Automatic driving taxi dynamic scheduling system for mixed traffic flow and collaborative decision-making method

Abstract

The invention discloses a mixed traffic flow-oriented automatic driving taxi dynamic scheduling system and a collaborative decision-making method, belongs to the field of intelligent traffic systems, and solves the problem that in the coexistence environment of manual driving vehicles and automatic driving taxies, the automatic driving taxies cannot be automatically scheduled. The technical problem of how to efficiently and cooperatively dispatch vehicles, accurately predict demands, optimize energy management and improve the overall operation efficiency of the system is solved. The system comprises a scheduling server which is connected with a road side sensing unit, a vehicle-mounted control unit and a charging station management platform. The scheduling server comprises a traffic flow analysis module; a demand prediction module; a dynamic scheduling module; and an energy collaboration module. The system is mainly used for realizing real-time, dynamic and intelligent scheduling and energy management of the automatic driving taxis in the mixed traffic flow, improving the operation efficiency, relieving the traffic jam and optimizing the charging resource utilization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation systems. More particularly, the present application relates to a dynamic dispatching system and a collaborative decision-making method for autonomous taxi in mixed traffic flow. BACKGROUND

[0002] In the mixed traffic flow environment, the behavior patterns of human-driven vehicles and autonomous taxis are significantly different, and the existing dispatching system faces three technical bottlenecks. First, there is a coordination obstacle between traffic data collection and dispatching decision-making. The macro traffic flow parameters (such as mixed average speed) obtained by the roadside sensing unit and the micro vehicle state data reported by the on-board unit are difficult to effectively fuse, resulting in that the dispatching instruction lags behind the dynamic traffic changes. The root cause lies in: the random behavior of human-driven vehicles interferes with the trajectory prediction of autonomous vehicles, and the traditional data fusion model does not distinguish the difference in dynamic characteristics of the two types of vehicles; at the same time, the road network level dispatching needs to process massive heterogeneous data, and the existing system cannot meet the real-time response demand due to the limitation of the computing architecture, usually causing the actual road conditions to have changed when the dispatching instruction is issued.

[0003] Secondly, the mixed traffic congestion propagation mechanism lacks precise modeling capability. The traditional traffic flow model (such as macro fluid dynamics model or micro follow-up model) has inherent defects when analyzing mixed traffic flow: the macro model cannot capture the suppression effect of autonomous vehicle cooperative control on congestion propagation, and the micro model is difficult to extend to large-scale road network due to high computational complexity. More importantly, the interaction between human-driven vehicles and autonomous vehicles produces a new type of congestion pattern (for example, disturbance wave superposition caused by human vehicles frequently changing lanes), but the existing model does not establish a mathematical representation method for such mixed interaction, resulting in that the error rate of congestion bottleneck point identification often exceeds 35%, and the propagation path of congestion along the road network cannot be predicted.

[0004] Thirdly, multi-source data conflict and communication failure lead to the decline of system reliability. When there is a significant difference between the traffic parameters (such as headway) of the same road segment reported by the roadside unit and the on-board unit, the existing system usually adopts a fixed priority arbitration mechanism (such as always using roadside data), but this mechanism does not consider dynamic factors such as real-time failure rate of sensors and historical data credibility weight, which is easy to output incorrect road condition diagnosis in bad weather or equipment abnormality. In addition, in the communication delay scenario, some systems try to use static historical data compensation, but ignore the spatiotemporal correlation characteristics of traffic flow (such as the difference in propagation mode between morning and evening peak), resulting in that the output road network state deviates from the actual value by more than 40% in degraded mode. These defects together cause the insufficient robustness of the dispatching system in the mixed traffic environment, especially in the complex topology area of urban road network, which is easy to cause chain decision-making errors. SUMMARY

[0005] An object of the present application is to provide a dynamic dispatching system and a collaborative decision-making method for autonomous taxis in mixed traffic flow, which can realize real-time, dynamic and intelligent dispatching and energy management of autonomous taxis in mixed traffic flow, improve operational efficiency, alleviate traffic congestion and optimize the use of charging resources.

[0006] To achieve these objects and other advantages of the present application, according to one aspect of the present application, a dynamic dispatching system for autonomous taxis in mixed traffic flow is provided, comprising: a dispatching server connected to a roadside perception unit, a vehicle-mounted control unit and a charging station management platform; the roadside perception unit is deployed on each road section in the target area to collect traffic flow data including the mixed driving speed of manually driven vehicles and autonomous taxis in real time; the charging station management platform provides real-time charging pile state data and time-of-use electricity price data; The dispatching server comprises: a traffic flow analysis module for inputting traffic flow data from the roadside perception unit, executing an ant colony pheromone diffusion algorithm to generate a road traffic capacity heat map, and inputting a traffic pulse wave propagation model to output a road network health status diagnosis result; a demand prediction module for fusing historical trip data from a ride-hailing platform, social media event label data and meteorological monitoring data, and generating a 15-minute granular trip demand distribution prediction through a spatio-temporal graph convolutional neural network model; a dynamic dispatching module for generating a dynamic dispatching instruction set including vehicle path re-planning instructions and empty vehicle dispatching coordinates based on the trip demand distribution prediction and the road network health status diagnosis result; an energy coordination module for combining data from the charging station management platform and vehicle battery remaining capacity data, calculating a charging demand elasticity index, dynamically adjusting the charging pricing strategy based on the index, and generating a charging dispatching scheme through a multi-objective programming model of battery degradation cost function and order revenue function; The vehicle-mounted control unit is configured to receive and execute the dynamic dispatching instruction set and the charging dispatching scheme, and is connected to an augmented reality display unit and a blockchain storage unit.

[0007] Preferably, when the traffic flow analysis module executes the traffic pulse wave propagation model, it is specifically configured to: real-time acquisition of basic traffic flow parameters of each road section in the target area, including the mixed average speed of manually driven vehicles and autonomous taxis, the headway and the lane occupancy rate; constructing a road network pulse wave propagation model: abstracting road sections as waveguide units, calculating traffic pressure fluctuation values based on the basic traffic flow parameters, and simulating the propagation process of traffic pressure waves between adjacent road sections through a preset fluctuation equation; Generating road network health diagnosis results: According to the attenuation rate and propagation delay of traffic pressure wave, the road network health state atlas containing road section level congestion index, bottleneck position and propagation path is dynamically output.

[0008] Preferably, the traffic flow analysis module further comprises a data conflict arbitration unit and a communication degradation processing unit, wherein: The data conflict arbitration unit is configured to: when the mixed average speed, headway or lane occupancy rate data reported by the roadside perception unit and the vehicle-mounted control unit on the same road section differs by more than a preset threshold, dynamically select the data source to be used preferentially based on the reliability weight of the historical data source of the road section and the real-time sensor fault diagnosis result; The communication degradation processing unit is configured to: when the communication delay with the roadside perception unit or the vehicle-mounted control unit exceeds a set value, automatically call the road network health state atlas of the same period in history for interpolation compensation, and use the latest effective vehicle flow pulse wave propagation model parameters cached locally by the vehicle-mounted control unit to perform calculation to generate an approximate road network health state diagnosis result.

[0009] Preferably, when the energy coordination module calculates the charging demand elasticity index, the following steps are performed: Real-time acquisition of power grid dynamic parameters: Obtain the time-of-use electricity price data, charging pile idle rate data provided by the charging station management platform, and the battery remaining capacity distribution data of autonomous taxis in the target area; Constructing a CDEI calculation model: dynamically generating a regional charging demand elasticity index based on the following formula: ; Where ΔP t is the electricity price fluctuation value relative to the reference electricity price P base at time t; Q idle,t is the number of idle charging piles at time t; Q total is the total number of charging piles; S critical,t is the number of vehicles with battery capacity below the warning threshold; N vehicle is the total number of autonomous taxis to be dispatched in the target area, which is obtained in real time from the dispatch server; α, β, γ are the weight coefficients of power grid load, pile resource scarcity, and vehicle urgency, respectively; Dynamic adjustment of charging strategy: When CDEI t > θ high , trigger the charging price discount strategy, θ high is the charging demand stimulation threshold; When CDEI t < θ low , start the charging demand suppression strategy, θ low is the charging demand suppression threshold; The linkage charging scheduling scheme: input the adjusted charging strategy into the multi-objective planning model, and optimize to generate a charging instruction set containing a charging pile allocation scheme and a vehicle scheduling path.

[0010] Preferably, the energy coordination module further comprises an elasticity coefficient dynamic correction unit, which is configured to: Real-time monitoring of the actual implementation effect of the charging strategy formulated based on the charging demand elasticity index, calculating the deviation rate of the actual charging capacity after the implementation of the strategy and the predicted charging capacity; When the deviation rate exceeds the preset threshold, automatically dynamically adjusting the grid load weight coefficient, the pile resource scarcity weight coefficient and the vehicle emergency degree weight coefficient according to the grid load mutation indication signal and the actual deviation condition; During the dynamic adjustment of the weight coefficient, the dynamic scheduling module is linked to temporarily freeze the charging pile allocation scheme generated based on the current CDEI t Until the new weight coefficient correction is completed.

[0011] Preferably, the multi-objective planning model of the energy coordination module generates a charging scheduling scheme according to the following steps: Construct a battery cycle attenuation cost function: ; Wherein, C bat (i,j) is the attenuation cost of vehicle i using charging pile j; r j is the charging rate of charging pile j; ΔT j =|T j env -T i opt | is the difference between the ambient temperature and the optimal battery temperature; ai, bi are the battery attenuation coefficients of vehicle i, respectively; Construct an order revenue function: ; Wherein, R order (i,k) is the revenue of vehicle i executing order k; F k is the basic transportation fee of the order; s k is the order mileage; θ is the distance sensitivity parameter, wherein the value of highway is 1.2 and the value of urban area is 0.8; P k is the time premium benchmark value; e k is the emergency level, wherein the value of medical treatment is 3, the value of flight is 2, and the value of ordinary is 1; δ is the emergency amplification factor, taking the value of 1.5; w k is the passenger waiting time; ρ is the waiting penalty coefficient, taking the value of 0.1; Generate a matching scheme: ; Dynamic weight factor: ; wherein, X ijk is a decision variable, taking value 1 when matching and value 0 when not matching; Λ0 is a basic weight, taking value 0.3; Y i is the service life of vehicle i; v is the age gain coefficient, taking value 0.15; t is the age threshold, taking value 2 years; Hard constraints are set as follows: Single-pile constraint, each vehicle can use at most one charging pile; Electricity guarantee, the electricity after charging ≥ the minimum demand of the order + redundancy; Time constraint, charging time + driving time ≤ order deadline.

[0012] Preferably, the energy coordination module further executes a battery temperature real-time compensation mechanism when generating the charging scheduling scheme using the multi-objective planning model, wherein: Real-time receiving of the actual working temperature of the vehicle battery reported by the vehicle control unit; Introducing the difference between the actual working temperature of the battery and the ambient temperature of the charging pile as a compensation factor to dynamically correct the decay cost calculation result; When the difference between the actual working temperature of the battery and the ambient temperature is ≥ 15℃, the maximum charging rate of the charging pile available to the vehicle is forcibly limited.

[0013] Preferably, the dynamic scheduling module is further connected with a road network topology mutation sensing unit, which is configured to: Real-time access to the municipal construction announcement interface and receive abnormal event information reported by the vehicle-mounted camera or the roadside unit; When detecting an event that causes a change in the road network topology structure, such as road closure, lane number change or newly added control area in the target area, generate the corresponding topology mutation event code; Trigger the traffic flow pulse wave propagation model in the traffic flow analysis module to dynamically reconstruct, and adjust the waveguide unit segmentation parameters of the corresponding road segment in the model according to the actual traffic capacity change; Send the path feasibility state flag to the energy coordination module and the dynamic scheduling module to prohibit scheduling vehicles from entering the area that cannot pass due to topology mutation.

[0014] Preferably, the dynamic scheduling module is connected with a meteorological warning interface, and when the thunderstorm warning level is ≥ yellow, it executes: Grouping autonomous taxis in the target area into a fleet; Generating a special traffic waveguide channel for the fleet based on the traffic flow pulse wave propagation model under the condition that the road network topology allows; Synchronizing the acceleration / braking instructions of the fleet vehicles through the vehicle-mounted unit.

[0015] The application also provides a cooperative decision-making method of the automatic driving taxi dynamic scheduling system for mixed traffic flow, which comprises the following steps executed by the scheduling server: Real-time acquisition of mixed traffic flow data collected by the roadside sensing unit, wherein the data includes driving speeds of manually driven vehicles and automatic driving taxis; Fusion of historical travel data of a network car-hailing platform, social media event label data and meteorological monitoring data, input of a spatio-temporal graph convolutional neural network model and output of a 15-minute-granularity travel demand distribution prediction; Based on the driving speeds in the mixed traffic flow data, an ant colony pheromone diffusion algorithm is used to construct a road traffic capacity heat map; The road traffic capacity heat map is input into a traffic flow pulse wave propagation model to diagnose the road network health status, which includes congestion propagation directions and bottleneck section labels; According to the travel demand distribution prediction and the road network health status, a dynamic scheduling instruction set of automatic driving taxis is generated, which includes path re-planning instructions and empty vehicle scheduling coordinates; The charging pile usage state and power grid time-of-use electricity price data sent by the charging station management platform are received, and a charging demand elasticity index is calculated in combination with the vehicle battery remaining capacity; Based on the charging demand elasticity index, the charging service pricing strategy is dynamically adjusted, and a multi-objective programming model is input with a battery attenuation cost function and an order revenue function to generate a charging scheduling scheme; The dynamic scheduling instruction set and the charging scheduling scheme are sent to the vehicle-mounted control unit to trigger the following operations: Control the automatic driving taxi to perform path change or charging station navigation; Through an augmented reality windshield, dynamic display of yielding intention labels, lane change trajectory previews and scheduling decision basis summaries is realized; The vehicle decision timestamp and scheduling basis parameters are written into a blockchain distributed ledger to generate an encrypted verification record updated every 0.1 seconds.

[0016] The application at least has the following beneficial effects: The application provides an automatic driving taxi dynamic scheduling system for mixed traffic flow, which cooperates with a roadside sensing unit, a vehicle-mounted control unit and a charging management platform through a scheduling server, integrates traffic flow analysis, demand prediction, dynamic scheduling and energy coordination modules, realizes real-time fusion processing of mixed traffic flow data and cross-module decision linkage, reduces system response delay, improves empty vehicle scheduling efficiency, and eliminates the conflict risk between charging scheduling and path planning; The vehicle flow pulse wave propagation model is constructed based on the waveguide unit abstraction and traffic pressure fluctuation calculation, can quantize the disturbance propagation effect caused by the random behavior of the manually driven vehicle in the mixed traffic flow, improves the congestion bottleneck point identification accuracy, and the congestion propagation direction prediction error is small; the data conflict arbitration unit dynamically selects the data source according to the historical reliability weight and real-time fault diagnosis, and the communication degradation processing unit adopts the space-time correlation interpolation compensation, so that the system still maintains more than 90% of the road network health state output accuracy under 20% data conflict or 300ms communication delay; The charging demand elasticity index (CDEI) model fuses the power grid load, the scarcity of pile resources and the dynamic weight of vehicle urgency, realizes the elastic regulation and control of charging demand: when CDEI>θ high , the charging amount is increased, and when CDEI<θ low , the power grid peak-valley difference is reduced, and the charging resource squeeze is avoided; the elastic coefficient dynamic correction unit automatically adjusts the weight coefficient by monitoring the actual charging deviation rate (threshold>15%) in real time, and freezes the charging pile allocation scheme during correction, so that the CDEI model prediction deviation rate is stabilized within ±5%; The multi-objective planning model balances the battery attenuation cost and order income function with a dynamic weight factor, prolongs the vehicle battery life and improves the total order income under the guarantee of single pile constraint, power and timeliness hard constraint; the battery temperature real-time compensation mechanism introduces the difference between working temperature and environmental temperature as the attenuation cost compensation factor, and when the temperature difference is greater than or equal to 15 DEG C, the charging rate is limited, so that the battery over-temperature failure rate is reduced; The road network topology mutation perception unit dynamically reconstructs the vehicle flow pulse wave model through the municipal interface and abnormal event detection, and marks the path feasibility state, so that the scheduling path failure times under road closure events are reduced; when the thunderstorm warning is greater than or equal to yellow, the vehicle grouping and special waveguide channel generation mechanism makes the vehicle speed in bad weather increase, and the brake synchronization error is controlled within 0.2 seconds; the collaborative decision method dynamically displays the scheduling basis through augmented reality and stores the blockchain (0.1 second / once update), so that the acceptance of the scheduling decision of passengers is improved, and the dispute tracing time is shortened.

[0017] Other advantages, objects, and features of the application will be apparent from the following specification, and will be appreciated by those skilled in the art. DETAILED DESCRIPTION

[0018] The application will be further described in detail below with reference to specific embodiments, so that those skilled in the art can implement it according to the description.

[0019] It should be understood that the terms such as "have", "contain" and "include" used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0020] It should be noted that the experimental methods in the following embodiments are conventional methods, and the reagents and materials are commercially available unless otherwise specified.

[0021] The application provides a dynamic scheduling system for autonomous taxi in mixed traffic flow, comprising: a scheduling server connected with a roadside sensing unit, a vehicle-mounted control unit and a charging station management platform; the roadside sensing unit is arranged on each road section in a target area to collect traffic flow data including mixed driving speed of artificial driving vehicles and autonomous taxis in real time; the charging station management platform provides charging pile state data and time-of-use electricity price data in real time; The scheduling server comprises: a traffic flow analysis module for inputting traffic flow data of the roadside sensing unit, executing an ant colony pheromone diffusion algorithm to generate a road traffic capacity heat map, and inputting a traffic pulse wave propagation model to output a road network health state diagnosis result; a demand prediction module for fusing historical travel data of a network car-hailing platform, social media event label data and meteorological monitoring data, and generating a 15-minute-granularity travel demand distribution prediction through a spatiotemporal graph convolutional neural network model; a dynamic scheduling module for generating a dynamic scheduling instruction set including vehicle path re-planning instructions and empty vehicle scheduling coordinates according to the travel demand distribution prediction and the road network health state diagnosis result; an energy coordination module for combining data of the charging station management platform and vehicle battery remaining capacity data, calculating a charging demand elasticity index, dynamically adjusting a charging pricing strategy based on the index, and generating a charging scheduling scheme through a multi-objective programming model of a battery attenuation cost function and an order income function; The vehicle-mounted control unit is configured to receive and execute the dynamic scheduling instruction set and the charging scheduling scheme, and is connected with an augmented reality display unit and a blockchain storage unit.

[0022] In the technical solution, the dispatch server can be connected to the RSU road side sensing unit deployed at the intersection through the optical fiber network. The unit uses a millimeter wave radar with a model number of IVP-10 and a camera module to collect mixed traffic flow data in real time, where the speed sampling interval of the manually driven vehicle and the autonomous driving taxi is 0.1 seconds. The charging station management platform can transmit the idle state of the charging pile (the idle rate threshold is set to 15%) and the time-of-use electricity price data (the peak-valley price difference is 0.8 yuan / kWh) through the 4G / 5G communication protocol. The vehicle-mounted control unit is integrated into the vehicle central control system and connected to the projection module of the augmented reality display unit and the encryption chip of the blockchain storage unit through the CAN bus. The system working process is as follows: the road side sensing unit uploads the lane occupancy rate of the target section (accuracy ± 2%) every 2 seconds; the charging station management platform updates the charging pile usage state every minute; the dispatch server integrates data and issues instructions every 5 seconds. In terms of assembly position, the road side sensing unit is installed at a distance of 5.5 meters from the ground on the street lamp pole, and the vehicle-mounted control unit is fixed on the metal bracket below the instrument panel.

[0023] The traffic flow analysis module uses an ant colony pheromone diffusion algorithm to process input data, where the pheromone evaporation coefficient is set to 0.3 and the pheromone enhancement coefficient is 1.2. The algorithm generates a road capacity heat map every 60 seconds, with a resolution of 20 meters / pixel. Based on the fluid mechanics continuity equation: where ρ is the traffic density, v is the average speed of the traffic flow, x is the one-dimensional position coordinate along the traffic flow direction, and t is the time; a speed-density linear relationship is introduced: v = v f (1-ρ / ρ max ); where v f is the free flow speed, i.e., the ideal speed at which the driver can and tends to travel when the traffic density on the road is very low and there is little interaction between vehicles; the traffic pressure fluctuation equation is derived: (β is the compression coefficient); the fluctuation equation is in the form (c is the wave speed, , and the value is 8-12 m / s). The traffic flow pulse wave propagation model uses the above fluctuation equation to output the road network health status diagnosis result. In the working process, the model receives the lane occupancy rate (threshold ≥ 25% to determine congestion) and the headway (threshold ≤ 1.5 seconds to determine high density), and calculates the traffic pressure fluctuation value range to be controlled within -10 kPa to +10 kPa. The module runs on the X86 architecture processor of the dispatch server with a memory configuration of 64GB DDR4.

[0024] The input layer size of the spatio-temporal graph convolutional neural network of the demand prediction module is 1024x1024x8 (lengthxwidthxchannel number), and the input data includes: historical order data of the online car-hailing platform (time slice 15 minutes), concert / sports event tags of microblog social media (keyword matching rate threshold 80%), and precipitation probability data released by the meteorological bureau (threshold > 30% triggers weight doubling). The network output is the travel demand distribution map in the next 15 minutes, and the grid division scale is 500m x 500m. The model training uses the TensorFlow framework and uses a Nvidia Tesla V100 graphics card for acceleration. The training data comes from the 2023 Shenzhen taxi GPS trajectory database, with a total of 120 million records. The prediction results are stored in JSON format on the SSD solid state disk of the dispatch server.

[0025] After receiving the demand prediction and road network health status, the dynamic scheduling module uses the A* algorithm (heuristic function weight 0.7) for path re-planning, and the positioning accuracy of the empty vehicle dispatch coordinates is 10 meters. When calculating the charging demand elasticity index CDEI, the power grid load weight a is 0.4, the pile resource scarcity weight β is 0.3, and the vehicle emergency degree weight γ is 0.3; the charging demand stimulation threshold θ high = 1.2, and the inhibition threshold θ low = 0.8.

[0026] The charging scheduling scheme realizes multi-objective planning through the Gurobi solver. In the battery attenuation cost function, the charging rate r j is limited to the range of 0.5C-2C, the optimal battery temperature T i opt is set to 25℃±3℃. When the vehicle control unit executes the charging instruction, the vehicle with a remaining battery capacity of less than 20% is preferentially allocated to the charging pile. The dynamic scheduling module is deployed in the virtual machine cluster of the dispatch server, and the battery attenuation parameters of the energy coordination module are stored in the encrypted database partition.

[0027] This scheme realizes real-time synchronous processing of mixed traffic flow data, shortens the scheduling response time, reduces the empty running rate of vehicles through multi-module collaborative optimization, and reduces the power grid load fluctuation through dynamic adjustment of the charging strategy. The system shows stable decision-making ability in complex urban road network environment.

[0028] In traditional autonomous taxi dispatching systems, the technical limitations in mixed traffic flow environments mainly manifest in three aspects. First, the road-side sensing unit usually only collects macroscopic traffic flow parameters such as overall average speed, without distinguishing the dynamic behavior differences between manually driven vehicles and autonomous taxis, resulting in the data received by the dispatch server failing to accurately reflect the dynamic characteristics of mixed traffic flow. Second, the charging management module and the vehicle dispatching module operate independently, forming a decision-making island. Charging pile allocation is based solely on static pricing strategies and pile idle status (with a fixed idle rate threshold of 20%), without considering the correlation between road network traffic state and vehicle real-time location. When there is a charging peak and traffic congestion in the same area, the system still guides vehicles to the charging station according to the preset path, which has been found to cause 17% of vehicles to be forced to cancel charging tasks due to road closures. The demand prediction module relies solely on historical travel data and does not integrate social media event tags and weather warning information, resulting in a prediction error of over 40% in heavy rain. Finally, the long response chain of the system causes dispatching delays. The vehicle control unit needs to pass through four independent subsystems to obtain dispatching instructions: road-side data is uploaded to the cloud server (with an average delay of 3 seconds), the traffic analysis module takes 8 seconds to process, and the dispatching instructions are sent to the vehicle unit through the base station (with a delay of 2 seconds). Tests show that the total delay from traffic state changes to vehicle path changes is up to 13 seconds, during which manually driven vehicles have changed the road conditions, resulting in 23% of dispatching instructions being invalid. The modules use HTTP protocol to transmit data, without establishing a dedicated communication channel, and the delay further worsens to more than 18 seconds during network congestion periods.

[0029] In another technical solution, when the traffic flow analysis module executes the traffic pulse wave propagation model, it is specifically configured as follows: Real-time acquisition of basic traffic flow parameters of each road segment in the target area, including the mixed average speed of manually driven vehicles and autonomous taxis, the vehicle headway, and the lane occupancy rate; Construction of a road network pulse wave propagation model: abstract the road segment as a waveguide unit, calculate the traffic pressure fluctuation value based on the basic traffic flow parameters, and simulate the propagation process of traffic pressure waves between adjacent road segments through a preset fluctuation equation; Generation of road network health diagnosis results: based on the attenuation rate and propagation time delay of traffic pressure waves, dynamically output the road network health state atlas containing road segment-level congestion index, bottleneck location, and propagation path.

[0030] In the above technical solution, the traffic flow analysis module obtains the mixed traffic flow parameters of each road segment in the target area in real time, including the average speed (sampling frequency 10 Hz) of manually driven vehicles and autonomous driving taxis, the time headway (measurement accuracy ±0.1 seconds) and the lane occupancy rate (detection error ≤3%). These data are collected by millimeter wave radars deployed on the signal light poles at intersections. The radar installation height can be 5.5 meters from the ground, and the horizontal viewing angle can cover 120 degrees. The average speed threshold is set to 30 km / h, and congestion is determined when the lane occupancy rate exceeds 25%.

[0031] In the data processing stage, the module abstracts each road segment as a waveguide unit with a length of 200 meters. Based on the real-time time headway data, the traffic pressure fluctuation value is calculated, and the calculation process uses the finite difference method to solve iteratively. For example, when the pressure difference between adjacent waveguide units reaches 5 kPa, it is determined to be a fluctuation propagation node. The model updates the propagation state every 5 seconds, and the output results include congestion index classification: 0-3 levels are smooth, 4-6 levels are slow, and 7-10 levels are severe congestion. The calculation task is assigned to a dual-path Xeon server with 128GB DDR4 ECC memory configuration.

[0032] In the traditional mixed traffic flow analysis system, there are significant limitations in modeling the congestion propagation mechanism. Existing technologies usually use macroscopic fluid dynamics models or microscopic vehicle following models to process traffic data, but they are difficult to adapt to the complex environment where manually driven vehicles and autonomous vehicles coexist. For example, a certain city's intelligent transportation platform uses a macroscopic model that treats road traffic flow as a homogeneous fluid, and only outputs congestion index based on average speed (sampling interval 5 seconds) and traffic density. This model cannot distinguish the behavior differences between the two types of vehicles: the disturbance wave generated by the sudden braking behavior of manual vehicles is smoothed in the model, resulting in a 40% error rate in predicting the actual congestion propagation direction. For example, at the interchange merging point, the chain reaction deceleration effect caused by manual vehicles frequently changing lanes is not captured, causing the model output bottleneck position to deviate from the real point by more than 300 meters.

[0033] Microscopic models attempt to track individual vehicle trajectories, but due to computational resource limitations, they can only cover 10% of the road network. A certain commercial system uses the IDM following model, which requires solving differential equations for each vehicle. In an area with 5000 vehicles, the simulation time exceeds 15 minutes, which cannot meet the real-time scheduling requirements. More seriously, the model parameters are fixed and cannot be changed: the maximum deceleration of manual vehicles is uniformly set to -3 m / s 2 , and the maximum deceleration of autonomous vehicles is set to -5 m / s 2 , ignoring the impact of dynamic changes in the proportion of mixed vehicles. Measured data shows that when the proportion of manual vehicles exceeds 60%, the prediction error of the model for congestion propagation delay is as high as ±8 minutes due to the increased complexity of interactive behavior.

[0034] In addition, the traditional system outputs results with coarse granularity. The road network health status is only presented in three levels of "smooth / slow / congestion", without quantifying the propagation path and decay rate. The road condition display screen in the traffic management center of a certain place can only mark the current congestion section, and cannot predict which branch road will be congested in the next 2 minutes. This lack of information directly leads to the dispatch system having a 32% probability of guiding vehicles to a new path that will soon be congested.

[0035] In another technical solution, the traffic flow analysis module further includes a data conflict arbitration unit and a communication degradation processing unit, wherein: The data conflict arbitration unit is configured to: when the mixed average speed, headway or lane occupancy rate data reported by the roadside sensing unit and the vehicle-mounted control unit on the same section differ by more than a preset threshold, dynamically select the data source to be used preferentially based on the reliability weight of the historical data source of the section and the real-time sensor fault diagnosis result; The communication degradation processing unit is configured to: when the communication delay with the roadside sensing unit or the vehicle-mounted control unit exceeds a set value, automatically call the historical road network health status atlas of the same period to perform interpolation compensation, and use the latest effective vehicle flow pulse wave propagation model parameters stored locally by the vehicle-mounted control unit to perform calculation, to generate an approximate road network health status diagnosis result.

[0036] In the above technical solution, the data conflict arbitration unit is deployed on the redundant computing nodes of the dispatch server, and uses a dual-core ARM Cortex-A72 processor. When the lane occupancy rate reported by the roadside sensing unit and the value reported by the vehicle-mounted control unit differ by more than 10% (for example, the roadside data is 35% and the vehicle-mounted data is 20%), the unit starts the arbitration process: first, call the data source reliability weight library of the past 24 hours of the section, the historical reliability weight of the roadside sensing unit is set to 0.6, and the vehicle-mounted unit is 0.4; second, access the real-time fault diagnosis interface, if the temperature drift alarm is returned by the self-checking code of the roadside millimeter wave radar, then automatically increase the weight of the vehicle-mounted data to 0.8. The arbitration result output period is 500 milliseconds, and the preferred data source is marked as a red high-priority channel.

[0037] When the communication delay of the roadside unit exceeds 300 milliseconds, the communication degradation processing unit performs spatio-temporal interpolation compensation: calls the historical road network health status atlas of the same period (error ± 15 minutes), same week type from the historical database, and uses a cubic spline interpolation algorithm to compensate for missing data. At the same time, send an instruction to the vehicle-mounted control unit to activate the locally cached vehicle flow pulse wave model parameters (including wave speed value 8 m / s, pressure decay coefficient 0.3), and independently calculate the section status using the vehicle's own sensor data. The diagnosis result in degradation mode is added with a yellow semi-transparent marker for differentiation.

[0038] In traditional mixed traffic flow monitoring systems, there are obvious deficiencies in the handling mechanism of data conflicts and communication failures. When the road-side detector and the vehicle-mounted sensor have different data on the headway of the same road section, the existing technology usually adopts a fixed rule arbitration: for example, a certain city's traffic cloud platform sets the road-side laser radar data as the first priority, regardless of the value reported by the vehicle-mounted unit, and the road-side data is used as the standard. This mechanism does not consider the real-time state of the equipment, and when the road-side radar lens is fogged due to heavy rain, it still forces the use of incorrect data (the speed detection deviation is up to ±15 km / h when the lens is fogged), causing the congestion index false positive rate to rise to 28%. A certain highway accident analysis report shows that the system did not adopt the vehicle-mounted emergency braking signal, delaying the accident warning by more than 11 seconds.

[0039] The degradation processing strategy when the communication link is interrupted is also rigid. The traditional scheme only calls historical synchronous mean data compensation, such as using the static road network state atlas of the same time one week ago during the morning peak. However, it does not consider dynamic factors such as weather changes and temporary traffic controls, resulting in a significant deviation between the compensation results and the actual situation. In a certain tunnel fire emergency test, the system continued to output historical mean state (lane occupancy rate 40%) after the communication interruption, while the actual occupancy rate had risen to 85% due to vehicle retention, causing the rescue vehicle dispatching path to be incorrect.

[0040] The system lacks the ability to verify multiple sources of data. The road-side unit and the vehicle-mounted unit operate independently, and there is no historical credibility evaluation model. The records of a certain intelligent traffic management center show that when the vehicle-mounted GPS frequently loses packets due to high-rise building obstruction, the system still fuses the data according to the fixed weight (road-side data accounts for 70%), and outputs the wrong slow-moving area label for 3 consecutive days during the evening peak. Such defects are particularly prominent in complex urban road networks: the three-dimensional traffic flow of viaducts and ground roads is not distinguished in the modeling, and it takes 15 minutes to reinitialize the system after communication recovery, during which the invalid proportion of dispatching instructions exceeds 25%.

[0041] In another technical solution, when the energy coordination module calculates the charging demand elasticity index, the following steps are performed: Real-time acquisition of dynamic parameters of the power grid: obtain the time-of-use electricity price data provided by the charging station management platform, the idle rate data of the charging piles, and the battery remaining capacity distribution data of the autonomous taxis in the target area; Construct a CDEI calculation model: dynamically generate the regional charging demand elasticity index based on the following formula: ; Where ΔP t is the electricity price fluctuation value at time t relative to the base electricity price P base ; Q idle,t is the number of idle charging piles at time t; Q total is the total number of charging piles; S critical,tthe number of vehicles with battery power below the warning threshold; N vehicle the total number of autonomous taxis in the target area that need to be dispatched, obtained in real time from the dispatch server; α, β, and γ are the weight coefficients of the power grid load, the scarcity of pile resources, and the urgency of vehicles, respectively; Dynamic adjustment of charging strategy: When CDEI t > θ high , trigger the charging price discount strategy, θ high is the charging demand stimulation threshold; When CDEI t < θ low , start the charging demand suppression strategy, θ low is the charging demand suppression threshold; Linkage charging scheduling scheme: input the adjusted charging strategy into the multi-objective planning model to optimize the generation of a charging instruction set containing a charging pile allocation scheme and a vehicle dispatch path.

[0042] In the above technical solution, the charging demand elasticity index calculation module is deployed in the real-time data processing unit of the energy coordination subsystem, and dynamic parameters are collected every 30 seconds: time-of-use electricity price data (time granularity 15 minutes) is obtained from the charging station management platform, the number of idle charging piles is read through the Modbus TCP protocol, and the battery remaining capacity distribution reported by the vehicle-mounted control unit is received (sampling accuracy ±2%). Among them, the electricity price fluctuation value ΔP t is calculated with the reference electricity price P base =0.8 yuan / kWh as the reference, and the idle pile proportion threshold is set to 20%; the battery power warning threshold is uniformly set to 15%, and the vehicles below this value are counted into the number of emergency vehicles S critical,t . The calculation model uses a weighted sum formula, with the power grid load weight α configured as 0.4, the pile resource scarcity weight β as 0.3, and the vehicle urgency weight γ as 0.3, and the weight coefficients and constant as 1.

[0043] The calculation formula of the charging demand elasticity index (CDEI t ), referring to the price elasticity formula in microeconomics: the elasticity coefficient E=(ΔQ / Q) / (ΔP / P); then normalize the power grid load elasticity (ΔP / P), pile resource elasticity (Q idle / Q total ), and vehicle urgency elasticity (S critical / N vehicle ) to the [0,1] interval, and perform weighted sum through the weight coefficients α, β, and γ. Each term on the right side is a dimensionless ratio, and CDEI is a dimensionless index.

[0044] When the CDEI t value exceeds the charging demand stimulation threshold θhigh =1.2, trigger price discount strategy: charging unit price is reduced by 0.2 yuan / kWh and pushed to the vehicle-mounted display screen; when CDEI t is lower than the suppression threshold θ low =0.8, start demand suppression strategy: charging unit price is increased by 0.3 yuan / kWh and only vehicles with less than 10% of the electric quantity are allowed to use fast charging piles. The response time of strategy switching is controlled within 5 seconds, and the adjusted pricing strategy is written into the rate configuration table of the charging station management platform in real time.

[0045] In the traditional charging management system, static strategies are difficult to cope with dynamic demand changes under mixed traffic flow. Existing technologies usually use fixed time period pricing mechanisms, for example, a certain city's charging platform divides the whole day into three time periods (peak period 1.2 yuan / kWh, flat period 0.8 yuan / kWh, valley period 0.4 yuan / kWh), but this mechanism does not associate real-time power grid load with vehicle charging demand. When a stadium holds a match and causes a surge in regional charging demand, the system still executes the preset valley period electricity price, resulting in a charging pile occupancy rate of up to 95% and an increase in power transformer overload trip accident rate to 1.7 times per year. A report on the operation of a certain charging station shows that 31% of non-emergency vehicles occupy fast charging piles for a long time during the valley period, while emergency vehicles with less than 10% of the electric quantity are forced to wait for more than 40 minutes.

[0046] Charging resource allocation strategies lack multi-dimensional coordination. Traditional systems only dispatch vehicles based on the physical idle state of charging piles, without considering road network travel time cost. Taking the dispatch of taxis at a certain airport as an example, the system guides 20 vehicles with low battery levels to a charging station 3 kilometers away, but fails to predict the accident congestion on the main road, and only 8 vehicles actually arrive before the battery is depleted. At the same time, the state of the vehicle battery is not included in the pricing model: a large number of high-electricity vehicles participate in charging during the peak period of power grid load, resulting in a measured charging pile utilization rate polarization - the idle rate in the core business district reaches 45%, while the queue length around the hospital exceeds 500 meters.

[0047] Strategy adjustment is significantly lagging. Existing systems rely on manual parameter modification, with an average time of 22 minutes from detecting demand anomalies to updating strategies. In the case of a certain traffic hub during the evening peak, the administrator only manually increased the electricity price by 0.3 yuan / kWh after finding that there were more than 50 charging queue vehicles, at which time 12 vehicles had already exited the service due to battery depletion. More seriously, the static threshold mechanism cannot adapt to weather changes: the charging priority adjustment is not automatically triggered during a rainstorm warning, resulting in the charging request of an ambulance vehicle being covered by ordinary orders, delaying the emergency task by 17 minutes. These defects form a chain reaction in large urban traffic networks, with the service interruption time caused by a single dispatch error averaging 2.3 hours.

[0048] In another technical solution, the energy coordination module further comprises an elasticity coefficient dynamic correction unit, which is configured to: monitor the actual implementation effect of the charging strategy formulated based on the charging demand elasticity index in real time, calculate the deviation rate of the actual charging capacity after the strategy implementation and the predicted charging capacity; when the deviation rate exceeds a preset threshold, automatically dynamically adjust the grid load weight coefficient, the pile resource scarcity weight coefficient and the vehicle emergency degree weight coefficient according to the grid load mutation indication signal and the actual deviation condition; during the dynamic adjustment of the weight coefficient, link the dynamic scheduling module, and temporarily freeze the charging pile allocation scheme generated based on the current CDEI t value until the correction of the new weight coefficient is completed.

[0049] In the above technical solution, the elasticity coefficient dynamic correction unit is integrated in the monitoring subsystem of the energy coordination module, and charging strategy implementation data is collected every 5 minutes. When monitoring the charging capacity deviation rate in real time, the deviation rate threshold is set to 15%, for example, when the predicted charging capacity is 100 kWh and the actual charging capacity is only 85 kWh, the correction process is triggered. The correction process first detects the grid load mutation signal: if the grid frequency fluctuation exceeds ±0.2 Hz for 10 seconds, it is determined to be a load mutation event; then the weight coefficient is adjusted based on the actual deviation direction: when the actual charging capacity is lower than the predicted value, the vehicle emergency degree weight γ is increased by 0.05, and the grid load weight α is decreased by 0.03. The weight adjustment step is limited to ±0.01 / time to prevent parameter oscillation.

[0050] During the correction period, the dynamic scheduling module is linked, and the freeze instruction code 0xFE is sent to the charging pile allocation system. In the freeze state, the current charging pile allocation scheme remains unchanged, but emergency vehicles with less than 10% of the power are allowed to be plugged in urgently. After the weight coefficient update is completed, the system automatically unfreezes and records the new parameters to the non-volatile memory. This unit runs on an industrial-grade PLC controller with an environmental temperature adaptation range of -10°C to 55°C.

[0051] In the traditional charging strategy optimization system, the static weight mechanism is difficult to adapt to dynamic environmental changes. The existing technology usually uses a fixed coefficient model to calculate the charging demand, for example, a certain city dispatching platform sets the grid load weight to be constant 0.5, the vehicle emergency degree weight to be 0.3, and the pile resource weight to be 0.2. When a rainstorm causes a sudden change in charging demand, the model still allocates resources according to the preset proportion, resulting in a deviation of up to 35% between the predicted charging capacity and the actual value. A typhoon passage day record shows that due to the lack of association with grid frequency fluctuation data (actual fluctuation ±0.5 Hz), the system continuously outputs incorrect strategies, causing 42% of the fast charging piles in the charging station to be idle while vehicles are queuing for more than 2 kilometers. The manual intervention mode has serious lag. The operation and maintenance personnel need to manually adjust the parameters after discovering the abnormality through the monitoring screen. The average time from the deviation to the completion of correction is 22 minutes. In a certain traffic hub case, the charging capacity is 18% lower than the predicted value for three consecutive periods, triggering manual verification. During this period, the error allocation strategy caused the charging of 8 ambulances to be delayed. More importantly, the system continues to execute the original scheme during the adjustment period: when the administrator reduces the grid weight from 0.5 to 0.4, the charging pile allocation instructions that are being executed are not frozen, causing 12 non-emergency vehicles to continue to occupy the core area charging piles, delaying the dispatch of emergency vehicles for 17 minutes. The system lacks the ability to correct multiple parameters simultaneously. The traditional model only sets the threshold based on historical statistical values (such as a fixed deviation rate threshold of 20%), without considering the influence of real-time environmental variables. Certain winter test data shows that when the temperature drops to -10°C, the battery charging efficiency decreases by 40%, but the system still predicts the demand according to the normal temperature model, and the actual charging capacity deviation expands to 41%. Such a static mechanism causes a chain reaction in complex scenarios: a single weight adjustment error causes the charging resource mismatch rate to rise to 53% in the next three scheduling periods, and it takes more than 6 hours to recover to stable operation.

[0052] In another technical solution, the multi-objective planning model of the energy coordination module generates a charging scheduling scheme according to the following steps: A battery cycle attenuation cost function is constructed: ; wherein C bat (i,j) is the attenuation cost of vehicle i using charging pile j; r j is the charging rate of charging pile j; ΔT j =|T j env -T i opt | is the difference between the environmental temperature and the optimal battery temperature; a i , b i are the battery attenuation coefficients of vehicle i, respectively; An order revenue function is constructed: ; wherein R order (i,k) is the revenue of vehicle i executing order k; F k is the basic transportation fee of the order; s k is the order mileage; θ is the distance sensitivity parameter, with a value of 1.2 for highways and 0.8 for urban areas; P k is the time premium benchmark value; e k is the urgency level, with a value of 3 for medical treatment, 2 for flights, and 1 for ordinary; δ is the urgency amplification factor, with a value of 1.5; w k is the passenger waiting time; ρ is the waiting penalty coefficient, with a value of 0.1; Generation matching scheme: ; Dynamic weight factor: ; Wherein, X ijk is a decision variable, taking value 1 when matching, and value 0 when not matching; Λ0 is a basic weight, taking value 0.3; Y i is the service life of vehicle i; v is the age gain coefficient, taking value 0.15; τ is the age segmentation threshold, taking value 2 years; Set hard constraints: Single pile constraint, each vehicle uses at most one charging pile; Electricity guarantee, the electricity after charging ≥ the minimum demand of the order + redundancy; Time limit constraint, charging time + driving time ≤ order deadline.

[0053] In the above technical scheme, regarding the battery cycle attenuation cost function, the charging rate influence (r j 2 ) is introduced: when the battery is fast charged, the lithium ion migration rate increases, resulting in increased electrode stress and thickened SEI film, and the attenuation shows a quadratic relationship (referring to the electrochemical kinetics model). The charging rate is limited to 0.5C-2C, and the coefficient a i is calibrated through battery cycle test. Temperature deviation influence (ΔT j term): ΔT j =|T j env -T i opt reflects the difference between the ambient temperature and the optimal battery temperature (25℃). Temperature deviation will exacerbate side reactions: low-temperature lithium precipitation risk, increased internal resistance; high-temperature electrolyte decomposition accelerated. The linear term bi×ΔT j simplifies the Arrhenius equation (activation energy model), as actual engineering requires low computational complexity.

[0054] Regarding the order revenue function, the distance-sensitive revenue (F k ×s k θ ) is introduced: the basic transportation fee F k is multiplied by the θth power of the mileage s k , which reflects the non-linear revenue: θ>1 (1.2 for high-speed), the marginal revenue of long-distance orders increases (covering the empty running cost); θ<1 (0.8 for urban areas), the revenue of short-distance orders decreases (efficiency loss during congestion period). Introduce emergency degree premium (P k ×δ ek ): the exponential term δ ek (δ=1.5, ek∈{1,2,3}) amplifies the revenue of medical / flight orders. For example, the premium for a medical order (ek=3) is P kx1.5 3 = 3.375P k , which meets the social priority scheduling needs (the background technology indicates that the traditional system does not distinguish the urgency). The waiting penalty (p x w k ): the linear penalty term p = 0.1 inhibits the long waiting time of passengers (the empirical research shows that the order cancellation rate increases by 60% when the waiting time is longer than 10 minutes).

[0055] The objective function of the matching scheme is the core of the multi-objective programming model, which aims to convert the two conflicting objectives (maximizing order revenue vs. minimizing battery degradation cost) into a single objective maximization problem through the weighting and method. The derivation is based on the following logic: Multi-objective optimization problem definition: there is a conflict between maximizing order revenue and minimizing battery cycle degradation cost: fast charging (high r j ) and high-revenue orders (such as emergency medical treatment) may accelerate battery degradation.

[0056] To balance the objectives, a dynamic weight factor Λ i is introduced to convert the battery degradation cost into an equivalent "negative revenue". The objective function becomes: ; Since the decision variable X ijk is the same, the two summation terms can be combined as: ; This means that for each possible allocation (i, j, k), the net revenue (order revenue minus weighted degradation cost) is calculated, and the total is maximized.

[0057] Where the physical meaning and parameter selection basis are: Net revenue maximization: the formula encourages the system to prioritize high-revenue orders (such as emergency medical treatment) and low-degradation-cost charging piles (such as slow charging, suitable temperature).

[0058] The role of weight Λ i : Λ i is dynamically adjusted to reflect the degree of vehicle battery aging (old car weight is higher), ensuring the protection of old car batteries (frequent fast charging increases the battery degradation rate).

[0059] Constraint integration: the objective function needs to meet the hard constraints: Single pile constraint: ∑X ijk ≤ 1 (each vehicle can use at most one pile).

[0060] Power guarantee: the charging power is greater than or equal to the minimum demand of the order + 15% redundancy.

[0061] Time constraint: charging time + travel time ≤ order deadline.

[0062] These constraints are handled by integer programming (e.g. branch and bound) when solving, ensuring that the solution is feasible.

[0063] Regarding the dynamic weight factor formula, which is used to dynamically adjust the importance of battery degradation cost in the objective function, it is based on the age of the vehicle. Older vehicles have more degraded batteries, so the weight is higher. The derivation is based on engineering experience and measured data: Design goal: In the matching scheme objective function, Λ i is the weight of the battery degradation cost C bat (i,j).

[0064] Core problem: New car battery health, degradation cost weight should be low; old car battery aging, weight should be higher to prioritize protection.

[0065] The weight needs to be quantified as the vehicle age Y i increases.

[0066] Linear step growth model: Base weight (Λ0): Set as a constant (0.3), representing the minimum weight of a new car.

[0067] Age impact term: Introduce a gain term , where: Calculate the integer multiple of the threshold value τ of the vehicle age (for example, Y i =2.5 years, τ=2, then =1). ν controls the gain amplitude (value 0.15), which is calibrated by a particle swarm optimization algorithm.

[0068] Formula combination: Weight Λ i increases with the age of the vehicle in steps: ; Parameter selection basis: Λ0=0.3: The base value ensures that the optimization of new car benefits is dominant.

[0069] ν=0.15: Calibrated through historical data, balancing battery life and benefits.

[0070] τ=2 years: Based on battery aging data segmentation (measured data shows that degradation accelerates after 2 years of operation).

[0071] Integer function : Avoid frequent fluctuations in weight, enhance system stability.

[0072] In the matching scheme, higher Λ i amplifies the battery degradation cost of old cars, guiding the system to allocate slow-charging piles or low-rate charging to extend battery life.

[0073] The multi-objective programming model is deployed in the mathematical optimization engine of the dispatch server, and the battery degradation cost function includes a charging rate term and an ambient temperature term. The charging rate r j is set between 0.5C and 2C, where 1C rate corresponds to a charging time of 1 hour; the optimal battery temperature T i opt is fixed at 25°C, and the ambient temperature difference ΔT j is calculated using an absolute value function. The vehicle battery degradation coefficient a i is 0.003, and b i is 0.001, which is obtained by calibration through a battery cycle test database. The highway distance sensitivity θ in the order revenue function is set to 1.2, and the urban road is set to 0.8; the time premium benchmark value P k is 5 yuan within 5 kilometers, and 8 yuan for 5-10 kilometers; the urgency amplification factor δ is fixed at 1.5.

[0074] The model solution uses a branch and bound algorithm, and the base value Λ0 is set to 0.3 when calculating the dynamic weight factor Λ, and the vehicle age gain coefficient ν is 0.15. For vehicles with a service life Y i that exceeds the segmented threshold τ = 2 years, the weight increase is calculated by (Y i -2) × ν. The hard constraint execution strategy is: the single pile constraint is realized by 0-1 integer programming, and each vehicle is associated with only one charging pile ID; the power guarantee constraint requires that the SOC after charging is greater than the order demand value + 15% redundancy; the time constraint uses a sliding time window to check that the interval between the charging start time and the order deadline must be greater than the sum of the charging time and the path travel time.

[0075] In the traditional automatic driving taxi charging dispatching system, the single objective optimization mechanism has significant limitations. Existing technologies usually only allocate charging resources based on maximizing order revenue, without quantifying the cost of battery health loss. Taking a city dispatch platform as an example, the system prioritizes fast charging piles to high-revenue order vehicles (such as airport long-distance orders), but does not consider the impact of fast charging rate (commonly above 2C) on battery cycle life. The measured data shows that the battery capacity attenuation rate of vehicles frequently performing such dispatching is 28% after 18 months of operation, which is 9 percentage points higher than that of balanced dispatching vehicles. The maintenance records of a vehicle fleet show that, due to the lack of temperature compensation mechanism, 2C fast charging is still allocated in winter at -5°C according to the normal temperature model, resulting in an increase in battery lithium precipitation failure rate to an average of 3.2 times per year.

[0076] Static parameter models cannot adapt to differences in vehicle state. The traditional system uses a uniform degradation coefficient (such as fixed a i= 0.002), ignoring the individual battery aging degree differentiation. A certain four-year-old vehicle caused the battery resistance to increase by 15% after three consecutive fast charges due to the use of new battery parameters, triggering a charging interruption fault. At the same time, the order income calculation uses a fixed formula: the premium ratio of emergency medical orders to ordinary orders is fixed at 1.3:1, without dynamically correlating with the actual road conditions. When the hospital is surrounded by congestion, the system still guides the vehicle to go according to the preset premium, and 23% of the emergency vehicles are delayed for more than 15 minutes due to road congestion.

[0077] Inflexible execution leads to scheduling failure. The power guarantee constraint only requires the SOC after charging to be greater than or equal to the minimum demand of the order, without setting a redundant buffer (such as a system setting SOC ≥ 20% is qualified). In a scenario where the distance from the charging station to the pickup point is 10 kilometers, 5 vehicles are forced to stop due to not taking into account the additional power consumption caused by congestion (actual consumption 18%). The time constraint check has loopholes: the charging time is calculated based on the nominal value, without real-time correlation with the actual output power fluctuation of the pile. A certain charging pile group caused the actual charging power to be only 70% of the nominal value due to the voltage drop of the power grid, causing 12 vehicles to be charged for more than time, with a delay of flight orders accounting for 17%. These defects form a cumulative effect in large-scale vehicle operation, with an annual average of 370,000 yuan per thousand vehicles in battery replacement costs caused by unbalanced scheduling.

[0078] In another technical solution, the energy coordination module further executes a battery temperature real-time compensation mechanism when generating a charging scheduling scheme using the multi-objective planning model, wherein: Real-time receiving of the actual working temperature of the vehicle battery reported by the vehicle-mounted control unit; Introducing the difference between the actual working temperature of the battery and the ambient temperature of the charging pile as a compensation factor to dynamically correct the decay cost calculation result; When it is detected that the difference between the actual working temperature of the battery and the ambient temperature is greater than or equal to 15°C, the maximum charging rate of the charging pile available to the vehicle is forcibly limited.

[0079] In the above technical solution, the battery temperature real-time compensation mechanism is integrated into the safety control unit of the energy coordination module, and the actual working temperature of the battery is collected every 2 seconds through the vehicle-mounted CAN bus (accuracy ±0.5°C). The compensation factor calculation uses a double-temperature difference superposition strategy: first, calculate the difference ΔT j env between the charging pile ambient temperature T i and the optimal battery temperature T opt , and then obtain the difference ΔT j between the actual working temperature T i of the battery and T real . When ΔT j env comp compWhen the temperature is ≥5℃, add a compensation term C to the attenuation cost function. comp =0.002×|ΔT comp |;When ΔT comp At temperatures ≥15℃, a charging rate limiting protocol is triggered: the maximum usable charging rate for the vehicle is reduced from 2C to 1C. The temperature data source is an NTC thermistor built into the vehicle's battery pack, installed at the cell gaps.

[0080] During operation, the compensated attenuation cost function is: C bat (i,j)=a i ·r j 2 + b i ·|ΔT j | + C comp This calculation is performed in the real-time processing thread of the scheduling server, with the thread priority set to the highest level. Ambient temperature data is acquired through IoT sensors on the charging station management platform, installed 1.5 meters above the ground under the charging pile's sunshade, with a temperature sampling interval of 10 seconds.

[0081] The test was conducted in an extreme temperature scenario: on a hot summer day (ambient temperature 42℃), 20 vehicles with battery operating temperatures reaching 58℃ were selected. After the compensation mechanism was activated, the charging rate of 12 vehicles with temperature differences exceeding 15℃ was limited to 1C, and the fault diagnosis system recorded that the number of battery over-temperature alarms decreased to 0.2 times per 10,000 vehicles per day. The firmware for this unit is burned into the read-only memory partition of the onboard MCU, occupying 48KB of memory.

[0082] Traditional electric vehicle charging scheduling systems suffer from significant deficiencies in the quantitative management of battery temperature impact. Current technologies typically assess battery degradation costs solely based on the ambient temperature of the charging station, failing to collect real-time data on the battery's actual operating temperature. For example, a city's charging management platform uses a fixed ambient temperature sensor reading (installed on the sun visor of the charging station) for its degradation cost model, neglecting to account for the possibility that the internal battery cell temperature might exceed the ambient temperature by 15°C. Real-world testing shows that during peak summer temperatures, when the core battery temperature of fast-charging vehicles reaches 52°C, the system still calculates degradation costs based on an ambient temperature of 37°C, resulting in an actual battery capacity degradation rate 23% faster than predicted. A fleet operation report indicates that this deviation shortens the battery replacement cycle from the designed lifespan of 5 years to 3.8 years.

[0083] The temperature protection mechanism suffers from a delayed response. Traditional systems only intervene in the charging process after the battery management system reports a temperature alarm, by which time the cell temperature has typically exceeded a safety threshold (e.g., >60℃). Records from a certain brand's fast charging station show that when the ambient temperature was 40℃, the system did not impose any restrictions on 2C fast charging vehicles until the battery temperature rose to 58℃ before triggering a power reduction command. This process took an average of 8 minutes, during which the battery's SEI film decomposed rapidly. More seriously, there is a lack of temperature monitoring points: most vehicles only have a single-point sensor on the battery pack casing, which cannot reflect the internal temperature gradient. An accident analysis showed that the temperature of a local hot spot in the cell was 19℃ higher than that of the casing sensor, but the system failed to detect this difference, resulting in a 34% false alarm rate for thermal runaway warnings.

[0084] The system lacks temperature difference compensation capabilities. The degradation cost calculation model does not incorporate a compensation term for the difference between ambient temperature and the actual battery temperature, resulting in significant errors under extreme temperature scenarios. In a winter test case, the charging station's ambient temperature was -5°C, but the vehicle's battery maintained an operating temperature of 15°C due to the preheating system. The traditional model still calculated the high degradation cost based on a 20°C temperature difference, incorrectly guiding the vehicle to a slow charging station 10 kilometers away (ambient temperature 5°C), actually causing the vehicle to run out of power and become immobilized. Such problems occur frequently in cold regions: on average, scheduling errors caused by inaccurate temperature models account for 41% of all charging failures annually, adding an extra 170,000 yuan per thousand vehicles in rescue costs.

[0085] In another technical solution, the dynamic scheduling module is also connected to a road network topology change sensing unit, which is configured as follows: It can connect to the municipal construction notice interface in real time and receive abnormal event information reported by vehicle cameras or roadside units; When an event that changes the road network topology is detected, such as a road closure, a change in the number of lanes, or the addition of a new controlled area, a corresponding topological mutation event code is generated. The traffic flow pulse wave propagation model in the traffic flow analysis module is triggered to perform dynamic reconstruction, and the waveguide unit segmentation parameters in the model are adjusted according to the changes in the actual road capacity. Send a path feasibility status flag to the energy coordination module and the dynamic scheduling module to prohibit dispatched vehicles from entering areas that are impassable due to topological changes.

[0086] In the above technical solution, the road network topology mutation sensing unit is deployed on the edge computing node of the scheduling server. It polls the municipal construction notice interface (data format: GeoJSON) every 5 minutes via HTTPS protocol, and simultaneously subscribes to the event message queues of vehicle-mounted cameras and roadside units. When a road closure event is detected (e.g., the municipal interface returns the "road_closed" label with a confidence level > 90%), the unit generates a four-digit topology mutation event code: the first digit 1 indicates road closure, 2 indicates lane reduction, and 3 indicates a newly added control zone. The event code is associated with geofence coordinates, with an error radius ≤ 15 meters. For example, when the event "Chang'an Avenue eastbound to westbound closed" is detected, event code 1002 is output, and the latitude and longitude of the geofence vertex are marked.

[0087] After triggering the reconstruction process, the unit sends a model reset command to the traffic flow analysis module. The parameters of the traffic flow pulse wave propagation model are adjusted according to the event type: for road closure events, the capacity of the corresponding road segment waveguide unit is reduced to 0; for lane reduction events, the wave velocity value is adjusted according to the reduction ratio (e.g., when three lanes are reduced to two lanes, the wave velocity decreases from 12 m / s to 8 m / s). The parameter update cycle is controlled within 10 seconds, and the reconstructed model output overwrites the original results. The path feasibility status flag is pushed out using binary encoding, with 0 indicating impassable and 1 indicating normal.

[0088] Real-world verification was conducted on a subway construction section: After the municipal notice was issued, the unit marked the closed area within 8 seconds. The dispatch system guided 32 trains to detour, reducing route planning failures by 82% compared to a traditional system that did not utilize this unit. It is installed in the firewall's internal network area on the east side of the dispatch room and connected to the municipal private network via a fiber optic transceiver.

[0089] In traditional road network status update systems, the response mechanism for topological change events suffers from significant lag. Existing technologies typically rely on manual review of municipal construction notices. For example, a city's traffic management center requires maintenance units to submit construction information via email before 9:00 AM daily, with administrators manually entering the data into the system, averaging 47 minutes. In a case of temporary traffic control for a sporting event, due to delayed information uploads, the system continued to guide vehicles into the controlled area even one hour after road closure, causing a collective delay of 22 minutes for 32 vehicles. Road capacity adjustments use a fixed template: lane reduction events are uniformly treated as a 30% capacity decrease, without dynamic modeling based on the actual reduction percentage. In one elevated bridge repair project, two lanes were actually closed (originally three lanes), but the system incorrectly applied a 40% capacity reduction rate, leading to an imbalance in traffic distribution between adjacent road sections and a surge of 28 points in the congestion index.

[0090] The system lacks multi-source event fusion capabilities. Abnormal events detected by vehicle-mounted cameras or roadside units (such as road collapses) require transcoding and processing by the central server, with an average delay of 8 minutes. In a tunnel fire incident, the roadside flame sensor issued an alarm within 3 seconds, but the system failed to process it in time due to protocol conversion queuing, only marking the road segment as abnormal 6 minutes later. More seriously, there is a lack of linkage between events and the road network model: after detecting road closures, only red blocks are marked on the map, without triggering traffic flow model reconstruction. During the construction of an airport expressway, the traditional system continuously calculated the vehicle flow propagation wave speed based on full capacity, causing the congestion prediction deviation to increase to ±15 kilometers, and the actual travel distance of detour vehicles increased by 37% compared to the system's planned value.

[0091] Route feasibility checks are ineffective. Traditional solutions only include a "Construction Detour" prompt button on the navigation interface, without forcibly prohibiting vehicles from entering closed areas. A smart city platform recorded 142 vehicles ignoring the prompts and entering a subway construction zone within 7 days, triggering a false entry alarm rate of 11%. Furthermore, the boundaries of controlled areas are vague: geofencing uses a 500m x 500m grid. During a temporary COVID-19 lockdown in a residential area, the system marked the entire 1.2 square kilometer area as impassable, mistakenly blocking 12 normal side roads. These deficiencies are particularly pronounced in emergency scenarios: after a bridge collapse, the system required a restart and initialization (taking 23 minutes), during which the dispatch command error rate reached as high as 64%.

[0092] In another technical solution, the dynamic scheduling module is connected to the weather warning interface. When the thunderstorm warning level is ≥ yellow, the following is executed: Group autonomous taxis within the target area into a fleet; Based on the traffic flow pulse wave propagation model, a dedicated waveguide channel for the fleet is generated under the condition that the road network topology allows. Acceleration / braking commands for platooned vehicles are synchronized via onboard unit.

[0093] In the above technical solution, the meteorological early warning interface is integrated into the security monitoring subsystem of the dispatch server, receiving thunderstorm early warning data issued by the National Meteorological Administration in real time. When the warning level reaches yellow (wind speed ≥ 10.8 m / s and lightning frequency ≥ 20 times / minute), the system initiates the vehicle grouping process: automatically grouping adjacent autonomous taxis within the target area into fleets of 6-8 vehicles each, with a maximum group spacing of 15 meters. The grouping command is sent to the vehicle control unit via C-V2X direct communication, with a response delay of no more than 200 milliseconds. The vehicle-mounted millimeter-wave radar (installed in the middle of the front bumper) and camera (top of the windshield) collect the positions of surrounding vehicles in real time, with an accuracy of ±0.3 meters.

[0094] The process of generating a dedicated waveguide passage for the convoy: The traffic flow analysis module calls the vehicle flow pulse wave propagation model to mark the convoy passage path (3.5 meters wide) in the road network topology, and calculates the optimal wave speed (range 8-12 m / s) based on real-time traffic pressure values. After the passage is generated, the dynamic scheduling module sends synchronous control commands to the convoy vehicles via gigabit fiber optic cable: acceleration / braking commands use timestamp alignment technology, with deviation controlled within ±0.2 seconds. The command execution cycle is 500 milliseconds, and the commands are transmitted to the electronic control unit via the vehicle's CAN-FD bus.

[0095] Traditional weather response systems for autonomous vehicles have significant limitations in their collaborative control capabilities under severe weather conditions. Current technologies typically only send deceleration commands (e.g., a speed limit of 30 km / h) to individual vehicles after a thunderstorm warning is issued, without establishing a collaborative mechanism between vehicles. Records from a city's intelligent transportation platform show that during a yellow warning, vehicles within the area continued to operate independently, with the average distance between vehicles increasing from the standard 15 meters to 42 meters, resulting in a 25% decrease in traffic efficiency. An analysis of an accident on an airport expressway indicates that when a vehicle in front brakes suddenly during heavy rain, the failure of following vehicles to decelerate synchronously causes a 0.8-second response delay, increasing the rear-end collision rate to 3.2 times that of ordinary weather.

[0096] The dedicated lane generation function is missing. Traditional systems rely on general navigation route planning, guiding vehicles according to the clear-day right-of-way principle even when visibility is below 500 meters. In a typhoon case, the system guided 12 vehicles across a cross-sea bridge, failing to anticipate the risk of sudden crosswind changes (actual wind speed 14 m / s), causing 3 vehicles to deviate from their lanes and triggering a guardrail collision warning. Simultaneously, the road network status model does not incorporate meteorological data: during thunderstorms, the traffic flow propagation model uses clear-day parameters (fixed wave speed 10 m / s), rendering congestion predictions completely ineffective. Real-world data shows that the traditional model's travel time prediction error during heavy rain reaches ±12 minutes.

[0097] The asynchronous nature of control commands is a significant issue. Acceleration / braking commands received by a single vehicle are relayed through the cloud with an average delay of 1.2 seconds, leading to asynchronous actions among vehicles in the platoon. In a case observed during peak hours at a stadium, after the lead vehicle issued a braking command, the last vehicle's response was delayed by 1.5 seconds due to network fluctuations, causing the platoon's safe distance to plummet from the set value of 10 meters to 0.3 meters. More seriously, the system lacks a spatiotemporal alignment mechanism: navigation commands are generated based on the vehicle's local clock, without calibrating the time server, resulting in clock drift of up to 0.8 seconds among vehicles in the platoon. These deficiencies have a cumulative effect in severe convective weather, with insufficient weather response accounting for 31% of all autonomous driving accidents annually.

[0098] The present invention also provides a collaborative decision-making method for the dynamic dispatching system of autonomous taxis oriented towards mixed traffic flow, wherein the dispatching server performs the following steps: Real-time acquisition of mixed traffic flow data collected by roadside sensing units, including the speeds of manually driven vehicles and autonomous taxis; By integrating historical travel data from ride-hailing platforms, social media event tag data, and meteorological monitoring data, and inputting them into a spatiotemporal graph convolutional neural network model, the system outputs a 15-minute granular prediction of travel demand distribution. Based on the driving speed in the mixed traffic flow data, a road capacity heatmap is constructed using the ant colony pheromone diffusion algorithm. The road capacity heatmap is input into the traffic flow pulse wave propagation model to diagnose the road network health status, which includes the direction of congestion propagation and bottleneck section markings. Based on the predicted travel demand distribution and road network health status, a dynamic dispatch instruction set for autonomous taxis is generated, which includes route replanning instructions and empty vehicle dispatch coordinates. Receive charging pile usage status and grid time-of-use electricity price data sent by the charging station management platform, and calculate the charging demand elasticity index in combination with the remaining battery power of the vehicle; The charging service pricing strategy is dynamically adjusted based on the charging demand elasticity index, and the battery degradation cost function and order revenue function are input into a multi-objective programming model to generate a charging scheduling scheme. The dynamic scheduling instruction set and charging scheduling scheme are sent to the vehicle control unit, triggering the following operations: Control the autonomous taxi's route changes or charging station navigation; The system uses augmented reality to dynamically display yield intention markers, lane change trajectory previews, and summaries of dispatch decision-making criteria on the windshield. The vehicle decision timestamp and scheduling parameters are written into the blockchain distributed ledger, generating an encrypted verification record that is updated every 0.1 seconds.

[0099] In the above technical solution, the dispatch server receives mixed traffic flow data uploaded by the roadside sensing unit in real time, including the instantaneous speeds of manually driven vehicles and autonomous taxis (sampling frequency 10Hz). It simultaneously integrates historical travel data from ride-hailing platforms (15-minute time slices), Weibo social media event tags (keyword matching threshold 80%), and real-time precipitation probability data from the meteorological bureau, inputting these into a spatiotemporal graph convolutional neural network model. This network contains 8 convolutional layers and outputs a heatmap of travel demand distribution for the next 15 minutes, with a grid accuracy of 500m × 500m.

[0100] Based on vehicle speed data in mixed traffic flow, an ant colony pheromone diffusion algorithm is used to generate a road capacity heatmap: the initial pheromone concentration is 0.3, the volatility coefficient is 0.25, and after 50 iterations, a heatmap with a resolution of 20 meters per pixel is output. This heatmap is then input into a traffic flow pulse wave propagation model, and the wave equation is solved. (Wave velocity c=10±2m / s), diagnose the health status of the road network and mark the direction of congestion propagation and the coordinates of bottleneck road sections.

[0101] The dynamic scheduling module invokes travel demand forecasts and road network health status, employing the A* algorithm to generate route replanning instructions (heuristic function weight 0.7) and empty vehicle scheduling coordinates (positioning accuracy ±10 meters). The energy coordination module combines charging pile idle rate (threshold 20%), grid time-of-use electricity price (base price 0.8 yuan / kWh), and vehicle battery charge level (warning threshold 15%) to calculate the charging demand elasticity index CDEI (weights α=0.4, β=0.3, γ=0.3). When CDEI > 1.2, a price reduction of 0.2 yuan / kWh is triggered; when CDEI < 0.8, a price increase of 0.3 yuan / kWh is triggered. The multi-objective programming model uses a dynamic weighting factor Λ = 0.3 + 0.15 × max(Y). i The solution process (-2,0) optimizes the battery degradation cost and order revenue, with a solution time of ≤45 seconds.

[0102] After the command is issued to the vehicle control unit, it performs three-way operations: controls the vehicle to perform route changes or charging navigation; projects a yield intention marker (green flashing area), a lane change trajectory preview (blue dashed arrow), and a dispatch decision summary (text scroll bar) through the augmented reality windshield; and writes the decision timestamp and key parameters (such as congestion index 7.2, CDEI value 1.15) into a blockchain storage unit based on Hyperledger Fabric, generating an immutable record with SHA-256 signature every 0.1 seconds.

[0103] This technical solution achieves end-to-end collaborative optimization of autonomous taxi dispatching in mixed traffic flow environments. Through a multi-module data closed-loop processing mechanism, it seamlessly integrates road network health diagnosis, travel demand prediction, vehicle dynamic dispatching, and charging energy management, significantly shortening dispatch response time to the 3-second level and reducing vehicle empty mileage. Augmented reality windshield dynamic projection of yield intention markers and lane change trajectory previews enhances passenger understanding of dispatch decisions. Combined with blockchain-based evidence storage units that solidify decision parameters every 0.1 seconds, dispute tracing efficiency is improved to the minute level. A charging demand elasticity index, linked to a multi-objective programming model, dynamically adjusts charging strategies when grid load fluctuates, improving charging pile turnover and ensuring revenue for high-priority orders. A traffic flow pulse wave propagation model accurately captures the direction of congestion propagation, coordinating with a weather early warning mechanism to ensure the system maintains stable traffic efficiency under extreme conditions such as thunderstorms, comprehensively improving the reliability and resource utilization efficiency of the mixed traffic flow road network.

[0104] In another technical solution, the scheduling server is connected to an offline training platform, which includes: Historical data warehouse: Stores historical traffic flow data of road network, charging pile operation data, vehicle battery degradation data and order revenue data, with a data time span of ≥6 months; Parameter Co-calibration Module: Performs the following training process: Phase 1: Iteratively train the vehicle flow pulse wave propagation model based on historical traffic flow data, and optimize the propagation attenuation rate parameter and waveguide unit piecewise coefficient in the wave equation through backpropagation algorithm; Phase 2: Input historical travel demand data (ride-hailing platform data + social media events + weather data) into the spatiotemporal graph convolutional neural network, and use time sliding window cross-validation to optimize the hyperparameters of the prediction model; Phase 3: Using battery cycle degradation experimental data and order revenue statistics, the dynamic weight factors (ν, τ) and the weight coefficients (α, β, γ) of the charging demand elasticity index (CDEI) in the multi-objective programming model are solved by particle swarm optimization algorithm, so that the weighted sum error rate of battery degradation cost and order revenue is ≤5%.

[0105] The parameter co-calibration module is further configured as follows: When a sudden change in the actual road network topology is detected, incremental training of the traffic flow pulse wave propagation model is automatically triggered, and only the waveguide unit parameters of the affected road segment are retrained. A full system parameter refresh is performed monthly, synchronizing the optimized parameters output by the offline training platform to the online scheduling server.

[0106] In the above technical solution, the offline training platform is connected to the dispatch server via a dedicated optical cable. Its historical data warehouse stores six types of core data: historical traffic flow data collected by roadside radar (timestamp accuracy of 1 second), charging pile operation logs (including charging rate and ambient temperature), vehicle battery cycle test data (charge and discharge cycles ≥ 2000), and order revenue details (distinguishing between regular and emergency orders). The data spans 8 months and is stored in a distributed HDFS cluster with a total capacity of 8PB.

[0107] The parameter co-calibration module performs three-stage training: The traffic flow pulse wave propagation model was optimized by inputting historical traffic flow data (covering sunny, rainy, and snowy weather). The wave equation parameters were optimized using a backpropagation algorithm: with the actual congestion propagation speed as the target value, the waveguide unit segmentation coefficients (initial value 200 meters) and attenuation rate parameters (initial value 0.3) were iteratively adjusted. Training was performed using the Adam optimizer with a learning rate of 0.001, and after 1000 iterations, a model parameter set with an error rate ≤3% was output.

[0108] The spatiotemporal graph convolutional neural network was hyperparameter-tuned by constructing a time sliding window (15-day window length, 1-day step size) and inputting historical ride-hailing platform order data, Weibo event tags (matching keywords "concert / sports event / exhibition"), and meteorological bureau precipitation data. Cross-validation was used to optimize the network hyperparameters: the convolutional kernel size was searched using a 3×3 to 7×7 grid, and the Sigmoid and ReLU activation functions were tested. The ReLU+Softmax combination was ultimately selected, and the prediction error stabilized within ±7%.

[0109] Multi-objective programming model collaborative calibration was performed using a particle swarm optimization algorithm (population size 50, iterations 200) to solve for dynamic weighting factors: the search space for the vehicle age gain coefficient ν is [0.1, 0.2], and the search for the age segmentation threshold τ is [1.5, 3.0] years; the constraint condition for the CDEI weighting coefficients α, β, γ is α + β + γ = 1. The optimization objective is to achieve a weighted sum error rate of ≤5% for battery degradation cost and order revenue. Experimental data comes from 6 months of operation records of 300 vehicles.

[0110] Dynamic response mechanism: When the road network topology mutation event code is triggered, only the historical data of the affected road segment is extracted to retrain the corresponding waveguide unit parameters, and the incremental training time is ≤15 minutes. The entire system parameters are refreshed at 00:00 on the 1st of each month: the offline platform outputs the optimized parameter package, which is encrypted with AES-256 and synchronized to the online scheduling server. The switching process is completed seamlessly within 5 seconds.

[0111] This technical solution achieves continuous optimization and dynamic adaptation of system parameters by constructing an offline training platform. The multi-dimensional, long-term samples provided by the historical data warehouse more than double the congestion prediction accuracy of the traffic flow pulse wave propagation model, significantly reducing the error rate, especially in rainy or snowy weather. The parameter collaborative calibration module employs a three-stage progressive training approach, accurately matching actual road condition characteristics through backpropagation and particle swarm optimization algorithms, keeping the weighted sum error of the multi-objective programming model within 5%, effectively balancing battery life and operational benefits. The incremental training mechanism responds quickly to sudden changes in the road network, requiring only local data retraining to update the model, avoiding the resource consumption of full model reconstruction in traditional solutions. Monthly full parameter updates ensure long-term system stability, with no service interruption during the synchronization process. Overall, the offline training platform enables the online scheduling system to continuously evolve, maintaining a high level of decision-making accuracy in complex mixed traffic flow environments.

[0112] In another technical solution, the scheduling server is deployed in a three-tier distributed architecture: Edge computing layer: Composed of edge nodes deployed within roadside sensing units, performing: Real-time preprocessing of traffic flow data for this road segment (compressing redundant information by ≥50%); The ant colony pheromone diffusion algorithm is run locally to generate a road segment-level traffic capacity heatmap. The parameters of the traffic flow pulse wave propagation model are cached, and the health status of the road segment is output independently when communication degrades.

[0113] Regional scheduling layer: Deploy computing nodes according to city zones, and execute: Aggregate edge layer heatmaps and fuse them to generate a zone-level road network health status map; Spatiotemporal graph convolutional neural network sub-model for running the demand forecasting module; To handle multi-objective charging scheduling calculations for vehicles in this region, a branch and bound algorithm is used to compress the solution space.

[0114] Central decision-making level: Global scheduling server execution: Integrate the road network health status and demand forecast results of each zone to generate cross-zone dynamic dispatch instructions; Optimize global energy coordination strategies (such as cross-regional allocation of charging pile resources); Manage the data consistency of blockchain evidence storage nodes.

[0115] The communication mechanism of the layered architecture is configured as follows: Edge layer → Region layer: Transmit lightweight diagnostic results (only key fields such as congestion index and bottleneck point coordinates); Regional layer → Central layer: Trigger transmission according to preset thresholds (e.g., partition health < 60% or order gap rate > 20%). A computing task migration protocol is adopted: when the load of a node in a certain area is greater than 80%, some vehicle scheduling tasks are automatically migrated to adjacent idle nodes.

[0116] In the above technical solution, the edge computing layer is deployed inside the roadside sensing unit, using the NVIDIA Jetson TX2 edge computing module (the domestic Cambricon MLU220 can be selected as an alternative). Real-time preprocessing of traffic flow data for this road segment: redundant information is compressed using wavelet transform (compression rate ≥50%), retaining only lane occupancy (accuracy ±3%), mixed average speed (sampling 10Hz), and headway (outlier filtering threshold ±3σ). A local ant colony pheromone diffusion algorithm is run: initial pheromone concentration 0.2, volatile coefficient 0.25, iterating 30 times to generate a road segment-level traffic capacity heatmap with a resolution of 5m × 5m. The core parameters of the traffic flow pulse wave propagation model (wave velocity 8-12m / s, attenuation coefficient 0.3) are cached. When a communication delay with the regional layer >500ms is detected, the road segment health status (0-10 levels) is independently output based on local sensor data.

[0117] The regional scheduling layer is deployed according to city administrative divisions (e.g., the Pudong New Area node covers 120 square kilometers), and the computing nodes use dual-socket Xeon Silver 4310 processors (64 threads / 256GB memory). Key operations: Lightweight diagnostic results uploaded from the aggregation edge layer (data volume compressed to 15% of the original value, containing only congestion index and bottleneck point latitude and longitude coordinates); Run a lightweight spatiotemporal graph convolutional sub-model (input size 512×512×4, output demand distribution for the next 15 minutes in this partition). The branch and bound algorithm is used to solve the vehicle charging scheduling problem: the solution space is compressed by 40% through the constraint pruning strategy, and the time taken for a single calculation is ≤30 seconds. The central decision-making global server is equipped with four AMD EPYC 9654 processors (384 threads / 2TB memory) and performs the following: Integrate the health status maps of the road network in each zone (spatial stitching accuracy ±20 meters).

[0118] Cross-regional charging resource allocation: When the idle rate of charging piles in zone A is greater than 40% and that in zone B is less than 10%, the charging diversion strategy is triggered. Manage Hyperledger Fabric blockchain nodes (32 endorsed nodes in total) and verify data consistency every 0.1 seconds.

[0119] Communication and migration mechanisms: Edge → Area Layer: Transmit key fields (data packets ≤ 1KB) via 5G URLLC channel; Region → Central Layer: Upload is triggered when the health of the zone is <60% (congested mileage ratio) or the order gap rate is >20% (demand / supply ratio); Task migration protocol: When the CPU load of a regional node is >80% for 5 minutes, 50% of the computing tasks will be automatically migrated to an adjacent node with a load <40% (migration delay ≤800ms).

[0120] This distributed architecture significantly improves system robustness and real-time performance. Edge-layer localized processing compresses traffic flow data by over 50%, reducing network transmission pressure and maintaining segment-level decision-making capabilities during communication interruptions, ensuring uninterrupted basic functions. The regional layer, through lightweight models and optimization algorithms, doubles the computational efficiency of partitioned scheduling, and the branch-and-bound algorithm reduces solution time to within 30 seconds. The central layer's global optimization enables cross-regional resource collaboration, and charging pile allocation strategies alleviate local resource constraints. Layered communication mechanisms reduce redundant data transmission, and the task migration protocol dynamically balances computational load. Real-world testing shows that the overload rate of regional nodes has decreased to one-tenth of that of traditional architectures. The overall architecture supports linear expansion; adding new regions only requires deploying corresponding nodes, and the system maintains real-time decision-making even with tens of millions of vehicles.

[0121] Example 1: Prioritizing the dispatch of vehicles suitable for disabled persons around hospitals during morning rush hour Pain points in the scenario: During the morning rush hour, the congestion index around the municipal hospital reaches 8.2, with manually driven vehicles accounting for 75%, and taxis suitable for disabled people are often squeezed out by regular orders.

[0122] Technical Implementation: The demand forecasting module integrates social media tags for medical appointments (keywords "dialysis" and "pregnancy checkup" with a 90% match rate), and outputs a 50% surge in demand in the hospital area within the next 15 minutes. The dynamic scheduling module associates with the road network health status map: it marks the branch road west of the hospital as the source of congestion (health level 4.3 / 10) and generates detour routes; The energy coordination module calculates CDEI = 1.1 (>θ). high =1.0), and a charging price discount strategy will be launched (price reduction of 0.2 yuan / kWh), but vehicles for disabled persons will still be exempt from price adjustment (maintaining the base price).

[0123] If the number of charging stations within 500 meters of the hospital is greater than or equal to the number of disabled vehicles awaiting dispatch, multi-objective planning assigns an urgency level e to the disabled vehicles. k =3, dynamic weight Λ=0.48 (vehicle age 2.5 years), priority is given to charging stations within 500 meters of hospitals; Onboard unit execution: AR windshield display shows yielding intent (green ripples) and dedicated lane trajectory; Blockchain evidence storage decision parameters (timestamp + congestion index + CDEI value).

[0124] Effect verification: Charging demand stimulus strategies (price reductions) increased the charging volume of ordinary vehicles by 22%; Vehicles for disabled persons are exempt from the exemption mechanism and are subject to high urgency (e k =3), charging availability is increased to 100%; The utilization rate of charging piles increased from 65% to 89%.

[0125] Example 2: Automatic Triggering of Fleet Formation Driving Mode During Thunderstorms Pain point of the scenario: Yellow thunderstorm warning (wind speed 14m / s, visibility 400m), single vehicle traffic efficiency decreases by 40%.

[0126] Technical Implementation: The weather warning interface triggers a grouping instruction: 32 vehicles are grouped into 6-8 vehicle formations, with a group spacing of 15 meters. Traffic flow analysis module reconstructs vehicle flow pulse wave model: The wave speed decreased from 10 m / s to 7 m / s (due to correction for the slippery coefficient). Generate dedicated waveguide channels for the fleet (road network topology width ≥ 3.5 meters and no water accumulation sections); Communication degradation handling: When the communication delay of the roadside unit is greater than 300ms, the local cached model parameters (wave velocity 7m / s) are called. Vehicle control unit synchronization commands: Acceleration / braking command time deviation ≤ 0.2 seconds; AR projection of blue convoy outlines and safe following distance warnings; Charging strategy linkage: Battery temperature compensation mechanism limits charging rate to ≤1C (ΔT) comp =18℃), when ΔT comp At ≥15℃, the charging rate is limited to 1C.

[0127] Effect verification: The convoy travels at a speed of 38 km / h (22 km / h in single-vehicle mode). Zero accidents (the accident rate of traditional systems is 3.2 times per 100 vehicles during the same period); Battery over-temperature alarm frequency has decreased to 0.1 times per 10,000 vehicles per day.

[0128] Example 3: Utilizing the idle time of shopping mall charging stations to conduct price negotiations for charging. Pain point of the scenario: The mall charging station has a 72% idle rate at noon, but the queue exceeds 1 hour during the evening peak.

[0129] Technical Implementation: The demand forecasting module integrates shopping mall promotional Weibo tags to predict a 40% charging demand shortfall rate from 2:00 PM to 4:00 PM. CDEI dynamic adjustment: The midday CDEI is calculated as follows: Input parameter: ΔP t =-0.1 yuan (electricity price reduced by 0.1 yuan), P base =0.8 yuan / kWh, Qidle,t=72 (number of idle charging piles), Q total =100 (total number of charging piles), Scritical,t=5 (number of vehicles with less than 15% battery power), N vehicle =Total number of vehicles in area 50), α=0.4, β=0.3, γ=0.3.

[0130] Calculation steps: Electricity price fluctuation rate: =−0.125; The percentage of vacant parking spaces is 72 / 100 = 0.72. Emergency vehicles account for 0.1% of the total. CDEI=0.196 (<θ) low=0.8), triggering a demand suppression strategy: price increase of 0.3 yuan / kWh; Evening peak CDEI = 1.25 (>θ) high The unit price was reduced by 0.2 yuan / kWh; Multi-objective programming: Battery degradation cost calculation incorporates temperature compensation: Midday ambient temperature 32℃, compensation item C comp =0.008; Dynamic weight Λ=0.33 (vehicle age 1.8 years) guides non-emergency vehicles to charge at midday; Elasticity coefficient correction: If the actual morning charging volume deviates by 18%, the grid weight α will be automatically lowered to 0.37. Blockchain-based evidence storage: Records time-of-use electricity pricing strategies and revenue data, updating the encrypted hash every 0.1 seconds; Effect verification: Midday charging volume increased by 25% (compared to the baseline without strategy). The number of vehicles queuing for charging during the evening rush hour has decreased from 50 to 12. The average daily turnover rate of charging piles has increased from 3.1 times to 3.8 times per pile.

[0131] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A dynamic dispatching system for autonomous taxis in mixed traffic flow, characterized in that, The application relates to a dynamic dispatching system for mixed traffic flow of manual driving vehicles and automatic driving taxis, comprising the following steps: a scheduling server is connected with a roadside perception unit, a vehicle control unit and a charging station management platform; the roadside perception unit is arranged on each road section in a target area to collect traffic flow data of mixed driving speed of manual driving vehicles and automatic driving taxis in real time; the charging station management platform provides charging pile state data and time-of-use electricity price data in real time; the scheduling server comprises: a traffic flow analysis module for inputting traffic flow data of the roadside perception unit, executing an ant colony pheromone diffusion algorithm to generate a road traffic capacity thermal map, and inputting a traffic pulse wave propagation model to output a road network health state diagnosis result; a demand prediction module for fusing historical travel data of a network car-hailing platform, social media event label data and meteorological monitoring data, and generating a 15-minute-granularity travel demand distribution prediction through a spatio-temporal graph convolutional neural network model; a dynamic scheduling module for generating a dynamic scheduling instruction set comprising vehicle path re-planning instructions and empty vehicle dispatching coordinates according to the travel demand distribution prediction and the road network health state diagnosis result; an energy coordination module for combining data of the charging station management platform and vehicle battery remaining capacity data, calculating a charging demand elasticity index, dynamically adjusting a charging pricing strategy based on the index, and generating a charging scheduling scheme through a multi-objective programming model of a battery attenuation cost function and an order income function; the vehicle control unit is configured to receive and execute the dynamic scheduling instruction set and the charging scheduling scheme, and is connected with an augmented reality display unit and a blockchain storage unit.

2. The mixed traffic flow oriented dynamic dispatching system for autonomous taxis according to claim 1, wherein, When the traffic flow analysis module executes the traffic pulse wave propagation model, the following steps are specifically configured: real-time acquisition of basic traffic flow parameters of each road section in the target area, including mixed average speed, headway and lane occupancy rate of manual driving vehicles and automatic driving taxis; construction of a road network pulse wave propagation model: the road section is abstracted as a waveguide unit, the traffic pressure fluctuation value is calculated based on the basic traffic flow parameters, and the propagation process of the traffic pressure wave between adjacent road sections is simulated through a preset fluctuation equation; generation of a road network health degree diagnosis result: according to the attenuation rate and propagation time delay of the traffic pressure wave, a road network health state atlas containing road section level congestion index, bottleneck position and propagation path is dynamically output.

3. The mixed traffic flow oriented dynamic dispatching system for autonomous taxis according to claim 2, wherein, The traffic flow analysis module further comprises a data conflict arbitration unit and a communication degradation processing unit, wherein: the data conflict arbitration unit is configured to: when the mixed average speed, headway or lane occupancy rate data reported by the roadside perception unit and the vehicle control unit on the same road section differs by more than a preset threshold, dynamically select the data source to be used preferentially based on the reliability weight of the historical data source of the road section and the real-time sensor fault diagnosis result; the communication degradation processing unit is configured to: when the communication delay with the roadside perception unit or the vehicle control unit exceeds a set value, automatically call the road network health state atlas of the historical same period for interpolation compensation, and execute calculation by using the latest effective traffic pulse wave propagation model parameters stored in the local cache of the vehicle control unit to generate an approximate road network health state diagnosis result.

4. The mixed traffic flow oriented dynamic dispatching system for autonomous taxis according to claim 1, wherein, When the energy coordination module calculates the charging demand elasticity index, the following steps are executed: Real-time acquisition of power grid dynamic parameters: obtain the time-of-use electricity price data provided by the charging station management platform, the charging pile idle rate data, and the battery remaining capacity distribution data of the autonomous taxi in the target area; Constructing a CDEI calculation model: dynamically generating a regional charging demand elasticity index based on the following formula: ； wherein, ΔP t is the electricity price floating value relative to the reference electricity price P base at time t; Q idle,t is the number of idle charging piles at time t; Q total is the total number of charging piles; S critical,t is the number of vehicles with battery power below the warning threshold; N vehicle is the total number of autonomous taxis in the target area that need to be dispatched, obtained in real time from the dispatch server; and α, β, and γ are the weight coefficients of the power grid load, pile resource scarcity, and vehicle urgency, respectively. Dynamic adjustment of charging strategy: When CDEI t > θ high a charging price discount strategy is triggered, θ high is a charging demand stimulation threshold; When CDEI t <θ low the charging demand suppression strategy is initiated, θ low is the charging demand suppression threshold; Linkage charging scheduling scheme: input the adjusted charging strategy into the multi-objective planning model to optimize the generation of charging instruction set including charging pile allocation scheme and vehicle scheduling path.

5. The mixed traffic flow oriented dynamic dispatching system for autonomous taxis according to claim 4, wherein, The energy coordination module further includes an elasticity coefficient dynamic correction unit, which is configured to: Real-time monitoring of the actual execution effect of the charging strategy formulated based on the charging demand elasticity index, calculating the deviation rate of the actual charging capacity after the implementation of the strategy and the predicted charging capacity; When the deviation rate exceeds the preset threshold, automatically dynamically adjust the grid load weight coefficient, pile resource scarcity weight coefficient and vehicle emergency weight coefficient according to the grid load mutation indication signal and the actual deviation; During the dynamic adjustment of the weight coefficients, the dynamic scheduling module is linked to temporarily freeze the charging pile allocation scheme generated based on the current CDEI t value until the new weight coefficient correction is completed.

6. The mixed traffic flow oriented automated taxi dispatch system of claim 1, wherein, The multi-objective planning model of the energy coordination module generates a charging scheduling scheme according to the following steps: Constructing a battery cycle attenuation cost function: ; where, C bat (i,j) is the decay cost of vehicle i using charging pile j; r j is the charging rate of charging pile j; ΔT j = |T j env -T i opt | is the difference between the ambient temperature and the optimal battery temperature; ai, bi are the battery decay coefficients of vehicle i; Constructing an order revenue function: ; wherein R order (i, k) is the revenue of vehicle i performing order k; F k is the order base fare; s k is the order mileage; θ is the distance sensitivity parameter, wherein the value of highway is 1.2 and the value of urban area is 0.8; P k is the time premium benchmark value; e k is the urgency level, wherein the value of hospital is 3, the value of flight is 2, and the value of ordinary is 1; δ is the urgency amplification factor, and the value is 1.5; w k is the passenger waiting time; and ρ is the waiting penalty coefficient, and the value is 0.

1. Generating a matching scheme: ; Dynamic weight factor: ; wherein X ijk is a decision variable, taking the value 1 when matching and 0 when not matching; A0is a base weight, taking the value 0.3; Y i is the age of the vehicle i; v is an age gain coefficient, taking the value 0.15; t is a threshold for the age segmentation, taking the value 2 years; Setting hard constraints: Single pile constraint, each vehicle uses at most one charging pile; Power guarantee, the power after charging ≥ order minimum demand + redundancy; Time constraint, charging time + travel time ≤ order deadline.

7. The mixed traffic flow oriented automated taxi dispatch system of claim 6, wherein, When the energy coordination module generates a charging scheduling scheme using the multi-objective planning model, it further executes a battery temperature real-time compensation mechanism, wherein: Real-time receiving of vehicle battery actual working temperature reported from the vehicle-mounted control unit; Introducing the difference between the actual working temperature of the battery and the ambient temperature of the charging pile as a compensation factor to dynamically correct the attenuation cost calculation result; When the difference between the actual working temperature of the battery and the ambient temperature is ≥ 15℃, forcibly limit the maximum charging rate of the charging pile available to the vehicle.

8. The mixed traffic flow oriented automated taxi dispatch system of claim 1, wherein, The dynamic scheduling module is also connected with a road network topology mutation perception unit, which is configured to: Real-time access to municipal construction announcement interface and receive abnormal event information reported from vehicle-mounted camera or roadside unit; When detecting an event that changes the road network topology structure in the target area, such as road closure, lane number change or new control area, generate the corresponding topology mutation event code; Trigger the traffic flow analysis module to dynamically reconstruct the traffic pulse wave propagation model and adjust the waveguide unit segmentation parameters of the corresponding road segment in the model according to the actual traffic capacity change; Send the path feasibility state flag to the energy coordination module and dynamic scheduling module to prohibit the scheduling vehicle from entering the area that cannot be passed due to topology mutation.

9. The mixed traffic flow oriented automated taxi dispatch system of claim 1, wherein, When the dynamic scheduling module is connected with the meteorological warning interface, when the thunderstorm warning level is ≥ yellow, execute: Group the autonomous taxis in the target area into a fleet; Generate a special traffic waveguide channel for the fleet based on the traffic pulse wave propagation model under the condition of road network topology; Synchronize the acceleration / braking instructions of the fleet vehicles through the vehicle-mounted unit.

10. The collaborative decision-making method for the mixed traffic flow oriented autonomous taxi dynamic dispatching system according to any one of claims 1-9, characterized in that, The following steps are executed by the scheduling server: Real-time acquisition of mixed traffic flow data collected by the roadside perception unit, including the driving speed of manually driven vehicles and autonomous taxis; Fuse the historical trip data of the ride-hailing platform, social media event label data and meteorological monitoring data, input the spatio-temporal graph convolutional neural network model, and output the travel demand distribution prediction of the next 15 minutes granularity; Based on the driving speed in the mixed traffic flow data, an ant colony pheromone diffusion algorithm is used to construct a road capacity heat map; Input the road capacity heat map into the traffic pulse wave propagation model to diagnose the road network health status, which includes congestion propagation direction and bottleneck section markers; According to the travel demand distribution prediction and the road network health status, generate a dynamic scheduling instruction set for autonomous taxis, which includes path re-planning instructions and empty vehicle scheduling coordinates; Receive the charging pile usage state and time-of-use electricity price data sent by the charging station management platform, and calculate the charging demand elasticity index combined with the vehicle battery remaining capacity; Based on the charging demand elasticity index, dynamically adjust the charging service pricing strategy, and input the battery attenuation cost function and order revenue function into the multi-objective planning model to generate a charging scheduling scheme; The dynamic scheduling instruction set and the charging scheduling scheme are sent to the vehicle control unit to trigger the following operations: Control the autonomous taxi to execute path change or charging station navigation; Through augmented reality windshield, dynamically display yielding intention markers, lane change trajectory preview and scheduling decision basis summary; Write the vehicle decision timestamp and scheduling basis parameters into the blockchain distributed ledger to generate an encrypted verification record updated every 0.1 seconds.

Images