diagram of the network structure of the improved LSTM+Attention prediction model described in Implementation Method 1.

[0029] Figure 3 This is a closed-loop control framework diagram for the generation and evaluation of composite routing schemes as described in Implementation Method 1.

[0030] Figure 4 This is a schematic diagram of the topology of the four basic route types described in Implementation Method 1.

[0031] Among them, (a) is a single route, (b) is a long and short route, (c) is a Y-shaped route, and (d) is a loop route.

[0032] Figure 5 This is a flowchart of the improved NSGA-II multi-objective genetic algorithm described in Implementation Method 1. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0034] Implementation Method 1, see [link] Figures 1 to 5 This embodiment describes a dynamic design method for composite routes based on passenger flow forecasting. The method specifically includes the following steps: Step 1: Real-time passenger flow forecast at the station level Historical passenger flow data (including passenger flow in and out of stations at different times, cross-sectional passenger flow, and transfer passenger flow) and real-time passenger flow data (collected in real-time by the AFC (Automatic Fare Collection) system) are collected from each station on the rail transit line. Time characteristic data (hour, day of the week, month, and holiday indicators) and external factor data (weather conditions, information on large-scale events around the line, and information on emergencies) are also collected to construct a sliding window input sequence. ,in The sliding window length (usually 12 to 24 time steps, corresponding to 1 to 2 hours of historical data). For feature dimensions.

[0035] An improved LSTM+Attention model is used for real-time passenger flow prediction at the station level. The model first encodes the input feature sequence layer by layer through the LSTM encoding layer to capture the long-term trend features and complex temporal dependencies of passenger flow data; then, the Attention layer is used to calculate the differentiated importance weights of the hidden states at each historical time point to the current prediction time point, effectively highlighting the influence of key historical information; finally, the output layer generates the station-level passenger flow prediction results.

[0036] The prediction function is:

[0037] The calculation formula for the Attention mechanism is as follows:

[0038] In the formula, For time t, the entire line Vector of cross-sectional passenger flow prediction results for each station; Input the feature matrix into the sliding window; For LSTM networks at time The hidden state vector, To hide the state dimension; This is the weight matrix of the output layer; This is the bias vector for the output layer; , , Each is in a hidden state Through different linear transformation matrices , , Mapped to obtain; The dimension of the key vector is used to scale the dot product result to prevent the softmax function from entering the saturation region.

[0039] Step 2: Generation of composite route schemes Based on the station-level cross-section passenger flow prediction results obtained in step 1 By combining the line topology, station connections, and operational constraints, a set of composite route schemes is dynamically generated.

[0040] The composite routing scheme includes four basic types: Single route: Trains run back and forth between the starting station and the terminal station of the line, which is suitable for lines with relatively even passenger flow or during off-peak hours.

[0041] Long and short routes: Long-route trains cover the entire line, while short-route trains only run in sections with high passenger flow. This is suitable for scenarios where passenger flow is significantly unevenly distributed in different sections of the line (such as tidal passenger flow from suburbs to the city center during the morning rush hour).

[0042] Y-type route: Suitable for scenarios where the line has a Y-type bifurcation topology, the main line trains are flexibly allocated at different branch junctions, and each branch runs independently. It is suitable for lines where the passenger flow characteristics of each branch are very different.

[0043] Loop route: Applicable to lines with a closed loop topology, where trains run in one or two directions along the loop, suitable for scenarios where passenger flow is circular.

[0044] The constraints generated by the composite route scheme include: maximum passenger capacity constraint, minimum departure interval constraint, maximum number of train formations constraint, line capacity constraint, and route connection constraint, to ensure the feasibility of the generated scheme in actual operation.

[0045] Step 3: Route Plan Evaluation An improved NSGA-II multi-objective genetic algorithm was used to comprehensively evaluate and select the Pareto optimal solution set generated in step 2.

[0046] The optimization objective includes four dimensions: Minimize passenger waiting time: By optimizing the departure frequency and stop schedule in the route plan, the average waiting time of passengers on the platform is minimized; Minimize train empty running rate: By matching transport capacity supply with passenger flow demand, reduce the proportion of empty running mileage of trains in low passenger flow sections or during off-peak hours; Maximize capacity matching: maximize the matching degree between the capacity supply and the actual passenger flow demand in each section and time period. The matching degree is defined as the ratio of actual capacity to predicted passenger flow. Minimize operating costs: Taking into account train kilometer costs, energy costs, labor costs, and maintenance costs, the total operating cost of the solution is minimized.

[0047] The fitness function is:

[0048] In the formula, The average passenger waiting time corresponding to the plan (unit: minutes); The maximum passenger waiting time allowed for the route; The empty running rate of a train is defined as the ratio of the empty running mileage to the total running mileage. For capacity matching degree; The corresponding operating cost of the plan (unit: yuan); The maximum permissible operating cost for the line; , , , These are the weighting coefficients for each optimization objective, set by operations management personnel based on route operation strategies and preferences, to meet [the following criteria]. , .

[0049] Each sub-objective is normalized to The interval makes the fitness value A larger value indicates a better overall performance of the solution.

[0050] Step 4: Dynamic adjustment of route plan Real-time monitoring and prediction of passenger flow density change rate and emergency event signals, establishing a three-level triggering mechanism: Level 1 Trigger (Emergency Trigger): When a train malfunction, a station emergency, extreme weather, or the end of a large event occurs, the dynamic adjustment process of the route plan is immediately initiated.

[0051] Secondary trigger (change trigger): When the rate of change of passenger flow density in any interval of the predicted cross-section This triggers a reassessment and adjustment of the route plan.

[0052] Level 3 Trigger (Periodic Trigger): The system triggers at preset time intervals. ( (Every minute) it proactively performs a passenger flow forecast and route plan evaluation to detect whether there is a better route plan.

[0053] When any trigger condition is met, the system automatically re-executes steps 1 to 3 to generate a new optimal route plan and prepares for switching.

[0054] Step 5: Implementation and Feedback of the Route Plan The optimal route plan, selected through evaluation and screening, is issued to the train control center via the ATS system. The train control center then coordinates all trains to execute the new route plan. During the execution of the plan, real-time feedback data is collected, including actual passenger flow at each station, actual passenger flow in each section, actual train load factor, train punctuality rate, average passenger waiting time, and passenger satisfaction survey data.

[0055] The feedback data, after preprocessing, is input into the improved LSTM+Attention model in step 1. An incremental learning strategy is used to update the model parameters online. After accumulating a preset batch of new data (e.g., every 5000 new samples), the model is fine-tuned once, so that the prediction model can continuously adapt to the evolution of passenger flow patterns, forming a closed-loop control system of "prediction-generation-evaluation-execution-feedback-optimization".

[0056] Example 1: This example further explains and illustrates Implementation Method 1. Application Scenario: A certain urban rail transit line is approximately 35 kilometers long with 25 stations, running east-west and connecting the suburbs and the city center. During the morning peak hours (7:00-9:00), the line exhibits a significant tidal passenger flow characteristic—passenger flow is concentrated in the direction from the suburbs to the city center, while passenger flow in the opposite direction is sparse.

[0057] Step 1: Real-time passenger flow forecast at the station level: Historical AFC passenger flow data, train operation data, and external factor data for the past 6 months were collected to construct a training dataset. The data was aggregated at 5-minute time steps, with a sliding window length of [missing information]. (Corresponding to 2 hours of historical data), feature dimensions (Including passenger flow at each station, cross-sectional passenger flow, hourly characteristics, weekday characteristics, holiday markings, weather temperature, weather type, and special event markings for the line, etc.)

[0058] An improved LSTM+Attention model was used for training. The model structure parameters are as follows: the LSTM encoding layer contains 3 LSTM layers, each with 128 hidden units and a dropout rate of 0.2; the Attention layer uses a single-head attention mechanism, and the key vector dimension is... The output layer is a fully connected layer with an output dimension of 25 (corresponding to the predicted cross-sectional passenger flow values ​​for 25 stations). The model uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, 200 training epochs, and an early stop patience of 20 epochs.

[0059] On the test set during the morning rush hour (7:00-9:00), the model achieved a prediction accuracy of 96.8%, a root mean square error (RMSE) of 12.3 people / 5 minutes, and a mean absolute percentage error (MAPE) of 3.2%.

[0060] The forecast results show that during the morning peak hours, the passenger flow at the section from station 8 (suburban transfer station) to station 18 (core station in the central urban area) is significantly higher than that of other sections, with the maximum passenger flow reaching 28,500 people / hour, while the maximum passenger flow in the opposite direction is only 4,200 people / hour, and the passenger flow imbalance coefficient is as high as 6.79.

[0061] Step 2: Generation of composite route schemes: Based on the above passenger flow forecast results, the system determines that the current passenger flow pattern meets the applicable conditions for long-short routes (significantly uneven passenger flow in different sections of the line), and automatically generates the following composite route scheme: Long-distance route plan: Trains will run on the entire line between station 1 and station 25, with a departure frequency of 10 trains per hour and a 6-car train formation, mainly serving the basic passenger flow needs of the entire line.

[0062] Short-route option: Trains run between stations 8 and 18, with a departure frequency of 12 trains per hour and a 6-car formation, mainly serving the peak demand in high-passenger-flow sections.

[0063] The departure times for long and short routes at connecting stations (station 8 and station 18) are staggered to ensure even departure intervals and meet passenger transfer needs.

[0064] Verification of constraints generated by the plan: The maximum cross-sectional capacity (long route + short route combined) is 32,400 passengers / hour, which is greater than the maximum cross-sectional passenger flow demand of 28,500 passengers / hour, with a capacity surplus rate of 13.7%; the minimum departure interval is 2.7 minutes, which is greater than the minimum departure interval allowed by the line of 2 minutes; the number of train sets used simultaneously is 22, which is less than the maximum number of available train sets of 30; the throughput capacity of each section is within the line design capacity.

[0065] Step 3, Route Plan Evaluation: An improved NSGA-II multi-objective genetic algorithm was used to evaluate the above composite crossover scheme. Algorithm parameters were set as follows: population size 100, maximum generations 200, initial crossover probability 0.9, initial mutation probability 0.1, and Pareto front size 50. Weight coefficients were set as follows: (Passenger waiting time) (empty run rate) (Capacity matching degree) (Operating costs).

[0066] Evaluation results: The average passenger waiting time is 2.4 minutes, a 40% reduction compared to the original single-route plan's 4.0 minutes; the train empty-running rate is 4.8%, a 60% reduction compared to the original single-route plan's 12.0%; the capacity matching degree is 93.5%, a 29.9% increase compared to the original plan's 72.0%; and the overall operating cost is 15% lower than the original plan. The fitness function value is 0.847, and the plan passes the evaluation.

[0067] Steps 4 and 5: The system issues the route plan through ATS and continuously monitors it. No adjustment conditions are triggered during the morning peak hours, and the plan executes smoothly. The collected feedback data is used for online model updates.

[0068] Example 2: Holiday Combined Routes Design for a Certain Urban Rail Line Application scenario: A certain urban rail transit line passes through commercial centers, tourist attractions and transportation hubs. During holidays, the passenger flow pattern is significantly different from that on weekdays, showing the characteristics of multiple peaks throughout the day, complex directional distribution and sudden large-scale passenger flow.

[0069] Step 1: Real-time passenger flow forecast at the station level: During holiday passenger flow patterns, the improved LSTM+Attention model, the same as in Example 1, is used for prediction. The model has been pre-trained on weekday data, and a transfer learning strategy is adopted for holiday scenarios. It is fine-tuned using data from the past three holidays, with 50 rounds of fine-tuning and a learning rate of 0.0005.

[0070] On the test set during the National Day holiday (October 1-7), the model achieved a prediction accuracy of 95.5%, an RMSE of 18.7 people / 5 minutes, and a MAPE of 4.5%. The prediction results show that during the holiday, passenger flow in the commercial center area (stations 10-15) and the tourist area area (stations 18-22) exhibits a dual-peak pattern (10:00-12:00 AM and 2:00-5:00 PM), with the peak passenger flow periods in the two areas staggered and the directional distribution being relatively complex.

[0071] Step 2: Generation of composite route schemes: Based on passenger flow forecasts, the system determines that the current passenger flow pattern and route topology meet the applicable conditions for a Y-shaped route. The route has a fork at station 12, leading to a commercial center branch line (stations 13-16) and a tourist area branch line (stations 17-23), and the passenger flow characteristics of the two branches are highly independent.

[0072] The system generates a Y-shaped route plan: mainline trains run from station 1 to station 12, and then enter two branch lines after reaching the fork station; the commercial center branch line has a departure frequency of 8 trains / hour, and the tourist area branch line has a departure frequency of 10 trains / hour. At the same time, short-route trains are superimposed on the mainline in the high-passenger-flow section (stations 5-12), with a departure frequency of 6 trains / hour, to increase the mainline's capacity.

[0073] Step 3, Route Plan Evaluation: The same NSGA-II algorithm parameters and weight settings as in Example 1 were used. Evaluation results: Capacity matching degree was 91.2%, an improvement of 30.3% compared to the original single-route scheme's 70.0%; average passenger waiting time was 2.8 minutes; train empty-running rate was 5.2%; operating costs were reduced by 12% compared to the original scheme. Passenger satisfaction surveys showed a 25% increase in passenger satisfaction scores during holidays compared to previous periods. The fitness function value was 0.821, and the scheme passed the evaluation.

[0074] Example 3: Emergency route design for a sudden incident on a certain urban rail line Application Scenario: A train malfunction occurred on a city rail transit line during the evening rush hour (18:00-20:00), causing a single-track interruption between stations 10 and 15. The estimated repair time is 45 minutes. During this period, the normal passenger flow in this section is approximately 22,000 passengers per hour.

[0075] Step 4: Dynamic adjustment of the route plan: The system receives a train fault alarm signal via ATS, which meets the first-level trigger condition (emergency event), and immediately initiates the dynamic adjustment process of the route plan.

[0076] Step 1 (Re-execute), station-level real-time cross-sectional passenger flow forecast: Considering the disruptive impact of emergencies on passenger flow, the model incorporates emergency event identification features (interval disruption type, expected duration of impact) into the forecast, and re-predicts passenger flow in the section from station 10 to station 15. The forecast results show that, affected by train malfunctions, some passengers in this section will be transferred to adjacent lines or have their travel postponed, with the revised cross-sectional passenger flow estimated at approximately 18,500 passengers per hour.

[0077] Step 2 (re-execute): Generate composite route scheme: Based on the revised passenger flow forecast, the system generates an emergency route plan: Short-route plan A: The service will operate from station 1 to station 9, with a departure frequency of 12 trains per hour, ensuring normal operation north of the faulty section.

[0078] Short-route plan B: Operating from station 16 to station 25, with a departure frequency of 12 trains per hour, ensuring normal operation south of the faulty section.

[0079] Shuttle bus connection plan: A bus connection will be launched between the 9th and 16th stops, with a departure interval of 3 minutes, a passenger capacity of 80 people per bus, and a connection capacity of 1,600 people per hour.

[0080] Meanwhile, through station announcements and mobile apps, passengers are guided to transfer to adjacent lines, diverting approximately 30% of the affected passenger flow.

[0081] Step 3 (Re-execute), Route Plan Evaluation: The evaluation results show that the emergency plan took 3 minutes from triggering to generation, meeting the requirement that the dynamic adjustment cycle should not exceed 5 minutes; passenger evacuation time was reduced by 50% compared to the scenario without a plan (from the occurrence of the accident to the complete evacuation of all affected passengers was shortened from 90 minutes to 45 minutes); and operation recovery time was reduced by 40% (from 30 minutes after the fault was repaired to 18 minutes). The fitness function value was 0.756, and considering the constraints of the emergency scenario, the plan passed the evaluation.

[0082] Step 4: Implementation and feedback of the route plan: The system rapidly disseminated emergency route plans via the ATS, which were then coordinated and executed by each train control center. After the fault was repaired, the system triggered the adjustment process again, gradually transitioning back to the original combined route plan. All feedback data from this emergency response (passenger flow transfer patterns, evacuation times, passenger satisfaction, etc.) was recorded and used for incremental model learning to improve the ability to handle similar scenarios in the future.

[0083] Model Training and System Deployment Instructions Model training process: The improved LSTM+Attention model is trained using a strategy that combines offline training with online updates. The offline training phase initializes the model using historical data from the past 6 to 12 months. The online update phase employs an incremental learning strategy, performing model fine-tuning every 5000 new samples for 10 epochs, with a learning rate of 1 / 10 (0.0001) of that used in the offline training phase, to ensure the model's stability and adaptability.

[0084] System deployment architecture: The composite route dynamic design system of this invention adopts a layered deployment architecture. The data acquisition layer collects passenger flow data, train operation data, and external factor data in real time through multiple source interfaces such as the AFC system, ATS system, and meteorological API; the computing layer is deployed on a cloud server and is responsible for executing passenger flow prediction models, route plan generation algorithms, and multi-objective evaluation and optimization; the application layer provides operation and management personnel with visual displays and human-computer interaction interfaces through web and mobile terminals; and the execution layer interfaces with the ATS system through standard interface protocols to realize the automatic distribution and execution monitoring of route plans.

[0085] Key performance indicators: In actual deployments on multiple urban rail lines, the system has achieved the expected targets for all key performance indicators: passenger flow prediction accuracy ≥95%, route plan generation time ≤30 seconds, plan evaluation time ≤60 seconds, dynamic adjustment response time ≤5 minutes, and plan execution success rate ≥98%.

[0086] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0087] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A dynamic design method for composite routes based on passenger flow forecasting, characterized in that, The method includes the following steps: Step 1: Collect historical passenger flow data, real-time passenger flow data, time feature data and external factor data of each station on the rail transit line, construct a sliding window input sequence, and use an improved LSTM+Attention model to predict real-time cross-sectional passenger flow at the station level. Step 2: Based on the station-level cross-section passenger flow prediction results obtained in Step 1, and combined with the line topology, station connection relationships and operational constraints, dynamically generate composite route schemes. The composite route schemes include four basic types: single route, long and short route, Y-shaped route and loop route. The optimal route combination mode is adaptively selected according to the spatiotemporal distribution characteristics of the predicted passenger flow. Step 3: Use the improved NSGA-II multi-objective genetic algorithm to comprehensively evaluate and select the Pareto optimal solution set generated in Step 2; Step 4: Monitor the change rate of passenger flow density at the predicted cross-section and the signal of sudden events in real time. When the preset triggering conditions are met, automatically start the dynamic adjustment process of the route plan, re-execute steps 1 to 3, generate and switch to the new optimal route plan. Step 5: The optimal route plan obtained after evaluation and screening is sent to the train control center for execution through the ATS system. At the same time, actual train operation data, passenger flow data and plan execution effect data are collected. The feedback data is input into the improved LSTM+Attention model in Step 1, and the parameters of the improved LSTM+Attention model are updated and continuously optimized online.

2. The composite route dynamic design method based on passenger flow forecasting according to claim 1, characterized in that, The improved LSTM+Attention model mentioned in step 1 includes: The input layer is used to receive a multi-dimensional feature sequence within a sliding window. The multi-dimensional features include: historical cross-sectional passenger flow of each station, real-time passenger flow in and out of the station, hourly features, weekday features, holiday indicators, weather conditions, and special event indicators for the line. The LSTM encoding layer, containing multiple LSTM units, is used to capture the temporal dependencies and long-term trend features of passenger flow data, outputting the hidden state sequence at each time step. ; The Attention layer is used to calculate the importance weights of the hidden states at each historical time step to the current prediction time step, and generate a context vector by weighted summation. The output layer is used to map the context vectors to station-level cross-sectional passenger flow prediction results.

3. The composite route dynamic design method based on passenger flow forecasting according to claim 2, characterized in that, The Attention layer employs a multi-head attention mechanism, which captures the multivariate dependencies of passenger flow features from different representation subspaces through multiple parallel attention heads. The outputs of each attention head are concatenated and linearly transformed to obtain the final context vector.

4. The composite route dynamic design method based on passenger flow forecasting according to claim 1, characterized in that, The constraints for generating the composite route scheme in step 2 include: Maximum passenger capacity constraint for trains: The number of passengers carried by a train in each route plan shall not exceed the maximum passenger capacity of the train. ; Minimum departure interval constraint: The departure interval between two adjacent trains shall not be less than the minimum departure interval allowed by the line. ; Maximum train formation constraint: The total number of train formations operating simultaneously on the line shall not exceed the maximum available train formation number. ; Line capacity constraints: Train traffic volume at each section shall not exceed the line's designed capacity. ; Connection constraints: The time coordination of each route at the connecting station meets the passenger transfer needs.

5. The composite route dynamic design method based on passenger flow forecasting according to claim 1, characterized in that, In step 2, the selection strategy for the four basic types of composite route schemes—single route, long-short route, Y-shaped route, and loop route—is as follows: When the predicted passenger flow is evenly distributed across all sections of the route and the total passenger flow is below a preset threshold, a single route is selected. When the predicted passenger flow shows a significant uneven distribution in different sections of the line, long and short routes are selected, with the long route covering the entire line and the short route covering the high passenger flow sections. When the line has a Y-shaped bifurcation topology and the passenger flow of each branch is highly independent, a Y-shaped route should be selected. When the route is a closed loop topology and the passenger flow has a circumferential flow characteristic, a loop route should be selected.

6. The composite route dynamic design method based on passenger flow forecasting according to claim 1, characterized in that, The improved NSGA-II multi-objective genetic algorithm described in step 3 includes: A chromosome representation method based on real number encoding is adopted, where each chromosome represents a complete composite route scheme, and the gene positions include route type, origin and destination stations, departure frequency and number of trains; An initial population generation strategy based on passenger flow distribution characteristics is introduced to determine the quality and diversity of the initial solution; Adaptive crossover and mutation probabilities are adopted and dynamically adjusted based on the population evolution generation and individual fitness to balance global and local search capabilities. An elite preservation strategy is introduced to determine the traces of Pareto optimal frontier solutions in the evolutionary process.

7. The composite route dynamic design method based on passenger flow forecasting according to claim 1, characterized in that, The triggering conditions for the dynamic adjustment of the route plan in step 4 include: Condition for predicting the rate of change of passenger flow density at a cross-section: When the rate of change of passenger flow density at a predicted cross-section in any interval... Adjustments are triggered at specific times, among which For the newly predicted cross-sectional passenger flow density, The current plan is based on the cross-sectional passenger flow density; Emergency conditions: Adjustments will be triggered immediately in the event of train malfunction, station emergencies, extreme weather, or the end of a large event; Periodic adjustment condition: every preset time interval The system proactively detects and adjusts needs every few minutes.

8. The composite route dynamic design method based on passenger flow forecasting according to claim 1, characterized in that, The feedback data mentioned in step 5 includes the actual passenger flow entering and exiting each station, the actual passenger flow of each section, the actual passenger load factor of the train, the train punctuality rate, the average waiting time of passengers, and passenger satisfaction survey data; the improved LSTM+Attention model is updated online using an incremental learning strategy, and the model is fine-tuned once after each preset batch of new data is accumulated.

9. A composite route dynamic design system based on passenger flow forecasting, characterized in that, The system includes: The station-level real-time cross-sectional passenger flow prediction module is used to collect historical passenger flow data, real-time passenger flow data, time feature data and external factor data of each station of the rail transit line, construct a sliding window input sequence, and use an improved LSTM+Attention model to predict station-level real-time cross-sectional passenger flow. The composite route scheme generation module is used to dynamically generate composite route schemes based on the station-level real-time cross-sectional passenger flow prediction results obtained by the station-level real-time cross-sectional passenger flow prediction module, combined with the line topology, station connection relationships and operational constraints. The composite route schemes include four basic types: single route, long and short route, Y-shaped route and loop route. The optimal route combination mode is adaptively selected according to the spatiotemporal distribution characteristics of the predicted passenger flow. The route selection module is used to comprehensively evaluate and select the Pareto optimal solution from the composite route selection module using the improved NSGA-II multi-objective genetic algorithm. The route planning dynamic adjustment module is used to monitor the change rate of passenger flow density at the predicted cross-section and emergency signals in real time. When the preset triggering conditions are met, the route planning dynamic adjustment process is automatically started, and the station-level real-time cross-section passenger flow prediction module is re-executed to the route planning evaluation module to generate and switch to a new optimal route plan. The route plan execution and feedback module is used to send the optimal route plan obtained after evaluation and screening to the train control center for execution through the ATS system. At the same time, it collects actual train operation data, passenger flow data and plan execution effect data, and inputs the feedback data into the improved LSTM+Attention model in the station-level real-time cross-sectional passenger flow prediction module to update and continuously optimize the parameters of the improved LSTM+Attention model online.

10. The composite route dynamic design system based on passenger flow forecasting according to claim 1, characterized in that, The system also includes a data preprocessing module, which is used to clean, normalize and perform feature engineering on the collected multi-source heterogeneous passenger flow data. The visualization module is used to display passenger flow forecast results, route plan operation status and evaluation indicators in real time; the human-computer interaction module allows operation and management personnel to manually intervene in the selection and adjustment of route plans.

Images