A method and system for bus departure scheduling

Abstract

This invention relates to the field of public transportation system technology, specifically disclosing a bus departure scheduling method, comprising the following steps: S1, integrating multi-dimensional data; S2, calculating future passenger demand using a dynamic prediction model incorporating machine learning parameters; S3, calculating the required number of vehicles based on predicted demand, vehicle capacity, and dynamic load coefficient; S4, constructing a multi-objective function including energy consumption optimization to solve for the optimal departure interval and route; S5, revising the scheduling plan based on real-time data, collecting real-time feedback data through passenger mobile applications, analyzing passenger emotions and needs using natural language processing technology, recognizing feedback intent based on an emotion analysis model, constructing a passenger demand knowledge base by combining historical complaint data, and optimizing the dynamic prediction model and scheduling strategy. Through technological integration and system innovation, this method breaks through the static and singular bottlenecks of traditional scheduling, providing an intelligent solution for urban public transportation that balances efficiency, low carbon emissions, and user experience.

Description

Technical Field

[0001] This invention relates to the field of public transportation system technology, specifically to a bus departure scheduling method and a bus departure scheduling system. Background Technology

[0002] In the field of urban public transportation, traditional bus dispatching methods generally suffer from the following technical bottlenecks:

[0003] 1. Limited data dimensions: Relying solely on historical passenger flow data or fixed timetables, the system lacks the ability to integrate multi-dimensional data such as real-time passenger requests, traffic conditions, environmental factors (e.g., heavy rain, high temperatures), and planned events (e.g., concerts, commuting peaks), leading to a disconnect between demand forecasts and actual passenger flow fluctuations.

[0004] 2. Static Prediction Models: Existing prediction models are mostly based on fixed parameters or simple statistical methods, which cannot dynamically adapt to the switching between holiday / weekday modes, the passenger flow transfer patterns of adjacent stations, and the impact of time and space events (such as sudden traffic congestion or a surge in passenger flow during large-scale events), resulting in scheduling plans lagging behind actual needs.

[0005] 3. Insufficient real-time response capability: The lack of a real-time decision-making mechanism that combines edge computing and cloud collaboration makes it difficult to respond to traffic events (such as traffic accidents) or emergency passenger feedback (such as vehicle breakdowns) on a minute-level time scale. Temporary dispatching relies on manual intervention, which is inefficient.

[0006] 4. Lack of multi-objective optimization: Traditional scheduling only takes passenger waiting time or operating cost as a single objective, without deeply coordinating energy consumption optimization, carbon emission reduction and route planning, which cannot meet the needs of green transportation development; at the same time, it does not consider the dynamic impact of the battery status of new energy vehicles on the load, which may lead to problems such as vehicles running out of power midway.

[0007] 5. Low passenger participation: Passenger feedback data (such as congestion complaints and route suggestions) has not been effectively integrated into the scheduling optimization process. There is a lack of a real-time feedback response mechanism based on natural language processing and sentiment analysis, making it difficult to form a closed-loop optimization of "demand perception - scheduling adjustment - effect verification". Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a bus departure scheduling method and a bus departure scheduling system, which solve the problems mentioned in the background.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a bus departure scheduling method, comprising the following steps:

[0010] S1. Integrate multi-dimensional data: collect real-time passenger request data, vehicle location information, traffic status data, environmental factor data, planned event logs, smart card swipe data, and subway connection station passenger flow data, and achieve second-level data synchronization through 5G communication technology;

[0011] S2. Calculate passenger demand for future periods using a dynamic prediction model that includes machine learning parameters;

[0012] S3. Calculate the required number of vehicles based on the predicted demand, vehicle capacity, and dynamic load factor;

[0013] S4. Construct a multi-objective function that includes energy consumption optimization, and solve for the optimal departure interval and route;

[0014] S5. Adjust the scheduling plan based on real-time data, collect real-time feedback data through passenger mobile applications, analyze passenger emotions and needs using natural language processing technology, identify feedback intent based on sentiment analysis model, build a passenger demand knowledge base by combining historical complaint data, and optimize dynamic prediction model and scheduling strategy.

[0015] Preferably, the dynamic prediction model described in S2 calculates the passenger demand for future periods using the following formula. :

[0016] ;

[0017] in: The real-time change in passenger requests per unit of time, expressed in passenger trips per minute; The time interval for prediction is in minutes; Dynamically optimized by real-time machine learning algorithms; For the site exist The spatial influence weighting factor of time is used to learn the probability of passenger flow transfer between adjacent stations through a spatiotemporal graph neural network, with a value range of... ; To measure the real-time social media check-in density within the region, natural language processing is used to extract activity keywords and quantify demand fluctuations. To determine the concentration of shared bicycle riding destinations, we can characterize the potential demand for public transportation connections. The attention mechanism weight matrix dynamically allocates the importance of each feature; the model supports dual-mode switching between holidays and weekdays, and automatically identifies the mode switching threshold through an incremental learning algorithm.

[0018] Preferably, the dynamic prediction model employs the logistic suppression function. The formula is:

[0019] ;

[0020] in: The congestion baseline coefficient is set to a value of [value missing]. , representing the normalized threshold of historical average congestion level; environmental factor data Traffic status data The coupling effect is dynamically calibrated by a fuzzy logic controller that includes an extreme weather emergency rule base, when When heavy rain is detected, the express train mode is automatically triggered to skip low-passenger-flow stations; the congestion level is identified based on the YOLOv8 object detection algorithm, and the congestion coefficient is output. ; LSTM models pre-trained using transfer learning techniques are used to analyze vehicle sensor data and generate environmental sensitivity coefficients. ; Fuzzy logic controller input Adjust parameters in real time .

[0021] Preferably, the dynamic prediction model employs a Gaussian kernel event influence function. The formula is:

[0022] ;

[0023] in:

[0024] The Euclidean distance between the location of the incident and the bus stop;

[0025] Spatial attenuation coefficient, unit: (This refers to training different event types using graph neural networks.) value;

[0026] As an event type influence factor, The time when the event occurred. This represents the standard deviation of time decay.

[0027] Preferably, in S3, the required number of vehicles is calculated using the following formula:

[0028] ;

[0029] in: For vehicle capacity;

[0030] ;

[0031] The baseline load factor represents the benchmark for the ideal full load rate of the vehicle. The level of crowding inside the vehicle is calculated using pressure sensors and camera image recognition. Passenger satisfaction scores are input after normalization. For new energy vehicle batteries, the state factor is determined by the battery management system (BMS) which collects the remaining battery power in real time. hour, Reduce the load limit; For passenger flow prediction confidence, when A conservative scheduling strategy is triggered, increasing the number of reserve vehicles by 10%; weights are dynamically adjusted through reinforcement learning. .

[0032] Preferably, the multi-objective function in S4 is:

[0033] ;

[0034] in: Total passenger waiting time, expressed in person-minutes, is the sum of the time from request to boarding for all passengers. Total operating cost of the vehicle; The vehicle's total energy consumption is monitored in real time by onboard sensors. As a baseline energy consumption, , , The weighting coefficients for multiple objectives are dynamically allocated using the analytic hierarchy process (AHP); carbon emission reduction benefits are quantified through regional carbon trading market prices; a hybrid genetic algorithm and particle swarm optimization algorithm are used to solve the problem, and the route is optimized by combining a digital twin simulation module.

[0035] The energy consumption optimization objective is deeply integrated with the route planning:

[0036] Predicting road segment energy consumption using driving behavior analysis models ;

[0037] A digital twin road network is constructed based on real-time traffic flow data, and the energy consumption and running time of candidate routes are simulated.

[0038] Low-energy routes are prioritized while meeting departure interval and passenger waiting time constraints.

[0039] Preferably, in step S5, the passenger feedback data processing flow includes: extracting keywords using natural language processing technology to generate an emotion rating vector. Negative values ​​indicate negative emotions;

[0040] Establish a feedback priority classification model:

[0041] Emergency feedback: Keywords include fault, danger, inability to board, and sentiment value <-0.5. Duplicate feedback is merged through semantic similarity matching algorithm, and temporary dispatch + manual intervention is triggered within 10 minutes.

[0042] Trend-based feedback: Five consecutive similar feedbacks will update the regional prediction model weight by 30% within 2 hours;

[0043] Suggested feedback: Stored in the knowledge graph, and the global model is updated at night using graph embedding technology;

[0044] A feedback processing effect evaluation mechanism is introduced, and the effect of scheduling adjustments is verified through secondary passenger ratings, forming a closed-loop optimization.

[0045] A bus dispatching system includes:

[0046] The edge-cloud collaborative data acquisition module is deployed in IoT sensors, edge computing nodes and roadside units on bus routes to achieve local data preprocessing and low-latency transmission, and supports real-time synchronization with the cloud. A new real-time decision engine is added to the edge layer, which can autonomously generate temporary scheduling plans based on RSU data.

[0047] The intelligent predictive analysis module integrates a dynamic prediction model and uses an incremental learning algorithm to continuously update model parameters to adapt to seasonal demand fluctuations.

[0048] The three-objective optimization scheduling module has a built-in genetic algorithm-particle swarm optimization hybrid solver, which outputs scheduling schemes in real time that include departure interval, route and energy consumption optimization, and supports scheduling at multiple time scales.

[0049] The passenger-participatory service module provides seat availability prediction and dynamic alternative route recommendations through a mobile application, and supports passenger feedback to trigger adjustments to scheduling parameters.

[0050] The green monitoring module monitors vehicle energy consumption data in real time and generates a carbon footprint report by combining traffic conditions and environmental factors.

[0051] The dynamic fare adjustment module, based on the supply and demand matching results, introduces a peak-valley period elasticity coefficient. A volatility of >20% is defined as a peak, and the fare is dynamically adjusted through the Q-learning algorithm.

[0052] Preferably, the collaboration mechanism between the edge computing node and the cloud includes:

[0053] The edge layer is responsible for handling tasks with high real-time requirements and encrypting and uploading sensor data to the blockchain via lightweight blockchain nodes;

[0054] Long-term trend analysis and global optimization are performed in the cloud, and the global model is updated while protecting data privacy through federated learning technology.

[0055] The cloud verifies data integrity through smart contracts, and abnormal data triggers an automatic audit process.

[0056] Preferably, the optimized scheduling module supports multi-timescale scheduling:

[0057] Minute-level response: Responds to real-time demand fluctuations and traffic events, adjusts departure intervals, and supports dynamic skip-stop strategies;

[0058] Hourly level: Based on the dynamic prediction model, vehicles are pre-allocated to hot spots to balance supply and demand, and tidal lanes are adjusted in conjunction with traffic management departments;

[0059] Daily level: Based on urban land use planning data, generate special dispatch plans 72 hours in advance.

[0060] This invention provides a bus departure scheduling method and a bus departure scheduling system, which have the following beneficial effects:

[0061] 1. Accurate prediction driven by multi-dimensional data: Integrating multi-source data such as real-time passenger requests, traffic conditions, environmental factors, social media check-ins, and shared bicycle connection demand, and achieving second-level synchronization through 5G, a dynamic prediction model is constructed. An attention mechanism is used to dynamically allocate the weights of each feature, and a spatiotemporal graph neural network is combined to learn the probability of passenger flow transfer between stations. A Gaussian kernel function is used to quantify the spatiotemporal impact of events, which reduces the demand prediction error by more than 30%. The prediction accuracy is significantly improved, especially in extreme weather (such as rainstorms triggering express train mode at major stations) or large-scale event scenarios.

[0062] 2. Dynamic Load and Multi-Objective Optimization Scheduling: The load coefficient is dynamically adjusted based on factors such as in-vehicle congestion, passenger satisfaction, and battery status. When the battery level is less than 20%, the load limit is automatically reduced to prevent vehicles from running out of power mid-journey. When the prediction confidence level is less than 0.6, conservative scheduling is triggered to add backup vehicles and improve operational reliability. A multi-objective function is constructed that includes passenger waiting time, operating costs, and energy consumption optimization. This function is solved using a genetic algorithm-particle swarm optimization hybrid algorithm and combined with digital twin simulation to select the lowest energy consumption path. It is expected to reduce the energy consumption of a single vehicle by 15%-20%. At the same time, the emission reduction benefits are quantified through the carbon trading market to help the transformation to green transportation.

[0063] 3. Real-time response through edge-cloud collaboration: A real-time decision engine is deployed at the edge layer to autonomously generate temporary scheduling schemes based on roadside unit data (such as adjusting departure intervals by ±2 minutes and dynamically skipping stops), reducing response time to within 10 minutes; the cloud updates the global model through federated learning, protecting data privacy while achieving long-term trend optimization, supporting multi-timescale scheduling at the minute level (real-time demand fluctuations), hour level (pre-allocation of hotspot areas), and day level (land use planning linkage), flexibly responding to scenarios such as commuting peak hours and tidal traffic, and working with traffic management departments to adjust tidal lanes to improve road network traffic efficiency.

[0064] 4. Passenger Participatory Closed-Loop Optimization: Passenger feedback is collected in real time through a mobile application, and emotion scores are generated using natural language processing. Emergency feedback triggers manual intervention and temporary dispatch within 10 minutes; trend feedback updates the regional model weights within 2 hours, enabling dynamic iteration of demand response. A feedback processing effect evaluation mechanism is introduced, and the effect of dispatch adjustments is verified through secondary passenger ratings. Combined with knowledge graph and graph embedding technology, the global model is automatically updated at night, forming a closed loop of perception, analysis, optimization, and verification to improve passenger satisfaction.

[0065] 5. Technological Architecture Innovation and Scalability: The system employs lightweight blockchain encryption to upload sensor data to the chain, and verifies data integrity through smart contracts to ensure the credibility of scheduling decisions. The dynamic fare adjustment module is based on the Q-learning algorithm, which automatically adjusts fares during peak periods when supply and demand volatility exceeds 20%, balancing supply and demand while improving operational revenue. The system is compatible with both new energy vehicles and traditional buses, and supports data linkage with subway connection stations, providing underlying technical support for building a multi-modal collaborative smart transportation system. It possesses strong industry scalability and social benefits.

[0066] In summary, this invention, through technological integration and system innovation, breaks through the static and singular bottlenecks of traditional scheduling, providing urban public transportation with an intelligent solution that balances efficiency, low carbon emissions, and user experience. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of a bus departure scheduling method according to the present invention;

[0068] Figure 2 This is a block diagram illustrating the principle of a bus departure scheduling system according to the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] like Figure 1 As shown, the present invention provides a technical solution: a bus departure scheduling method, comprising the following steps:

[0071] S1. Integrate multi-dimensional data: collect real-time passenger request data, vehicle location information, traffic status data, environmental factor data, planned event logs, smart card swipe data, and subway connection station passenger flow data, and achieve second-level data synchronization through 5G communication technology;

[0072] S2. Calculate passenger demand for future periods using a dynamic prediction model that includes machine learning parameters;

[0073] S3. Calculate the required number of vehicles based on the predicted demand, vehicle capacity, and dynamic load factor;

[0074] S4. Construct a multi-objective function that includes energy consumption optimization, and solve for the optimal departure interval and route;

[0075] S5. Adjust the scheduling plan based on real-time data, collect real-time feedback data through passenger mobile applications, analyze passenger emotions and needs using natural language processing technology, identify feedback intent based on sentiment analysis model, build a passenger demand knowledge base by combining historical complaint data, and optimize dynamic prediction model and scheduling strategy.

[0076] More specifically, the dynamic prediction model described in S2 calculates passenger demand for future periods using the following formula. :

[0077] ;

[0078] in: The real-time change in passenger requests per unit of time, expressed in passenger trips per minute; The time interval for prediction is in minutes; Dynamically optimized by real-time machine learning algorithms; For the site exist The spatial influence weighting factor of time is used to learn the probability of passenger flow transfer between adjacent stations through a spatiotemporal graph neural network, with a value range of... ; To measure the real-time social media check-in density within the region, natural language processing is used to extract activity keywords and quantify demand fluctuations. To determine the concentration of shared bicycle riding destinations, we can characterize the potential demand for public transportation connections. The attention mechanism weight matrix dynamically assigns the importance of each feature; the model supports dual-mode switching between holidays and weekdays, automatically identifying mode transition thresholds (e.g., weekday morning rush hour features are weighted) through an incremental learning algorithm. Keywords related to commuting account for >60%.

[0079] More specifically, the dynamic prediction model employs the logistic suppression function. The formula is:

[0080] ;

[0081] in: The congestion baseline coefficient is set to a value of [value missing]. , representing the normalized threshold of the historical average congestion level, such as Indicates a baseline for moderate congestion; environmental factor data. Traffic status data The coupling effect is dynamically calibrated by a fuzzy logic controller that includes an extreme weather emergency rule base, when When heavy rain (rainfall > 50 mm / h) is detected, the express train mode is automatically triggered to skip low-passenger-flow stations; the congestion level is identified based on the YOLOv8 object detection algorithm, and the congestion coefficient is output. ; LSTM models pre-trained using transfer learning techniques are used to analyze vehicle sensor data and generate environmental sensitivity coefficients. ; Fuzzy logic controller input Adjust parameters in real time .

[0082] More specifically, the dynamic prediction model employs a Gaussian kernel event influence function. The formula is:

[0083] ;

[0084] in:

[0085] The Euclidean distance between the location of the incident and the bus stop;

[0086] Spatial attenuation coefficient, unit: ), trained using graph neural networks for different event types (such as concerts, rush hour commutes), Value (such as large events) =0.001, daily events ( =0.005);

[0087] As an event type influence factor, The time when the event occurred. This represents the standard deviation of time decay.

[0088] In this embodiment, passenger demand in future time periods The calculation is performed through the following steps:

[0089] Step 1: Real-time acquisition and preprocessing of multi-source data:

[0090] Data interface definition: Real-time changes in passenger requests are obtained through the backend interface of the bus APP. (Unit: person-times / minute), according to Minute-cycle statistical increment;

[0091] Historical passenger flow data obtained from smart card ticketing systems The number of passengers boarding the vehicle is calibrated by using onboard pressure sensors and camera image recognition (based on the YOLOv8 target detection algorithm).

[0092] Traffic status data Generated from video stream analysis of roadside units (RSUs), lane vehicle density is detected using the YOLOv8 algorithm, and a congestion coefficient is output. (like To ensure smooth flow, (due to severe congestion)

[0093] Environmental factor data Environmental sensitivity coefficients are output by data collected from vehicle-mounted weather sensors (rain gauges, thermometers, and hygrometers) and processed through a pre-trained LSTM model using transfer learning. (such as heavy rain weather) Sunny weather );

[0094] Planned event data (such as concerts, peak commuting times) is obtained through the city event management platform API, and the event type is analyzed. Time of occurrence and geographical coordinates;

[0095] Social media check-in data The system crawls check-in information from platforms such as Weibo and WeChat within the region, extracts keywords such as "commuting," "office," and "travel" using the BERT natural language processing model, and calculates keyword density (e.g., the proportion of commuting keywords). (Identified as having weekday morning rush hour characteristics).

[0096] Shared bike connection demand By connecting to the API of shared bicycle platforms, the number of cycling destinations within 500 meters of bus stops is counted and normalized. The clustering value of the interval;

[0097] Site space affects weight Trained using a Spatiotemporal Graph Neural Network (STGNN), it takes historical passenger flow data from adjacent stations as input, learns the passenger flow transfer probability matrix between stations, and outputs... (such as between adjacent stations) (Indicates a strong association).

[0098] Step 2: Dynamic weight allocation for the attention mechanism:

[0099] Attention mechanism weight matrix using Transformer architecture The weights of each feature are dynamically adjusted based on the real-time scenario.

[0100] Weekday morning rush hour ( The proportion of keywords related to commuting ):promote ( (weight) and ( (Weight) to 0.3, 0.25;

[0101] Extreme weather (such as) (And heavy rain was detected): Enhanced ( (Weight) reduced to 0.2, while decreasing ( The weight was set to 0.1 to suppress demand at low-traffic stations;

[0102] Holiday pattern (identified via incremental learning algorithm) (The proportion of tourism keywords in China has increased sharply): Improvement ( The weighting is set to 0.2 to reflect changes in connection demand.

[0103] Step 3: Optimize model parameters using machine learning algorithms:

[0104] Parameters are optimized in real time using the Adaptive Gradient Descent (AdaGrad) algorithm. The objective function is the root mean square error (RMSE):

[0105] ;

[0106] The iteration period is The parameters are updated every minute, once per cycle, to ensure the prediction error. .

[0107] Step 4: Switch between holiday / weekday modes:

[0108] Automatically identify mode transition thresholds using incremental learning algorithms:

[0109] Detected in 3 consecutive cycles The proportion of keywords related to commuting and If the system is determined to be in weekday mode, peak scheduling parameters (such as...) will be enabled. , );

[0110] Detected in two consecutive cycles Percentage of keywords related to leisure and It has been determined to be in holiday mode, and the parameters have been adjusted as follows: , .

[0111] Logistic suppression function Implementation details:

[0112] Step 1: Coupling and processing congestion and environmental data:

[0113] Define the congestion baseline coefficient (Corresponding to the historical average level of moderate congestion), a fuzzy logic controller is constructed to handle the situation. and Coupling effects:

[0114] Input variable: Normalized congestion coefficient Environmental sensitivity coefficient ;

[0115] Output variables: Logistic function parameters , The rules can be dynamically adjusted using a fuzzy rule table (example rules are as follows):

[0116]

[0117] Step 2: Extreme Weather Emergency Response:

[0118] when And heavy rain was detected (rainfall amount) When this occurs, the following mechanism is triggered:

[0119] Automatically skip daily passenger flow Reduce stop time at stations with low passenger traffic;

[0120] Passengers are notified in real time via onboard announcements of route adjustments for express buses to major stops, and the dynamic route information is updated simultaneously on the bus app.

[0121] Gaussian kernel event influence function Implementation details:

[0122] Step 1: Quantification of the spatiotemporal impact of the event:

[0123] For each event Calculate its impact on bus stops demand impact :

[0124]

[0125] Parameter definition:

[0126] Event type influencing factor (e.g., concert) , commuting peak Temporary traffic control );

[0127] Standard deviation of time decay (short-term events) Minutes, long-term events minute);

[0128] Spatial decay coefficient, trained using a graph neural network for different event types. Value (large-scale event) Daily events );

[0129] Location and site of the incident Euclidean distance (unit: meters).

[0130] Step 2: Calculation of the cumulative impact of multiple events:

[0131] Accumulate all events for the site Impact:

[0132] ;

[0133] Example: A concert will be held 800 meters from station A. , , If so, the impact value one hour before the start of the event is:

[0134] ;

[0135] This indicates that the event increased the predicted demand for site A by 42%.

[0136] More specifically, in S3, the required number of vehicles is calculated using the following formula:

[0137] ;

[0138] in: Vehicle capacity (60 people for regular buses, 120 people for BRT).

[0139] ;

[0140] The baseline load factor represents the benchmark for the ideal full load rate of the vehicle. The level of crowding inside the vehicle is calculated using pressure sensors and camera image recognition. Passenger satisfaction scores are input after normalization. For new energy vehicle batteries, the state factor is determined by the battery management system (BMS) which collects the remaining battery power in real time. hour, Reduce the load limit; For passenger flow prediction confidence, when A conservative scheduling strategy is triggered, increasing the number of reserve vehicles by 10%; weights are dynamically adjusted through reinforcement learning. .

[0141] In this embodiment, the required number of vehicles is calculated using the following formula:

[0142] ;

[0143] Vehicle capacity is preset based on vehicle type: Regular buses: People; BRT (Bus Rapid Transit): people;

[0144] : Passenger demand for future periods (unit: passenger trips) output by the S2 dynamic prediction model.

[0145] The dynamic load factor reflects the adjustment of vehicle occupancy rate by real-time operating conditions. This is a battery state factor for new energy vehicles, which dynamically adjusts the load limit based on the remaining battery power.

[0146] Dynamic load factor Calculation:

[0147] Parameter acquisition and processing:

[0148] Baseline load factor Defined as the benchmark for the ideal full load rate of a vehicle, with a value range of... ,default value (i.e., an ideal load factor of 80%); optimization can be achieved through historical operational data analysis, such as setting peak hours for morning and evening rush hours. Off-peak period .

[0149] crampedness inside the car The system uses onboard pressure sensors (deployed on the passenger compartment floor) to collect real-time pressure distribution data in the standing area, combined with camera image recognition (based on the YOLOv8 algorithm to detect passenger density), to calculate the average area occupied per person inside the vehicle.

[0150] ;

[0151] Example: The carriage has a maximum capacity of 60 people, and there are currently 45 people. .

[0152] Passenger satisfaction rating Passenger ratings (1-5 stars) are collected through the bus app and normalized. Interval:

[0153] ;

[0154] Example: When the average rating is 3 stars, .

[0155] Passenger flow prediction confidence level Output from the S2 dynamic prediction model, reflecting the reliability of the prediction results (based on historical prediction error statistics), with values ​​ranging from [value missing]. ;when If the prediction result is deemed unreliable, a conservative scheduling strategy is triggered.

[0156] Dynamic optimization of weight parameters: (Crowding weight) (Satisfaction weight) (Confidence weights) are dynamically adjusted using reinforcement learning algorithms (such as Q-Learning), with the goal of minimizing passenger complaint rates and operating costs.

[0157] ;

[0158] Initial value settings: , , .

[0159] Battery state factor Conservative scheduling strategy:

[0160] Battery state factor The remaining battery power of new energy vehicles is collected in real time through the Battery Management System (BMS). Threshold determination logic:

[0161] ;

[0162] when At this time, the load limit is reduced to 10% of the ideal capacity to prevent the vehicle from breaking down due to insufficient power.

[0163] Conservative scheduling strategy: when Automatically increase the number of reserve vehicles by 10% at that time.

[0164] Example: When the calculated base number of vehicles is 10, 11 vehicles are scheduled after the conservative strategy is triggered.

[0165] Example: Calculation of vehicle count during peak hours:

[0166] Scenario: Weekday morning rush hour (7:30-8:30), predicted demand for a certain bus route. Passenger volume, using regular public transportation ( people).

[0167] Parameter values:

[0168] (Peak occupancy rate benchmark);

[0169] (The interior of the vehicle was quite crowded);

[0170] (Passenger satisfaction is moderate);

[0171] (The prediction confidence is reliable and does not trigger a conservative strategy);

[0172] (Power is sufficient) ).

[0173] Dynamic load factor calculation:

[0174] ;

[0175] Number of vehicles required: ;

[0176] Low power scenario scheduling adjustments:

[0177] Scenario: Remaining battery power of a new energy bus (trigger ), forecast demand Passenger volume, using regular public transportation ( people).

[0178] Key parameters: (Off-peak season baseline load factor); ;

[0179] Vehicle count calculation: .

[0180] Note: Due to insufficient battery power, the load limit has been significantly reduced, and the number of vehicles needs to be increased to avoid overloading.

[0181] More specifically, the multi-objective function described in S4 is:

[0182] ;

[0183] in: Total passenger waiting time, expressed in person-minutes, is the sum of the time from request to boarding for all passengers. Total operating cost of the vehicle; The vehicle's total energy consumption is monitored in real time by onboard sensors. As a baseline energy consumption, , , The weighting coefficients for multiple objectives are dynamically allocated using the analytic hierarchy process (AHP); carbon emission reduction benefits are quantified through regional carbon trading market prices; a hybrid genetic algorithm and particle swarm optimization algorithm are used to solve the problem, and the route is optimized by combining a digital twin simulation module.

[0184] The energy consumption optimization objective is deeply integrated with the route planning:

[0185] Predicting road segment energy consumption using driving behavior analysis models ;

[0186] A digital twin road network is constructed based on real-time traffic flow data, and the energy consumption and running time of candidate routes are simulated.

[0187] Low-energy routes are prioritized while meeting departure interval and passenger waiting time constraints.

[0188] In this embodiment, the multi-objective function aims to minimize the overall cost, and the formula is as follows:

[0189] Parameter description:

[0190] Total passenger waiting time :

[0191] Definition: The total time from when all passengers submit their ride request to when they board the bus (unit: person-minutes), calculated by recording the request time and the vehicle arrival time through the bus APP;

[0192] Data source: Real-time request data and vehicle GPS arrival time are synchronized.

[0193] Total vehicle operating cost :

[0194] Composition: Fuel cost (or electricity cost), labor cost, and vehicle depreciation, as shown in the formula:

[0195] ;

[0196] in: Fuel cost per unit mileage (e.g., 0.8 yuan / km for diesel vehicles, 0.3 yuan / km for electric vehicles). For vehicles Mileage; The driver's hourly wage (e.g., 25 yuan / hour); For operating time; Cost per vehicle (e.g., 0.5 yuan / km).

[0197] Total energy consumption of vehicles :

[0198] definition: Real-time monitoring of motor power via onboard sensors Or engine fuel consumption;

[0199] Baseline energy consumption Average energy consumption calculated based on historical operating data of the same time period and route, used to measure the effectiveness of carbon emission reduction.

[0200] Weighting coefficient :

[0201] The steps for dynamically assigning priorities using the Analytic Hierarchy Process (AHP) are as follows:

[0202] Establish a hierarchical structure: target layer (minimize overall cost) and criteria layer (waiting time, operating costs, energy consumption).

[0203] Constructing a judgment matrix: Inviting transportation planning experts to score the importance of the criteria layer (1-9 scale method), such as morning and evening rush hours. Weighted priority ( Off-peak period Weighted priority ( );

[0204] Calculate the weight vector and perform a consistency check, then output the dynamic weights.

[0205] Hybrid algorithm solution process:

[0206] A hybrid framework combining genetic algorithms (GA) and particle swarm optimization (PSO):

[0207] Encoding: Encode the departure interval (minutes) and route node sequence as chromosome / particle positions (e.g., the departure interval is an integer of 5-15 minutes, and the route is a sequence of station IDs).

[0208] Fitness function: Fitness is calculated based on a multi-objective function, and weight normalization is introduced.

[0209] ;

[0210] Genetic operations: crossover (uniform crossover), mutation (random replacement of route nodes), combined with the PSO speed update formula (e.g.) Improve local search capabilities;

[0211] Termination criteria: No significant improvement in fitness for 10 consecutive generations or reaching the maximum number of iterations (e.g., 200 generations).

[0212] Coordination mechanism between energy consumption optimization and route planning:

[0213] Step 1: Driving Behavior Analysis Model:

[0214] Road segment energy consumption prediction model trained based on LSTM neural network:

[0215] Input features: road gradient, traffic light density, real-time vehicle speed, driving habits (frequency of rapid acceleration / deceleration);

[0216] Output: Predicted power (Unit: kW), with the error rate controlled within 8%.

[0217] Step 2: Construction and Simulation of Digital Twin Road Network

[0218] Building a dynamic digital twin road network based on real-time traffic flow data (such as Amap API):

[0219] Nodes: Bus stops and intersections;

[0220] Edge: Road segment attributes (length, speed limit, historical congestion probability);

[0221] Simulation module: Input candidate routes and departure intervals, output energy consumption. Running time Passenger waiting time .

[0222] Step 3: Low-energy path selection:

[0223] Establish energy consumption-time constraints:

[0224] ;

[0225] Prioritize the path that satisfies the constraints and has the lowest energy consumption, for example:

[0226] Standard Route A: Energy consumption 200kWh, running time 45 minutes;

[0227] Low-energy route B: Energy consumption 170kWh, running time 50 minutes (within the time constraint), then choose route B.

[0228] Example: Multi-objective optimization during peak hours:

[0229] Scenario: During the morning rush hour on a weekday (8:00-9:00), a bus route needs to optimize its departure intervals and routes.

[0230] Input parameters: Forecasted demand: 800 people; Weighting coefficient: (Waiting time takes priority) , ;

[0231] Candidate routes: Route 1: Conventional route (12 km, estimated congestion time 15 minutes); Route 2: Low-energy route (10 km, bypassing bridges, no congestion).

[0232] Hybrid algorithm output results: Departure interval: 8 minutes; Optimal route: Route 2 (energy consumption reduced by 18%, waiting time reduced by 25%); Overall cost: 12% lower than traditional scheduling.

[0233] Off-peak energy consumption priority dispatch:

[0234] Scenario: Off-peak hours (10:00-11:00), weight adjusted to (Energy consumption priority).

[0235] More specifically, in S5, the passenger feedback data processing flow includes: extracting keywords using natural language processing techniques (Jieba word segmentation, BERT sentiment analysis model) to generate sentiment score vectors. Negative values ​​indicate negative emotions;

[0236] Establish a feedback priority classification model:

[0237] Emergency feedback: Keywords include fault, danger, inability to board, and sentiment value <-0.5. Duplicate feedback is merged through semantic similarity matching algorithm, and temporary dispatch + manual intervention is triggered within 10 minutes.

[0238] Trend-based feedback: Five consecutive similar feedbacks (such as morning rush hour congestion) will trigger a 30% weighting update to the regional prediction model within two hours.

[0239] Suggested feedback: Stored in the knowledge graph, and the global model is updated at night using graph embedding technology;

[0240] A feedback processing effect evaluation mechanism is introduced, and the effect of scheduling adjustments is verified through secondary passenger ratings, forming a closed-loop optimization.

[0241] like Figure 2As shown, a bus departure scheduling system includes: an edge-cloud collaborative data acquisition module, an intelligent predictive analysis module, a three-objective optimization scheduling module, a passenger participatory service module, a green monitoring module, and a dynamic fare adjustment module. The edge-cloud collaborative data acquisition module is used to deploy IoT sensors (pressure sensors, cameras), edge computing nodes, and roadside units on bus routes to achieve local data preprocessing and low-latency transmission, supporting real-time synchronization with the cloud. A new real-time decision engine is added to the edge layer, which can autonomously generate temporary scheduling plans (adjusting departure intervals by ±2 minutes, dynamically skipping stops) based on RSU data (such as traffic accidents). The intelligent predictive analysis module integrates a dynamic prediction model, continuously updating model parameters using an incremental learning algorithm. To adapt to seasonal demand fluctuations, the three-objective optimization scheduling module incorporates a hybrid solver of genetic algorithm and particle swarm optimization, providing real-time output of scheduling schemes that include departure intervals, routes, and energy consumption optimization, supporting multi-timescale scheduling. The passenger participatory service module provides seat availability prediction and dynamic alternative route recommendations through a mobile application, supporting passenger feedback to trigger scheduling parameter adjustments. The green monitoring module monitors vehicle energy consumption data in real time and generates carbon footprint reports by combining traffic conditions and environmental factors. The dynamic fare adjustment module introduces peak-valley time elasticity coefficients based on supply and demand matching results, defining peak periods as volatility >20%, and dynamically adjusts fares using a Q-learning algorithm (state space: supply-demand ratio, time period, weather; action space: fare ±10%).

[0242] More specifically, the collaboration mechanism between the edge computing nodes and the cloud includes:

[0243] The edge layer is responsible for handling tasks with high real-time requirements and encrypting and uploading sensor data to the blockchain via lightweight blockchain nodes;

[0244] Long-term trend analysis and global optimization are performed in the cloud, and the global model is updated while protecting data privacy through federated learning technology.

[0245] The cloud verifies data integrity through smart contracts, and abnormal data triggers an automatic audit process.

[0246] More specifically, the optimized scheduling module supports multi-timescale scheduling:

[0247] Minute-level response: Responds to real-time demand fluctuations and traffic events, adjusts departure intervals, and supports dynamic skip-stop strategies;

[0248] Hourly level: Based on the dynamic prediction model, vehicles are pre-allocated to hot spots to balance supply and demand, and tidal lanes are adjusted in conjunction with traffic management departments;

[0249] Daily level: Based on urban land use planning data, generate special dispatch plans 72 hours in advance.

[0250] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for scheduling bus departures, characterized by, Comprise the following steps: S1, integrate multi-dimensional data: collect real-time passenger request data, vehicle location information, traffic state data, environmental factor data, planned event log, smart card swipe data, subway transfer site passenger flow data, realize second-level data synchronization through 5G communication technology; S2, adopt a dynamic prediction model containing machine learning parameters to calculate the passenger demand in the future period; S3, according to the predicted demand, vehicle capacity and dynamic load coefficient, calculate the required number of vehicles; S4, construct a multi-objective function containing energy consumption optimization, solve the optimal departure interval and route; S5, according to the real-time data, correct the scheduling scheme, collect real-time feedback data through the passenger mobile application, analyze the passenger emotion and demand by using natural language processing technology, identify the feedback intention based on the sentiment analysis model, combine the historical complaint data to construct the passenger demand knowledge base, optimize the dynamic prediction model and scheduling strategy; The dynamic prediction model in S2 calculates the passenger demand in the future time period by the following formula : ； Wherein: is the real-time passenger request change amount per unit time, with units of person times / minute; is the prediction time interval, with units of minutes; is dynamically optimized by a real-time machine learning algorithm; is the station ; is the spatial influence weight factor at the moment, the passenger flow transfer probability between adjacent stations is learned through a spatio-temporal graph neural network, and the value range is ; is the real-time social media check-in density in the region, the demand fluctuation is quantified by extracting activity keywords through natural language processing; is the shared bicycle riding end gathering degree, which describes the potential demand of public transport transfer; is the attention mechanism weight matrix, which dynamically allocates the importance of each feature; the model supports holiday / weekday dual mode switching, and automatically identifies the mode conversion threshold through incremental learning algorithm; The dynamic prediction model employs a logistic inhibition function with the formula: ； wherein: is the congestion reference coefficient, taking values , representing the normalized threshold of the historical average congestion level; the coupling effect of the environmental factor data and the traffic state data is dynamically calibrated through a fuzzy logic controller containing an extreme weather emergency rule base, when heavy rain is detected, the express station skip mode is automatically triggered, skipping low passenger flow sites; the congestion coefficient is output based on the YOLOv8 target detection algorithm to identify the congestion level ; the LSTM model pre-trained using the transfer learning technology analyzes the vehicle-mounted sensor data to generate the environment-sensitive coefficient ; the fuzzy logic controller inputs real-time adjustment parameters ; The dynamic prediction model employs a Gaussian kernel event influence function with the formula: ； Among them: is the Euclidean distance from the event occurrence to the bus stop; for spatial attenuation coefficients, in units of , values for different event types are trained by a graph neural network ; an event type influence factor, an event occurrence time, a time decay standard deviation; In S3, the required number of vehicles is calculated by the following formula: ； wherein: is the vehicle capacity; ； is a reference load coefficient, representing the ideal state of the vehicle full load rate reference; is the degree of congestion in the vehicle, calculated by pressure sensor and camera image recognition; is the passenger satisfaction score, which is normalized and input; is the new energy vehicle battery state factor, which is collected in real time by the battery management system when the remaining power is , , the upper limit of the load is reduced; is the passenger flow prediction confidence, which triggers the conservative scheduling strategy when , increasing 10% of the standby vehicles; dynamically adjust the weight through reinforcement learning .

2. The bus dispatching method of claim 1, wherein, The multi-objective function in S4 is: ； wherein: is the total waiting time of passengers, in person・minutes, the total time of all passengers from request to boarding; is the total operating cost of vehicles; is the total energy consumption of vehicles, monitored in real time by on-board sensors; is the baseline energy consumption, , , is the multi-objective weight coefficient, dynamically assigning priorities through the analytic hierarchy process; carbon emission reduction benefits are quantified through regional carbon trading market prices; a genetic algorithm, particle swarm hybrid algorithm is used to solve, combined with a digital twin simulation module to optimize the route; The energy consumption optimization target is coordinated with the route planning depth: Predicting road segment energy power through driving behavior analysis model ; Based on real-time traffic flow data, a digital twin road network is constructed to simulate the energy consumption and running time of the candidate route; Prioritize low-energy paths while meeting departure interval and passenger waiting time constraints.

3. The bus dispatching method of claim 2, wherein, In S5, the processing flow of the passenger feedback data includes: extracting keywords through natural language processing technology, and generating an emotion score vector , and a negative value represents a negative emotion. Establish a feedback priority classification model: Emergency feedback: keywords include failure, danger, unable to board, and emotion value <-0.5, merge repeated feedback through semantic similarity matching algorithm, trigger temporary scheduling + manual intervention within 10 minutes; Trend feedback: 5 consecutive feedback of the same type, update the regional prediction model weight by 30% within 2 hours; Suggestion feedback: stored in the knowledge graph, updated globally through graph embedding technology at night; Introduce feedback processing effect evaluation mechanism, verify the scheduling adjustment effect through passenger secondary scoring, form a closed-loop optimization.

4. A bus dispatching system for use in a bus dispatching method according to claim 3, wherein Including: Edge-cloud collaborative data acquisition module, deployed in public transportation line Internet of Things sensors, edge computing nodes and roadside units, realizes local data preprocessing and low-latency transmission, supports real-time synchronization with the cloud, adds real-time decision engine in the edge layer, which can generate temporary scheduling scheme based on RSU data; Intelligent prediction analysis module, integrated with dynamic prediction model, continuously updates model parameters using incremental learning algorithm to adapt to seasonal demand fluctuations; Three-objective optimization scheduling module, built-in genetic algorithm-particle swarm optimization hybrid solver, real-time output of scheduling scheme containing departure interval, route and energy consumption optimization, supports multi-time scale scheduling; Passenger participation service module, provides seat availability prediction and dynamic alternative route recommendation through mobile application, supports passenger feedback triggered scheduling parameter adjustment; Green monitoring module, real-time monitoring of vehicle energy consumption data, combined with traffic state and environmental factors, generates carbon footprint report; Dynamic fare adjustment module, based on supply and demand matching results, introduces peak and valley period elasticity coefficient, defines peak as fluctuation rate >20%, dynamically adjusts fare through Q-learning algorithm.

5. The bus dispatching system of claim 4, wherein, The collaborative mechanism of edge computing nodes and the cloud includes: The edge layer is responsible for processing tasks with high real-time requirements, and the sensor data is encrypted and uploaded to the chain through a lightweight blockchain node; Cloud performs long-term trend analysis and global optimization, and updates the global model while protecting data privacy through federated learning technology; Cloud verifies data integrity through smart contracts, and triggers automatic audit processes for abnormal data.

6. The bus dispatching system of claim 5, wherein, The optimization scheduling module supports multi-time scale scheduling: Minute level: Respond to real-time demand fluctuations and traffic incidents, adjust the departure interval, and support dynamic station skipping strategy; Hour level: According to the dynamic prediction model, pre-allocate vehicles to hot areas, balance supply and demand, and adjust the tidal lane in coordination with the traffic management department; Daily level: Combine with urban land use planning data to generate special scheduling plans 72 hours in advance.

Images