A method and system for dynamic bus capacity adaptation based on multi-source data fusion

Abstract

This invention relates to the field of intelligent public transportation scheduling technology, specifically a method and system for dynamic bus capacity adaptation based on multi-source data fusion. Real-time data collection of station passenger flow, vehicle operation, road conditions, and environmental and operational constraints is performed. After noise filtering, anomaly removal, time alignment, and standardization, the data is deeply fused to generate representation vectors. Based on these vectors, a waiting risk function and a congestion-stop delay coupling function are constructed to obtain waiting risk indicators and congestion delay indicators. Under constraints such as fleet size, departure intervals, and road capacity, a network-level ensemble optimization model is established to solve decisions such as departure interval adjustment, adding or removing vehicles, and cross-line / station allocation, and these decisions are then implemented. The model parameters are continuously updated using execution effect data to improve supply-demand matching and operational stability in highly volatile scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent public transportation scheduling technology, specifically a method and system for dynamic bus capacity adaptation based on multi-source data fusion. Background Technology

[0002] Public transport capacity organization typically relies on a pre-determined schedule, combined with the experience of dispatchers, to handle daily departures and adjustments. To enhance operational visibility and emergency response capabilities, the industry has widely deployed information technology tools such as vehicle location tracking, arrival and departure event collection, card or QR code boarding records, station video passenger flow statistics, and external road condition and weather services. Based on these tools, several dynamic dispatching schemes have been developed. These schemes include measures such as triggering temporary additional buses, increasing departure intervals, short-line turnarounds, and cross-line support based on the number of people waiting at stations, occupancy rates, route delays, or congestion levels, in order to alleviate peak-hour congestion and the spread of delays.

[0003] However, from the perspective of objective operational mechanisms and engineering implementation, existing technologies still have significant shortcomings in dynamic capacity adaptation driven by multi-source data. Station passenger flow statistics are mostly updated at the second or event level; vehicle positioning and speed are subject to noise and jumps; road condition information has delays and inconsistent granularity, and is prone to missing or drifting when there is partial obstruction, communication interruption, or equipment failure. If only single-source or small-source data is simply spliced ​​or thresholded, it is easy to cause false triggers or missed triggers under conditions of inconsistent data caliber, timestamp misalignment, and outlier interference, leading to deviations between scheduling instructions and actual needs, and even causing operational disturbances due to frequent adjustments.

[0004] Current dynamic scheduling demand assessments often employ fixed-weighted indicators or single-indicator threshold triggers, making it difficult to characterize the most critical sources of uncertainty in public transport operations. In reality, passenger waiting experience is not only related to the average departure interval but also closely linked to the intensity of fluctuations in headway. Under the influence of factors such as congestion fluctuations, signal control, station entry and exit interference, and temporary traffic control, the variance of headway increases significantly, leading to unbalanced supply with long intervals or overlapping buses under the same average interval, thus creating sudden waiting congestion at the platform. If the assessment model ignores the fluctuation term of headway or implements jitter, it often underestimates the risks at the platform, resulting in a delayed response. Conversely, if strong intervention is carried out based solely on the instantaneous number of waiting passengers, short-term noise is easily mistaken for real demand, inducing unnecessary additional buses and interval disturbances.

[0005] Existing technologies often oversimplify the modeling of carriage congestion and stop delays, frequently using occupancy rates or stop times as sole criteria, or treating stop times as a linear function of passenger volume. In reality, stop delays exhibit significant congestion-coupled characteristics: as congestion increases, obstructed passageways, door areas, and passenger weaving introduce additional frictional delays, significantly increasing stop times for the same passenger volume. This further reduces intervals between trains, amplifies headway fluctuations, and creates a chain reaction of delays. Without a quantitative characterization of this congestion frictional delay, dispatching systems are prone to underestimating delay risks in congested scenarios, leading to counterproductive effects such as increased train frequency but more severe stop delays, or localized supplementary trains causing further congestion on main lines.

[0006] Existing solutions primarily employ single-line, localized strategies, lacking network-wide feasibility constraints and collaborative optimization mechanisms. Bus network operation is constrained by hard constraints such as fleet size, available vehicles at depots, upper and lower limits of route departure intervals, and road capacity. In trunk corridors or multi-route shared sections, increasing the frequency of a single route may trigger corridor capacity bottlenecks, leading to queuing and speed reductions, thus exacerbating overall delays. Cross-line support and depot allocation are also limited by resource boundaries. Without a unified ensemble optimization framework, problems such as unexecutable decisions, resource conflicts, or repeated oscillations under multi-objective conflicts can easily arise, making it difficult to maintain stable service levels under highly volatile conditions. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for dynamic public transport capacity adaptation based on multi-source data fusion, so as to solve the technical problems mentioned in the background art.

[0008] Based on the above ideas, the present invention provides the following technical solution:

[0009] A method for dynamic public transport capacity adaptation based on multi-source data fusion includes the following steps:

[0010] S1. Real-time collection of multi-source operation data of the bus network, wherein the multi-source operation data includes at least station passenger flow data, vehicle operation data, road condition data, and environmental and operational constraint data;

[0011] S2. Preprocess the multi-source operational data, including at least noise filtering, outlier removal, timestamp alignment, and feature standardization.

[0012] S3. Perform deep fusion on the preprocessed multi-source operational data and output a unified representation vector;

[0013] S4. Based on the representation vector, extract the station passenger flow arrival intensity, line departure interval, and travel time fluctuation intensity caused by road conditions, and construct a waiting risk function accordingly to obtain the waiting risk index.

[0014] S5. Based on the representation vector, extract the parameters of the station passenger flow arrival intensity, the line departure interval, and the vehicle's effective passenger carrying capacity and passenger boarding / alighting service capacity, and construct a congestion-stop delay coupling function accordingly to obtain the congestion delay index.

[0015] S6. Based on the waiting risk index and the congestion delay index, construct a network-level capacity adaptation set optimization model, solve for the adaptation decision set including at least departure interval adjustment, adding or removing vehicles, cross-line coordination or station vehicle allocation, and send the adaptation decision set to the dispatch terminal and vehicle terminal. Collect the effect data after execution to update the parameters of the waiting risk function, the congestion-stop delay coupling function or the set optimization model.

[0016] Through a closed-loop process of multi-source real-time data acquisition, preprocessing, joint representation generation, waiting risk assessment, congestion and station delay assessment, network-level set optimization, and execution feedback updates, platform-side risks and in-vehicle risks are integrated into the same decision-making chain. This enables the continuous output of executable capacity adaptation decisions in dynamic scenarios such as road condition fluctuations and sudden increases in passenger flow, reducing the risk of local optima and scheduling oscillations caused by single rule triggers, while improving the interpretability and sustainable iteration capability of decisions.

[0017] Preferably, step S4 specifically includes:

[0018] Within a preset statistical time window, the station passenger flow arrival intensity per unit time is calculated based on the station waiting number sequence and boarding records.

[0019] Calculate the departure interval of the line at the station based on the departure time difference between two adjacent trains;

[0020] Based on the length of the road segment between adjacent stations, the vehicle speed and its fluctuations, and combined with the filtering residuals from the preprocessing stage, the fluctuation intensity of the travel time between adjacent stations is calculated, and a plan execution jitter term is introduced to characterize the random disturbances in the scheduling execution and arrival / departure processes.

[0021] Based on the platform capacity or the maximum number of people that can be accommodated, the waiting risk is normalized, and the waiting risk index is output.

[0022] By refining the modeling of waiting risk into the acquisition and normalization of passenger flow arrival intensity within a statistical window, station departure intervals, travel time fluctuation intensity between adjacent stations, and execution random disturbances, waiting risk no longer relies solely on static thresholds but can automatically reflect the sources of uncertainty as road congestion and execution deviations occur. This improves the stability and robustness of station-level risk assessment and reduces the misleading influence of noisy data on scheduling conclusions.

[0023] Preferably, the determination of the waiting area risk index satisfies the following relationship:

[0024] The fluctuation intensity of travel time is obtained by converting the length of the road segment between adjacent stations, the average travel speed, and the speed fluctuation. The fluctuation intensity of travel time is combined with the planned execution jitter term to obtain the headway fluctuation intensity. The expected arrival volume of a single trip is calculated based on the station passenger flow arrival intensity and the departure interval. The second-order statistic of waiting time is determined by combining the average level of the departure interval and the headway fluctuation intensity. The second-order statistic of waiting time is then coupled with the station passenger flow arrival intensity and normalized by the platform capacity to obtain the waiting risk index. The waiting risk increases with the increase of station passenger flow arrival intensity, decreases with the increase of platform capacity, and increases nonlinearly with the increase of departure interval and the enhancement of headway fluctuation.

[0025] First, the time-distance fluctuation is synthesized from the fluctuation of travel time on the route segment and the execution disturbance. Then, the risk profile, which includes second-order statistics, is formed by combining passenger arrival and departure intervals and normalizing it with platform capacity. This relationship can simultaneously characterize the superposition and nonlinear amplification effect of two types of problems on waiting risk: the increase in average departure interval and the increase in frequency fluctuation. This makes the risk indicator more sensitive to congestion fluctuations and more equitable in terms of platform capacity differences, which is conducive to triggering targeted capacity intervention earlier.

[0026] Preferably, step S5 includes:

[0027] Obtain the vehicle's rated passenger capacity and real-time passenger capacity, and combine them with the upper limit of the comfort occupancy rate to determine the vehicle's effective passenger capacity;

[0028] The passenger boarding and alighting service capacity parameters are determined based on the number of doors, passage conditions, or average boarding and alighting time per passenger. These parameters include at least the boarding service rate and the alighting service rate.

[0029] The estimated number of passengers boarding a single vehicle at the station is calculated based on the passenger flow intensity at the station and the departure interval, and the estimated number of passengers alighting is calculated based on the alighting ratio.

[0030] Based on parameters such as vehicle effective passenger capacity, in-vehicle passenger volume, passenger boarding and alighting service capacity, and expected boarding and alighting volume, a congestion-stopping delay coupling function is constructed to output a congestion delay index.

[0031] The assessment process for congestion and stop delays is concretized into a joint modeling of effective passenger capacity, passenger boarding and alighting service capacity, expected passenger boarding and alighting volume, and in-vehicle passenger capacity. It can simultaneously incorporate vehicle load and station boarding and alighting speeds, avoiding judgments based solely on occupancy rate or stop time. This improves the ability to identify the chain reaction of congestion leading to extended stops, and how extended stops further affect headway, thereby supporting more refined departure intervals and vehicle deployment strategies.

[0032] Preferably, the determination of the congestion delay index satisfies the following relationship:

[0033] The estimated boarding volume is obtained by multiplying the station passenger flow intensity by the departure interval; the estimated alighting volume is determined by the alighting ratio and the estimated boarding volume; the dwell time is obtained by superimposing fixed time items such as vehicle entry and door opening / closing, the ratio of estimated boarding volume to boarding service rate, the ratio of estimated alighting volume to alighting service rate, and a congestion friction delay item related to the degree of congestion inside the vehicle. The congestion friction delay item increases as the proportion of passenger volume inside the vehicle to the vehicle's effective passenger capacity increases, and also increases as the total number of passengers boarding and alighting increases; the congestion delay index is composed of the degree of congestion after arrival and the relative proportion of dwell time to departure interval, wherein the degree of congestion is determined by the relationship between passenger volume inside the vehicle, estimated boarding volume, estimated alighting volume, and the vehicle's effective passenger capacity.

[0034] The congestion delay index is broken down into two physically distinct components: one reflecting the degree of congestion upon arrival (the relationship between passenger capacity and effective passenger capacity, and the number of passengers boarding and alighting), and the other reflecting the proportion of stop time relative to the interval between trains. A frictional delay term, which increases with increasing congestion, is also introduced. This modeling approach can more realistically reflect the stop tail effect in highly congested scenarios, reduce over-scheduling in mildly congested scenarios, and improve the discriminative power and reliability of delay assessments.

[0035] Preferably, step S6 includes:

[0036] Construct a network-level decision set to be solved, which includes at least the departure interval adjustment amount of each line, the number of standby vehicles deployed or recalled, cross-line collaborative replenishment strategy, or depot vehicle allocation strategy.

[0037] Construct network-level constraints, which include at least fleet size constraints, station available vehicle constraints, departure interval upper and lower limits constraints, and road capacity constraints.

[0038] The network-level decision set is solved by jointly using the waiting risk index and the congestion delay index, and the constraint tightness is adaptively updated based on the execution effect data to suppress scheduling instability under congestion fluctuation conditions.

[0039] Network-level decision-making is defined as a set of executable actions, including adjusting departure intervals, deploying / recovering standby vehicles, cross-line coordinated support, and depot vehicle allocation. These actions are accompanied by constraints such as fleet size, available vehicles at depots, interval limits, and road capacity. Furthermore, the tightness of these constraints is adaptively updated to mitigate instability under congestion fluctuations. This approach avoids resource conflicts and infeasible solutions resulting from relying solely on single-line experience for scheduling, making scheduling actions more robust and feasible under global constraints.

[0040] Preferably, the solution to the wire mesh-level ensemble optimization model satisfies the following relationship:

[0041] The optimization cost is set by taking the degree of exceeding the limits of waiting risk indicators and congestion delay indicators on the station set as the objective, and by taking the upper and lower limits of departure interval, total number of vehicles, available vehicles at the station, and the upper limit of road capacity related to road density or congestion level as constraints. The optimization is carried out on the network-level decision set to obtain the optimal decision that satisfies the constraints and minimizes the cost of exceeding the limits. Based on the optimal decision, an adaptation decision set is generated, which includes at least departure interval adjustment, adding or removing vehicles, cross-line collaborative supplementation, or station vehicle allocation.

[0042] Using excessive waiting time and congestion delays as the core optimization objectives, and seeking the optimal decision under multiple hard constraints, the output solution naturally satisfies the operational and road capacity boundaries. Reducing the number of buses could exacerbate congestion, or shortening intervals could lead to train collisions. Compared to fixed-weighted approaches or simple rule-triggered methods, this solution is more conducive to obtaining a compromise-optimal and interpretable solution in multi-objective conflict scenarios, improving network-level coordination efficiency.

[0043] A public transport dynamic capacity adaptation system based on multi-source data fusion is provided to implement the aforementioned public transport dynamic capacity adaptation method based on multi-source data fusion.

[0044] The system includes a processor, a memory, and a multi-source data interface communicatively connected to the processor. The memory stores a computer program, which, when executed by the processor, includes the following functional modules:

[0045] The multi-source data acquisition module is used to collect multi-source operational data of the public transport network in real time.

[0046] The data preprocessing module is used to perform noise filtering, outlier removal, timestamp alignment, and feature standardization on the multi-source running data.

[0047] The deep fusion module is used to perform deep fusion on the preprocessed data and output a unified representation vector.

[0048] The waiting risk modeling module is used to extract the station passenger flow arrival intensity, line departure interval, travel time fluctuation intensity and plan execution jitter based on the representation vector, and construct the waiting risk function to obtain the waiting risk index.

[0049] The delay coupling modeling module is used to extract parameters such as station passenger flow arrival intensity, line departure interval, vehicle effective passenger capacity, in-vehicle passenger volume, alighting ratio, and passenger boarding / alighting service capacity based on the representation vector, and to construct a congestion-stop delay coupling function to obtain congestion delay index.

[0050] The network-level set optimization solution module is used to construct and solve the network-level set optimization model based on the waiting risk index and the congestion delay index, and generate an adaptive decision set under the conditions of satisfying the constraints of fleet size, available vehicles at the station, upper and lower limits of departure interval and road capacity.

[0051] The decision distribution and feedback update module is used to distribute the adapted decision set to the scheduling terminal and vehicle terminal and trigger execution, collect the effect data after execution, and update the model parameters, threshold parameters or constraint tightness parameters of the waiting risk modeling module, the delay coupling modeling module or the network-level set optimization solution module based on the effect data.

[0052] The data acquisition, preprocessing, joint representation generation, waiting risk modeling, congestion-stop delay modeling, network-level ensemble optimization solution, and decision issuance and feedback update are broken down into deployable functional units, which facilitates integration with existing dispatch terminals, vehicle terminals, and multi-source data interfaces to achieve project implementation and expansion. At the same time, continuous self-calibration is achieved through the feedback update module, which can maintain performance stability and decision consistency in scenarios such as seasonal passenger flow changes, road construction, and emergencies.

[0053] The technical solution of the present invention may include the following beneficial effects:

[0054] This invention collects and preprocesses multi-source operational data of the public transport network in real time, and generates heterogeneous information through cross-source alignment and joint representation. This enables passenger flow, vehicles, road conditions, and operational constraints to form a unified input that can be used for evaluation and decision-making within the same data semantic space. This improves data availability and consistency from the source, reduces interference from noise, time sequence misalignment, and inconsistent standards in scheduling judgments, and ensures that subsequent capacity adaptation calculations still have a stable input foundation and sustainable iteration capability in dynamic scenarios.

[0055] At the assessment level, this invention constructs a coupled index system for waiting risk and congestion-stop delay: waiting risk considers passenger arrival intensity, departure interval and its fluctuation uncertainty, and normalizes it in combination with platform carrying capacity; congestion-stop delay considers effective passenger capacity, in-vehicle passenger capacity, passenger boarding and alighting service capacity and expected passenger boarding and alighting volume, and introduces a delay effect that increases with the degree of congestion, thereby incorporating platform-side risk and carriage-side risk into the same logical link, improving the sensitivity and differentiation of scenarios such as congestion fluctuations and sudden increases in passenger flow, and avoiding misjudgment and over-scheduling caused by a single threshold or single index.

[0056] At the decision-making and implementation level, this invention drives network-level ensemble optimization based on the degree of exceeding the limits of waiting risk and congestion delay. Under constraints such as fleet size, available vehicles at depots, departure interval boundaries, and road capacity, it outputs a set of executable decisions, including departure interval adjustment, adding or removing vehicles, cross-line collaborative replenishment, and depot vehicle allocation. The model parameters and the tightness of constraints are updated in a closed loop through the execution effect data, making the scheduling scheme more feasible and robust under global constraints, reducing the risk of train collisions and scheduling oscillations, and improving the overall service level and resource utilization efficiency of the network. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the workflow of a dynamic public transport capacity adaptation method based on multi-source data fusion according to the present invention.

[0058] Figure 2 This is a block diagram of a public transport dynamic capacity adaptation system based on multi-source data fusion according to the present invention. Detailed Implementation

[0059] Example 1

[0060] like Figure 1 A method for dynamic public transport capacity adaptation based on multi-source data fusion includes the following steps:

[0061] S1: Collects multi-source operational data such as passenger flow, vehicles, road conditions, and operational constraints of the public transport network.

[0062] S2: Cleaning and aligning multi-source data, including noise reduction, anomaly removal, time synchronization, and standardization.

[0063] S3: Perform deep fusion on the preprocessed multi-source data to generate a unified representation vector.

[0064] S4: Extract arrival intensity, departure interval and travel fluctuation from the representation vector to calculate waiting risk indicators.

[0065] S5: Extract parameters related to capacity and passenger boarding / alighting ability from the representation vector, and calculate the congestion delay index.

[0066] S6: Based on two types of indicators, perform network-level set optimization and solve the problem, and issue scheduling decisions. Collect effect data to update model parameters on a rolling basis.

[0067] In this embodiment, the method employs a rolling cycle online operation, with the rolling cycle Δt selectable from 30 seconds to 300 seconds, preferably 60 seconds; the statistical time window W selectable from 5 minutes to 30 minutes, preferably 10 minutes. Multi-source operational data includes at least: station passenger flow data (station number, timestamp, number of waiting passengers or estimated number of waiting passengers, number of boarding passengers, optional number of alighting passengers / transfer passengers), vehicle operation data (vehicle number, route number, latitude and longitude, speed, arrival / departure event time, door opening / closing time), road condition data (road segment number, average speed, speed standard deviation or congestion level, optional construction / control signs), and environmental and operational constraint data (platform capacity, minimum / maximum allowable departure interval, number of available vehicles in the fleet, number of available vehicles in the depot, weather conditions, etc.). The sampling frequency of station passenger flow data can be selected from 1 second to 10 seconds (video / infrared counting) or updated according to arrival events (card swiping / code scanning), the sampling frequency of vehicle positioning can be selected from 1 second to 5 seconds, and the sampling frequency of road conditions can be selected from 30 seconds to 5 minutes. Missing data and outliers are handled as follows: Negative waiting numbers or short-term spikes (difference between two consecutive values ​​exceeding 30% of platform capacity and lasting <20 seconds) are considered outliers and corrected using median filtering or neighborhood interpolation. Vehicle position jumps (displacement and speed inconsistent within Δt) are removed using Kalman filtering and filled in using uniform extrapolation. For consecutive missing data ≤30 seconds, extrapolation is preferred; for >30 seconds, historical speed statistics for the same time period are used to fill in the missing data, and the confidence level of the source is lowered. When road condition data is missing, the average speed and speed standard deviation are preferably calculated using speed statistics of multiple vehicles on the same road segment. If the sample size is insufficient (e.g., <3 vehicles / cycle), it degenerates to historical statistics for the same time period. During preprocessing, all source data are aligned to a unified time raster. Numerical fields are standardized by route / station grouping, with the standardization parameter preferably updated using a sliding window over the last 7 days. Outlier removal (IQR or 3σ rule optional) and timestamp alignment (events aligned to the nearest raster or interval aggregation) are performed. The deep fusion described in step S3 is used to output a unified representation vector. The deep fusion can be implemented by "cross-source feature alignment + joint representation generation": the passenger flow source, vehicle source, road condition source, and constraint source are encoded into source vectors of the same dimension (discrete fields are embedded and numerical fields are directly concatenated), and then a unified representation vector is output through a joint representation model. The joint representation model can be a multi-layer attention encoder or a gated concatenation network, and its output dimension d can be selected from 32 to 256, preferably 64 or 128.This representation vector serves as the common input for subsequent S4 and S5, robustly extracting parameters such as arrival intensity, departure interval, travel time fluctuation intensity, vehicle passenger load, and passenger pick-up and drop-off service capacity. Subsequently, in S6, waiting risk indicators and congestion delay indicators are used as optimization drivers to solve for the appropriate decision set, which is then distributed to the dispatch terminal and vehicle terminal for execution. After execution, effect data such as actual headway, actual stop time, changes in waiting number, and changes in vehicle passenger load are collected to update parameters in the waiting risk function, congestion-stop delay coupling function, or ensemble optimization model (e.g., execution jitter statistics, congestion friction coefficient, threshold, and constraint tightness parameters) to support online calculation for the next rolling cycle.

[0068] In this embodiment, step S4 is specifically implemented as follows: Within the statistical time window W, based on the station waiting number sequence Q... i (t) and boarding record B i (t) Calculate the station passenger arrival intensity λ per unit time. i , where λ i The preferred unit is "people / minute"; when both waiting and boarding records are available, λ i The preferred method is to calculate using the conservation approximation, which is obtained by dividing the sum of the increase in the number of people waiting for the train and the number of people boarding the train by the time interval, and then applying λ. i Exponential smoothing is performed to suppress noise, with a smoothing coefficient of 0.2–0.5. When only card / QR code boarding is available, the number of new boarders between two vehicle arrivals is counted and converted into arrival intensity per unit time. Prior compensation can be made by incorporating historical arrival intensity data from the same week and time period. Departure interval h. i The time difference between the departure times of two adjacent trains is used, with the preferred unit being "minutes"; to enhance resilience against anomalies, h i The median of the head-time distance samples within the time window W is preferred, but the mean can also be used. The intensity of travel time fluctuation between adjacent stations is obtained from the road segment length of adjacent stations, the average vehicle speed, and its fluctuation: road segment length L i−1,i Average speed v (obtained from GIS road network, unit: meters) i−1,i With the standard deviation of velocity σ v,i The speeds of multiple vehicles within the road segment are statistically analyzed (in meters per second), and combined with the filtering residuals from the preprocessing stage to adjust σ. v,i Corrections are made to remove measurement noise. A jitter term is included to characterize random disturbances in scheduling execution and arrival / departure processes, with its standard deviation σ. s,i The preferred method is to obtain the standard deviation of the difference between the planned departure time and the actual departure time within a time window W; if there is no planned timetable, the standard deviation of the error sequence between the target departure interval and the actual departure interval generated in the previous cycle can be used instead. Platform capacity P iThe maximum number of people a station can accommodate (per person) can be determined by operational configuration or calculated based on platform area and safety density (selectable density: 2-4 people / square meter). Finally, the waiting risk is normalized and output as a waiting risk indicator, facilitating comparisons across stations and lines and incorporating it into the S6 optimization solution.

[0069] Specifically, the determination of the waiting area risk index satisfies the following equation:

[0070]

[0071]

[0072]

[0073] in:

[0074] X i λ is the waiting risk indicator for station i; i h represents the passenger arrival intensity at station i. i P represents the average departure interval at station i; i Let i be the platform capacity of station i;

[0075] L i−1,i v is the length of the road segment from adjacent station i−1 to station i; i−1,i With σ v,i These are the average driving speed and the speed standard deviation of the aforementioned road segment, respectively.

[0076] σ T,i σ is the standard deviation of the travel time for the aforementioned road segment; s,i The standard deviation of the planned jitter term; σ h,i Let be the standard deviation of the head-time distance fluctuation intensity at station i.

[0077] In this embodiment, step S5 is implemented as follows: The effective passenger capacity of the vehicle is determined by the vehicle's rated passenger capacity and the upper limit of the comfortable occupancy rate: Vehicle rated passenger capacity C rated As provided in the vehicle file, the upper limit of the comfort occupancy rate η is configured by the operating rules, and η can be selected from 0.85 to 1.0. Therefore, the effective passenger capacity Ci = ηC rated (Unit: Person). Passenger capacity n in the vehicle. on,i The optimal passenger flow count is the cumulative number of passengers boarding and alighting. If the count drifts, it can be corrected at the terminal station using a closed-loop calibration constraint that "the number of passengers after alighting should be close to 0" (e.g., if the terminal deviation is >5 people, the correction should be evenly distributed across the trip based on mileage or number of stops). Passenger boarding and alighting service capacity parameters should include at least the boarding service rate μ. b With disembarkation service rate μ a The unit is people per minute; μ b ,μ aOptional methods of obtaining the data: First, based on historical stop data fitting, in low-crowding samples (passenger load ratio <0.6), the stop time is linearly regressed on the number of passengers getting on and off, and the inverse of the slope is used as an estimate of the service rate.

[0078] Specifically, the congestion delay index is determined by the following equation:

[0079]

[0080]

[0081]

[0082]

[0083] in:

[0084] Y i Here, n is the congestion delay indicator for station i; b,i With n a,i These represent the estimated number of bicycles boarding and alighting at station i, respectively; θ i The percentage of passengers disembarking;

[0085] n on, i represents the passenger capacity inside the vehicle before it arrives at station i; C i μ represents the vehicle's effective passenger capacity at station i at time i; b With μ a These represent the boarding service rate and the alighting service rate per unit time, respectively; τ 00 κ represents the fixed time term for vehicle entry into the station, door opening and closing, and departure; κ is the congestion friction coefficient.

[0086] T d,i Let be the dwell time of station i.

[0087] In this embodiment, the decision set at the wire mesh level uses a decision variable vector. express, At least include: departure intervals for each line (or its adjustment amount), number of vehicles deployed on each route (or its increase or decrease), and optional discrete action variables for cross-line coordinated replenishment or depot vehicle allocation; discrete action variables can be expressed as "allocating k vehicles from depot s to the line". The term is represented in integer form and is constrained by the number of available vehicles at the depot. Network-level constraints include at least: fleet size constraint (total number of available vehicles is N), depot available vehicle constraint, and upper and lower limits constraint on departure interval h. min ,h maxThe configuration is determined by operational rules, prioritizing peak-hour intervals of 2-8 minutes and off-peak intervals of 5-15 minutes, along with road capacity constraints. The required road density / congestion level is specified within the road capacity constraints. It can be derived from road condition sources or vehicle speeds; the capacity limit function qmax(ρℓ) can be obtained by looking up the corresponding maximum allowable departure frequency according to the congestion level (0-5) or by fitting it based on historical flow, speed, and density data. The fitted parameters are updated weekly or monthly. The adaptive update of constraint tightness is implemented online as follows: when a significant decrease in road speed and increased queuing are detected for M consecutive rolling cycles (optional 3-10), the constraint tightness is reduced. The effective upper limit (can be lowered by 5% to 20%); when the waiting risk continues to exceed the limit for M consecutive cycles and there is a surplus of vehicle resources, the effective value of the proportion of dispatchable vehicles or the upper limit of departure interval can be increased without exceeding the actual available vehicle boundary, so as to avoid long-term unreachable solutions caused by fixed constraints; the above updates are all used as updates to the constraint tightness parameters in ensemble optimization, without changing the optimization structure.

[0088] Specifically, the solution to the wire mesh-level ensemble optimization model satisfies the following equation:

[0089]

[0090]

[0091]

[0092]

[0093] Let be the vector of network-level decision variables; u is the feasible region that satisfies the constraints of depot, fleet size, and operation; S is the set of depots; and L is the set of routes. and Decision-making The following are the waiting risk indicators and congestion delay indicators; X max With Y max These are the maximum allowed limits;

[0094] and The lines are respectively In decision-making The departure interval and number of vehicles deployed are specified below; N is the total number of available vehicles. In order to match road density The relevant capacity limit function is used to constrain the frequency of train departures on the route to not exceed the road's carrying capacity.

[0095] The upper limit threshold is obtained by selecting a safe sample set based on historical calibration data, where no platform congestion interventions have occurred / no full-load skipping of stations or significant train congestion has taken X values ​​respectively. i With Y i The 90th to 95th quantiles can be used as the threshold; or the threshold can be derived inversely based on the service level target (e.g., using the constraint that "the number of people waiting does not exceed a certain proportion of the platform capacity and the waiting time does not exceed the target value" to deduce X). max Y is derived by using the constraint that "the passenger load ratio does not exceed the comfortable full load rate and the proportion of stops to intervals does not exceed the target value". max It allows configuration by line and time period. The online solution employs a feasible iterative algorithm: using... Using initial values, calculate the values ​​for each station in the current period. , And form the target value; for continuous variables (such as...) , The gradient can be approximated using analytical dependency or finite difference, and local enumeration or neighborhood search can be used for discrete allocation variables; updates are made with a step size η. The iterations are then projected onto the feasible region u (repairing the violated interval upper and lower limits, total vehicle volume, and road frequency constraints to the boundary). The number of iterations can be selected from 10 to 50, or the upper limit of the computation time can be selected from 0.5 to 2 seconds. Output is given after the stopping condition is met. Output This is further transformed into a set of adaptable decisions that can be issued: including adjusting departure intervals (forming a new departure plan), adding or removing trains (forming vehicle deployment / retrieval instructions), and cross-line coordinated replenishment or depot vehicle allocation (forming allocation instructions). To illustrate how the final optimized value affects system operation, taking the calculation results of the aforementioned example station as an example, if a certain line calculates Y in the current period... i Exceeding Y max The optimizer will shorten the line. Or add this line (Within the constrained feasible region) make n b,i =λ i h i Decrease or make n on,i By diverting traffic down, the dwell time T is reduced. d,i Compared to the crowding ratio, this ultimately reduces the over-limit squared penalty and outputs an executable decision.

[0096] like Figure 2In this embodiment, the system includes a processor, a memory, and a multi-source data interface communicatively connected to the processor. The memory stores a computer program, which, when executed by the processor, enables the system to complete data aggregation, calculation, and distribution on a rolling cycle, forming a closed-loop update. The multi-source data interface is used to interface with vehicle terminals, station counting devices, traffic service, and operation rules and resource configuration libraries. The multi-source data interface can be implemented using message subscription, queues, or interface calls. The collected raw data includes at least the following fields: the number of people waiting at stations or their estimates, boarding records, vehicle location and speed, arrival and departure times and door opening / closing events, average speed and speed fluctuation of the road segment, platform capacity, minimum and maximum allowable departure intervals, number of vehicles available in the fleet and depot, and weather conditions. The raw data is written to a cache or time-series storage for subsequent calculations.

[0097] The system includes a multi-source data acquisition module for real-time acquisition of multi-source operational data of the bus network through the multi-source data interface; a data preprocessing module for aligning the data from each source to a unified time grid and performing noise filtering, outlier removal, missing data completion, timestamp alignment, and feature standardization, outputting a rasterized feature package. Abnormal waiting numbers, vehicle position jumps, and speed anomalies are corrected by filtering or interpolation. Short-term missing data is preferably filled by extrapolation, and long-term missing data is preferably filled by historical statistics from the same time period, with confidence markers added to the completion results; and a deep fusion module for performing cross-source feature alignment and joint representation generation on the preprocessed heterogeneous data, outputting a unified representation vector. The joint representation generation can be implemented using a multi-layer attention coding network or a gated concatenation network.

[0098] The system also includes a waiting risk modeling module, used to obtain station passenger flow arrival intensity, line departure interval, travel time fluctuation intensity caused by road conditions, and execution jitter statistics representing random disturbances in scheduling execution and arrival / departure processes based on the representation vector and the rasterized feature package, and to calculate waiting risk indicators according to the waiting risk function, wherein the travel time fluctuation intensity is obtained by converting the length of adjacent station road segments, average speed, and speed fluctuation degree, and the execution jitter statistics are obtained by the difference between the planned departure time and the actual departure time or by the statistics of the error between the target departure interval and the actual departure interval; a delay coupling modeling module, used to obtain the station passenger flow arrival intensity, line departure interval, travel time fluctuation intensity caused by road conditions, and execution jitter statistics based on the representation vector and the rasterized feature package, and to calculate waiting risk indicators according to the waiting risk function, wherein the travel time fluctuation intensity is obtained by converting the length of adjacent station road segments, average speed, and speed fluctuation degree, and the execution jitter statistics are obtained by converting the length of adjacent station road segments, average speed, and speed fluctuation intensity ... The representation vector and the rasterized feature package are used to obtain parameters such as the vehicle's effective passenger capacity, in-vehicle passenger volume, alighting ratio, and passenger boarding / alighting service capacity. The congestion delay index is calculated according to the congestion-stop delay coupling function. The vehicle's effective passenger capacity is determined by the rated passenger volume and the upper limit of the comfort occupancy rate. The passenger boarding / alighting service capacity parameters include at least the boarding service rate and the alighting service rate. The boarding service rate and the alighting service rate can be obtained by fitting historical stop data or by converting the average boarding / alighting time per passenger. The parameter representing congestion friction delay in the coupling function is obtained by regressing the historical stop time residuals onto the congestion ratio and the total number of passengers boarding / alighting.

[0099] The system also includes a network-level ensemble optimization solution module, which aggregates waiting risk indicators and congestion delay indicators on the station set. Under constraints of fleet size, available vehicles at the station, upper and lower limits of departure intervals, and road capacity, it constructs and solves the network-level ensemble optimization model, outputs the optimal decision variables, and generates an adaptive decision set. The adaptive decision set includes at least departure interval adjustment, adding or removing vehicles, cross-line coordinated replenishment, or station vehicle allocation. The road capacity limit required in the road capacity constraint can be obtained by looking up a table based on the congestion level or by fitting historical flow-speed-density data, and allows adaptive updates to the constraint tightness based on the execution effect. The system also includes a decision distribution and feedback update module, which distributes the adaptive decision set to the dispatch terminal and vehicle terminal and triggers execution. Simultaneously, it collects effect data such as actual headway, actual stop time, changes in waiting passengers, passenger load changes, and full-load skipping events after execution. Based on this effect data, it updates the execution jitter statistics in the waiting risk modeling module, the congestion friction delay parameters in the delay coupling modeling module, and the threshold parameters and constraint tightness parameters in the network-level set optimization solution module to form a closed-loop adaptive update. When external road condition data is missing or the deep fusion module is unavailable, the system can degenerate into directly calculating arrival intensity, departure interval, and speed fluctuation based on preprocessed basic features and continue to complete the waiting risk index, congestion delay index, and network-level set optimization solution, thereby ensuring that the system can still output an executable adaptive decision set even in the event of abnormal data sources.

Claims

1. A bus dynamic capacity adaptation method based on multi-source data fusion, characterized in that, Includes the following steps: S1. Real-time collection of multi-source operation data of the bus network, wherein the multi-source operation data includes at least station passenger flow data, vehicle operation data, road condition data, and environmental and operational constraint data; S2. Preprocess the multi-source operational data, including at least noise filtering, outlier removal, timestamp alignment, and feature standardization. S3. Perform deep fusion on the preprocessed multi-source operational data and output a unified representation vector; S4. Based on the representation vector, extract the station passenger flow arrival intensity, line departure interval, and travel time fluctuation intensity caused by road conditions, and construct a waiting risk function accordingly to obtain the waiting risk index. S5. Based on the representation vector, extract the parameters of the station passenger flow arrival intensity, the line departure interval, and the vehicle's effective passenger carrying capacity and passenger boarding / alighting service capacity, and construct a congestion-stop delay coupling function accordingly to obtain the congestion delay index. S6. Based on the waiting risk index and the congestion delay index, construct a network-level capacity adaptation set optimization model, solve for the adaptation decision set including at least departure interval adjustment, adding or removing vehicles, cross-line coordination or station vehicle allocation, and send the adaptation decision set to the dispatch terminal and vehicle terminal. Collect the effect data after execution to update the parameters of the waiting risk function, the congestion-stop delay coupling function or the set optimization model.

2. The method for dynamic bus capacity adaptation based on multi-source data fusion according to claim 1, characterized in that, Step S4 specifically includes: Within a preset statistical time window, the station passenger flow arrival intensity per unit time is calculated based on the station waiting number sequence and boarding records. Calculate the departure interval of the line at the station based on the departure time difference between two adjacent trains; Based on the length of the road segment between adjacent stations, the vehicle speed and its fluctuations, and combined with the filtering residuals from the preprocessing stage, the fluctuation intensity of the travel time between adjacent stations is calculated, and a plan execution jitter term is introduced to characterize the random disturbances in the scheduling execution and arrival / departure processes. Based on the platform capacity or the maximum number of people that can be accommodated, the waiting risk is normalized, and the waiting risk index is output.

3. The method for dynamic bus capacity adaptation based on multi-source data fusion according to claim 2, characterized in that, The determination of the waiting area risk indicators satisfies the following relationship: The fluctuation intensity of travel time is calculated from the length of the road segment between adjacent stations, the average travel speed, and the speed fluctuation; the headway fluctuation intensity is obtained by combining the fluctuation intensity of travel time with the planned execution jitter term. The estimated arrival volume of a single train is calculated based on the station passenger flow intensity and departure interval, and the second-order statistic of waiting time is determined by combining the average level of departure interval and the intensity of head-time distance fluctuation. The second-order statistic of the waiting time is then coupled with the passenger flow arrival intensity at the station and normalized by the platform capacity to obtain the waiting risk index. The waiting risk increases with the increase of passenger flow arrival intensity at the station, decreases with the increase of platform capacity, and increases nonlinearly with the increase of departure interval and the increase of headway fluctuation.

4. The method for dynamic bus capacity adaptation based on multi-source data fusion according to claim 3, characterized in that, Step S5 includes: Obtain the vehicle's rated passenger capacity and real-time passenger capacity, and combine them with the upper limit of the comfort occupancy rate to determine the vehicle's effective passenger capacity; The passenger boarding and alighting service capacity parameters are determined based on the number of doors, passage conditions, or average boarding and alighting time per passenger. These parameters include at least the boarding service rate and the alighting service rate. The estimated number of passengers boarding a single vehicle at the station is calculated based on the passenger flow intensity at the station and the departure interval, and the estimated number of passengers alighting is calculated based on the alighting ratio. Based on parameters such as vehicle effective passenger capacity, in-vehicle passenger volume, passenger boarding and alighting service capacity, and expected boarding and alighting volume, a congestion-stopping delay coupling function is constructed to output a congestion delay index.

5. The method for dynamic bus capacity adaptation based on multi-source data fusion according to claim 4, characterized in that, The determination of the congestion delay index satisfies the following relationship: The estimated boarding volume is obtained by multiplying the station passenger flow intensity by the departure interval; the estimated alighting volume is determined by the alighting ratio and the estimated boarding volume; the dwell time is obtained by superimposing fixed time items such as vehicle entry and door opening / closing, the ratio of estimated boarding volume to boarding service rate, the ratio of estimated alighting volume to alighting service rate, and a congestion friction delay item related to the degree of congestion inside the vehicle. The congestion friction delay item increases as the proportion of passenger volume inside the vehicle to the vehicle's effective passenger capacity increases, and also increases as the total number of passengers boarding and alighting increases; the congestion delay index is composed of the degree of congestion after arrival and the relative proportion of dwell time to departure interval, wherein the degree of congestion is determined by the relationship between passenger volume inside the vehicle, estimated boarding volume, estimated alighting volume, and the vehicle's effective passenger capacity.

6. The method for dynamic bus capacity adaptation based on multi-source data fusion according to claim 5, characterized in that, Step S6 includes: Construct a network-level decision set to be solved, which includes at least the departure interval adjustment amount of each line, the number of standby vehicles deployed or recalled, cross-line collaborative replenishment strategy, or depot vehicle allocation strategy. Construct network-level constraints, which include at least fleet size constraints, station available vehicle constraints, departure interval upper and lower limits constraints, and road capacity constraints. The network-level decision set is solved by jointly using the waiting risk index and the congestion delay index, and the constraint tightness is adaptively updated based on the execution effect data to suppress scheduling instability under congestion fluctuation conditions.

7. The method for dynamic bus capacity adaptation based on multi-source data fusion according to claim 6, characterized in that, The solution to the wire mesh-level ensemble optimization model satisfies the following relationship: The optimization cost is set by taking the degree of exceeding the limits of waiting risk indicators and congestion delay indicators on the station set as the objective, and by taking the upper and lower limits of departure interval, total number of vehicles, available vehicles at the station, and the upper limit of road capacity related to road density or congestion level as constraints. The optimization is carried out on the network-level decision set to obtain the optimal decision that satisfies the constraints and minimizes the cost of exceeding the limits. Based on the optimal decision, an adaptation decision set is generated, which includes at least departure interval adjustment, adding or removing vehicles, cross-line collaborative supplementation, or station vehicle allocation.

8. A public transport dynamic capacity adaptation system based on multi-source data fusion, characterized in that: This method is used to implement the public transport dynamic capacity adaptation method based on multi-source data fusion as described in any one of claims 1-7. The system includes a processor, a memory, and a multi-source data interface communicatively connected to the processor. The memory stores a computer program, which, when executed by the processor, includes the following functional modules: The multi-source data acquisition module is used to collect multi-source operational data of the public transport network in real time. The data preprocessing module is used to perform noise filtering, outlier removal, timestamp alignment, and feature standardization on the multi-source running data. The deep fusion module is used to perform deep fusion on the preprocessed data and output a unified representation vector. The waiting risk modeling module is used to extract the station passenger flow arrival intensity, line departure interval, travel time fluctuation intensity and plan execution jitter based on the representation vector, and construct the waiting risk function to obtain the waiting risk index. The delay coupling modeling module is used to extract parameters such as station passenger flow arrival intensity, line departure interval, vehicle effective passenger capacity, in-vehicle passenger volume, alighting ratio, and passenger boarding / alighting service capacity based on the representation vector, and to construct a congestion-stop delay coupling function to obtain congestion delay index. The network-level set optimization solution module is used to construct and solve the network-level set optimization model based on the waiting risk index and the congestion delay index, and generate an adaptive decision set under the conditions of satisfying the constraints of fleet size, available vehicles at the station, upper and lower limits of departure interval and road capacity. The decision distribution and feedback update module is used to distribute the adapted decision set to the scheduling terminal and vehicle terminal and trigger execution, collect the effect data after execution, and update the model parameters, threshold parameters or constraint tightness parameters of the waiting risk modeling module, the delay coupling modeling module or the network-level set optimization solution module based on the effect data.

Images