Traffic management method considering combination of causal inference and dynamic path induction

Abstract

The invention provides a traffic management method considering the combination of causal inference and dynamic path induction, and a globally optimal traffic induction scheme is generated by quantitatively analyzing the time-space diffusion effect of an accident and combining real-time monitoring data. According to the invention, local congestion caused by traffic accidents can be quickly relieved, the stability of a road network is recovered to the maximum extent, and the efficiency of urban emergency traffic management and the overall traffic capacity of the system are improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation technology, and relates to a traffic flow dynamic induction technology considering the temporal and spatial influence of accidents, and in particular to a traffic management method combining causal inference with dynamic path induction. Background Art

[0002] With the development of social economy and the acceleration of urbanization, the contradictions brought about by the limited urban traffic resources have become increasingly prominent, specifically manifested in the imbalance between traffic demand and supply of urban road networks, environmental pollution, and traffic safety issues. In order to cope with complex traffic problems, a series of traffic control and management solutions are urgently needed. Urban traffic status reflects the time-varying, multi-scale, multi-variable and random characteristics of traffic congestion in urban road networks. Quantifying the degree of traffic congestion can enable traffic management departments to grasp the spatiotemporal information of congested sections, which is the basis for implementing effective traffic control and management. In complex urban road networks, the complexity and dynamics of traffic flow make it extremely challenging to evaluate the congestion level of each section. Therefore, it is urgent to develop a scientific and dynamic urban road section traffic induction technology to provide a reference for efficient decision-making in traffic management.

[0003] Research in related fields has made some progress. For example, the paper A Real-Time Traffic Flow Prediction System Based on Deep Learning Models proposed a real-time traffic flow prediction system based on a deep learning model, which uses historical traffic data and real-time monitoring information to predict the traffic flow of urban roads, thereby realizing the evaluation of future traffic conditions. Although this method performs well in traffic flow prediction, it is not adequate for dealing with congestion caused by emergencies, and it is difficult to accurately capture changes in traffic flow after an accident, resulting in the induction scheme being unable to be quickly adjusted in an emergency.

[0004] The paper Dynamic Route Guidance Using Reinforcement Learning in Urban Road Networks proposes a dynamic route guidance method for urban road networks based on reinforcement learning. This method uses reinforcement learning technology to provide vehicles with optimal routes based on changes in road traffic conditions in order to minimize individual travel. Although this method can achieve dynamic route guidance to a certain extent, it is limited to the selection of the optimal individual route and lacks consideration of the overall optimization of the global road network. When a traffic accident occurs, the selection of the optimal individual route may lead to new traffic bottlenecks, further exacerbating traffic congestion and failing to achieve the overall optimization of the entire road network system.

[0005] In addition, the paper Urban Traffic State Evaluation by Integrating Multiple DataSources proposed an urban traffic state evaluation method based on multi-source data fusion, using traffic monitoring data, environmental information and other sensor data to evaluate traffic congestion through multi-level information fusion. Although this technical solution improves the accuracy of traffic state evaluation through data fusion, the information fusion process is highly complex, and the evaluation method lacks consideration of the spatiotemporal diffusion effect of accidents, and cannot respond in a timely manner to complex road network environment changes caused by sudden accidents. Summary of the invention

[0006] Technical problem: In order to solve the above problems, the present invention proposes a method based on the combination of causal inference and dynamic path induction, which can effectively deal with the spatiotemporal diffusion effect and global impact of road network caused by traffic accidents. The method first quantitatively analyzes the impact of traffic accidents through a causal inference model, and quantifies the spatiotemporal diffusion effect of accidents on traffic flow in combination with real-time monitoring information of the global road network. On this basis, a system-optimal dynamic traffic allocation model is constructed to generate a globally optimal dynamic path induction scheme with the goal of minimizing the total travel time of the entire road network. By combining the road network data calculated by the causal inference model with the actual traffic flow monitoring information, the changes in traffic flow after the accident can be analyzed and predicted more accurately, thereby achieving more scientific traffic induction. This method provides more scientific and effective support for traffic management and emergency response, and improves the ability to respond to emergencies and the overall traffic efficiency of the road network.

[0007] Technical solution:

[0008] A traffic management method combining causal inference and dynamic path induction includes the following steps:

[0009] Step 1. Obtain the topological structure information of the urban road network G from public channels, including basic data such as nodes (i, j), edges, road capacity c, number of lanes n and length L, in preparation for step 2;

[0010] Step 2: Construct a spatiotemporal impact inference model based on causal inference;

[0011] Step 3: Construct a dynamic traffic assignment model.

[0012] As a further preferred solution, step 2 includes:

[0013] S1. Obtain the accident impact area, define the initial position s and the end position t, denote the initial position s as the source node, and denote the end position t as the target node; obtain the GIS geographic information data in the evaluation road network area through networking, obtain the dynamic traffic flow information of the road network based on the historical road network information, including the corresponding vehicle flow and other traffic status information on all road sections in the accident impact area, and use the obtained data and information to model the traffic network in the evaluation road network area;

[0014] S2. Preprocess the vehicle GPS trajectory data in the evaluation road network area, extract the vehicle density and vehicle flow information of each road section, and calculate the speed information of each road section including the current speed v of each road section within the time interval t ij (t), prepare for calculating the travel time T for S3 (i,j)

[0015]

[0016] v ij : The average speed on the road segment (i, j) (unit: km / h or m / s);

[0017] q ij : The flow rate on the road segment (i, j) (unit: vehicles / hour);

[0018] ρ ij : Traffic density on road segment (i, j) (unit: vehicles / hour);

[0019] S3. Based on the road network constructed in S1 and the speed information in S2, calculate the travel time T of the vehicle on each road section (i,j) , as the initial road resistance of each road section;

[0020]

[0021] L (i,j) : Indicates the length of the road section;

[0022] T (i,j) : Indicates the travel time of the vehicles affected by the accident within the road section;

[0023] S4. Definitions is the travel time of the road section connected by the i-th and j-th intersections after the accident dur, R (i,j) ∈(0,1) indicates whether an accident occurs, where dur indicates the time (min) after the accident occurs, and the value corresponding to dur is used as an update cycle; multiple control variables X are introduced to eliminate the influence of heterogeneity, and the conditional average treatment effect (CATE) is calculated;

[0024] S5. Pre-select a series of variables as control variables. Control variables are a type of variable that affects the type, location, and time of an accident. According to the variable type, they can be divided into basic attribute variables. Traffic state variables Road Linear Variable Traffic signal variables Screen the control variables to reduce the bias; use the Pearson correlation coefficient to eliminate variables with strong correlation, delete variables that are not conducive to estimation (non-confounding variables), and use the borutaSHAP method as the basis for judging variable elimination; this step is to preliminarily screen the control variables (including environmental attributes, traffic status and road attribute variables), and use the Pearson correlation coefficient and borutaSHAP method to eliminate irrelevant or interfering variables, aiming to reduce the bias of the model and ensure the scientificity and explanatory power of the estimation results;

[0025] S6. Construct a causal relationship diagram based on the causal relationship between the variables;

[0026] S7. Construct a Doubly Robust Learning (DRL) causal inference model and proceed to S8. This step uses the DRL method combined with control variables to construct a causal inference model and calculate CATE, thereby accurately estimating the causal effect of the accident, solving the model bias problem, and improving the robustness and explanatory power of causal inference.

[0027]

[0028] in, Therefore, the estimated value of the accident causal effect CATE is:

[0029]

[0030] S8. Based on the causal relationship diagram constructed in S6, Doubly Robust Learning (DRL) causal inference is performed on the speed impact values ​​under different dur, and based on the estimated value of the speed impact on different roads under the influence of the accident, the speed of each road under different dur is calculated. Through the real-time changes in traffic flow and speed, the travel time T1 of the vehicles affected by the accident under different dur is obtained to prepare for step three.

[0031] As a further preferred solution, S4 of step 2 includes:

[0032] S4-1. Calculate the time impact effect based on the parameters in S4: ATE is the average treatment effect of the entire sample, reflecting the overall impact of the accident on the travel time;

[0033]

[0034]

[0035] The calculated value represents the impact of the accident on travel time. The larger the ATE, the greater the negative impact of the accident on travel time. This means that the congestion and delay caused by the accident are more serious, and stronger intervention measures (such as improved traffic management) are needed.

[0036] S4-2. In order to solve the problem of heterogeneity, multiple control variables are usually introduced represents the control variable for the ith and jth intersections under dur; therefore, the estimated result is the conditional average treatment effect (CATE):

[0037]

[0038] The larger the value of CATE under x, the more serious the impact of the accident.

[0039] As a further preferred solution, S8 of step 2 includes:

[0040] S8-1. Calculate the propensity score under different conditional variable combinations:

[0041]

[0042] S8-2. Construct inverse propensity weighting (IPW) to solve the problem of sample imbalance, that is, use propensity scores to solve the problem of uneven sample distribution (IPW):

[0043]

[0044] S8-3. At this time, the ATE estimate of IPW is (the robustness of the estimate is further improved by the DRL method):

[0045]

[0046] Since the sensitivity of IPW estimates is too high, an additional The regression model forms the DRL estimate:

[0047]

[0048] in, Using classification machine learning for estimation, Using regression machine learning for estimation:

[0049]

[0050] in, is the travel time between the ith intersection and the jth intersection under dur;

[0051] By comparing the travel time of the "treatment group" (with some intervention) and the travel time of the "control group" (without intervention), a specific CATE value is obtained.

[0052] As a further preferred solution, step three includes:

[0053] Based on the results of causal inference, the system constructs dynamic traffic allocation to optimize the global traffic flow distribution; the objective function of the model is set to minimize the total travel time or total travel cost of the entire network, especially in emergency situations to ensure the balance and rapid evacuation of traffic; the decision variable is defined as the traffic flow on each path, which is subject to multiple constraints, including flow conservation constraints, road capacity constraints, accident impact constraints and path uniqueness constraints;

[0054] S1. Dynamic traffic distribution (calculate traffic distribution between all intersections);

[0055] For a given road network, E is the set of the i-th and j-th intersections, the i-th and j-th intersections must be two adjacent intersections, P is the p(i,j) set, p(i,j) represents the road section from node i to node j, and TT represents the total travel time of all traffic demands; the optimal dynamic traffic assignment model based on the system of individual vehicles with path constraints can be expressed as: minTT = ∑ (i,j) T1(i,j), that is, the total travel time of vehicles in the area affected by the accident is minimized; the dynamic traffic assignment model is solved using mathematical programming software GAMS (The General Algebraic Modeling System), and the output path set Ω is prepared for step three S2; (the goal of S1 is to minimize the overall traffic time TT, which belongs to global optimization);

[0056] Objective function: min TT = ∑ (i,j) T1(i,j)

[0057] Add constraints:

[0058] Flow conservation constraint: the incoming flow of each road section is equal to the outgoing flow; q is the flow, and p represents the road section between the i-th and j-th intersections;

[0059]

[0060] Road capacity constraint: the traffic flow of each road segment does not exceed its maximum capacity; C ij is the maximum capacity of road traffic flow

[0061]

[0062] Accident impact constraint: the capacity of the road segment affected by the accident is reduced; γ1 represents the parameter affecting the capacity of the road segment;

[0063] Road sections affected by the accident

[0064] Rescue impact constraint: The capacity of the road segment affected by the emergency rescue path is greatly reduced; γ2 represents the parameter that affects the capacity of the road segment.

[0065] Road sections affected by the accident

[0066] Non-negativity constraint: All traffic flow variables must be non-negative;

[0067]

[0068] S2. Generate a traffic induction plan (based on the results of S1, further generate an executable traffic induction plan);

[0069] Design a traffic induction scheme, where Z is the sum of the squares of the difference between the actual service traffic demand of all paths and the expected allocated traffic demand. The smaller Z is, the better the induction scheme is, and the allocated demand meets the actual service traffic; minimize the sum of the squares of the difference between the actual service traffic demand of all paths and the expected allocated traffic demand, that is, is the objective function, and Z is minimized. For a specific path, the traffic demand it actually serves consists of four parts, namely, the number of vehicles that choose path i among vehicles without on-board induction terminals installed, the number of vehicles that comply with the induction plan and do not choose emergency rescue path i among vehicles with on-board induction terminals installed, the number of vehicles that do not comply with the induction plan but do not choose emergency rescue path i among vehicles with on-board induction terminals installed, and the number of vehicles that do not comply with the induction plan and choose emergency rescue path i among vehicles with on-board induction terminals installed; GAMS software is used to solve the above problem and generate an induction plan for implementation (the goal of S2 is to minimize the difference between the traffic induction plan and the actual traffic demand, which belongs to the local optimization and implementation level. Finally, a set of regional induction plans after an accident is obtained, which can effectively and quickly solve the congestion caused by the accident);

[0070] Objective function:

[0071] Constraints:

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] Where: Z represents the sum of squares of the difference between the actual service traffic demand of each path and the expected allocated traffic demand;

[0079] i and j represent the path numbers;

[0080] Ω represents the path set;

[0081] It represents the number of vehicles that choose path i among the vehicles without vehicle guidance terminals installed;

[0082] represents the number of vehicles installed with on-board guidance terminals that comply with the guidance plan and do not choose emergency rescue path i;

[0083] It indicates the number of vehicles installed with vehicle-mounted guidance terminals that did not comply with the guidance plan but did not choose emergency rescue path i;

[0084] The number of vehicles that are equipped with vehicle-mounted guidance terminals but do not comply with the guidance plan and choose emergency rescue path i;

[0085] represents the expected distribution of traffic demand for path i;

[0086] P represents the total traffic demand between the OD point pairs;

[0087] θ represents the coverage rate of vehicle-mounted induced terminals;

[0088] α i It represents the proportion of vehicles that choose route i in the total demand under the condition of no induced information;

[0089] γ i Indicates the proportion of induced path i in all the induced information sent;

[0090] η represents the induced compliance rate;

[0091] β i It represents the proportion of the expected traffic demand of path i in the total demand;

[0092] a represents the proportion of vehicles that choose emergency rescue route i

[0093] k represents a route in the Ω path set.

[0094] Beneficial effects:

[0095] 1. The present invention analyzes the spatiotemporal diffusion effect of traffic accidents through a causal inference model, and combined with real-time monitoring information of the global road network, can more accurately evaluate the impact of accidents on the traffic flow of the entire road network, thereby realizing a more scientific traffic induction strategy and improving the ability to respond to emergencies.

[0096] 2. The present invention introduces the Doubly Robust Learning (DRL) model in the path induction process. By combining the propensity score model with the outcome regression model, the robustness and stability of the induction scheme are enhanced, effectively solving the problem of poor induction effect caused by uncertainty in traditional path induction methods.

[0097] 3. The method of the present invention can achieve accurate induction of the global road network by comprehensively analyzing vehicle GPS trajectory data, historical traffic flow data and real-time traffic status without relying on high-density sensors, reducing the complexity and cost of sensor settings, making the solution feasible for application in a wider range.

[0098] 4. The dynamic traffic allocation model of the present invention adopts the system-optimal induction strategy with the goal of minimizing the total travel time of the entire road network. It can effectively alleviate local traffic congestion caused by accidents and restore the stable operation of the entire traffic system in a short time, thereby maximizing the overall traffic efficiency of the road network.

[0099] 5. The present invention continuously optimizes the traffic induction scheme through real-time traffic data feedback, and can make rapid adjustments when traffic conditions change, ensuring the real-time and adaptability of the induction effect, which is particularly suitable for emergency traffic management environments after accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 Flow chart of the method of the present invention;

[0101] Figure 2 Flowchart of spatiotemporal impact inference based on causal inference;

[0102] Figure 3 Dynamic traffic assignment flow chart;

[0103] Figure 4 Cause and effect diagrams; DETAILED DESCRIPTION

[0104] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0105] like Figure 1As shown, a traffic management method combining causal inference and dynamic path induction of the present invention aims at the problems that the traffic flow in the traffic network under the influence of an accident has a single path induction and limited effect. A dynamic traffic assignment model is constructed that considers the causal effect of the accident on the traffic demand and is guided by the system optimization of the traffic network. The method guides the social vehicles under the influence of the accident in real time, and includes the following steps:

[0106] Step 1. Obtain the topological structure information of the urban road network G from public channels, including basic data such as nodes (i, j), edges, road capacity c, number of lanes n and length L, in preparation for step 2.

[0107] Step 2: Construct a spatiotemporal impact inference model based on causal inference, such as Figure 2 shown.

[0108] Among them, step two specifically includes:

[0109] S1. Obtain the accident impact area, define the initial position s and the end position t, denote the initial position s as the source node, and denote the end position t as the target node; obtain the GIS geographic information data in the evaluation road network area through networking, obtain the dynamic traffic flow information of the road network based on the historical road network information, including the corresponding vehicle flow and other traffic status information on all road sections in the accident impact area, and use the obtained data and information to model the traffic network in the evaluation road network area;

[0110] S2. Preprocess the vehicle GPS trajectory data in the evaluation road network area, extract the vehicle density and vehicle flow information of each road section, and calculate the speed information of each road section including the current speed v of each road section within the time interval t ij (t), prepare for calculating the travel time T for S3 (i,j)

[0111]

[0112] v ij : The average speed on the road segment (i, j) (unit: km / h or m / s);

[0113] q ij : The flow rate on the road segment (i, j) (unit: vehicles / hour);

[0114] ρ ij : Traffic density on road segment (i, j) (unit: vehicles / hour);

[0115] S3. Based on the road network constructed in S1 and the speed information in S2, calculate the travel time T of the vehicle on each road section (i,j) , as the initial road resistance of each road section;

[0116]

[0117] L (i,j) : Indicates the length of the road section;

[0118] T (i,j) : Indicates the travel time of the vehicles affected by the accident within the road section;

[0119] S4. Definitions is the travel time of the road section connected by the i-th and j-th intersections after the accident dur, R (i,j) ∈(0,1) indicates whether an accident occurs, where dur indicates the time (min) after the accident occurs, and the value corresponding to dur is used as an update cycle; multiple control variables X are introduced to eliminate the influence of heterogeneity, and the conditional average treatment effect (CATE) is calculated;

[0120] Step 2 S4 includes:

[0121] S4-1. Calculate the time impact effect based on the parameters in S4: ATE is the average treatment effect of the entire sample, reflecting the overall impact of the accident on the travel time;

[0122]

[0123]

[0124] The calculated value represents the impact of the accident on travel time. The larger the ATE, the greater the negative impact of the accident on travel time. This means that the congestion and delay caused by the accident are more serious, and stronger intervention measures (such as improved traffic management) are needed.

[0125] S4-2. In order to solve the problem of heterogeneity, multiple control variables are usually introduced represents the control variable for the ith and jth intersections under dur; therefore, the estimated result is the conditional average treatment effect (CATE):

[0126]

[0127] The larger the value of CATE under x, the more serious the impact of the accident;

[0128] S5. Pre-select a series of variables as control variables. Control variables are a type of variable that affects the type, location, and time of an accident. According to the variable type, they can be divided into basic attribute variables. Traffic state variables Road Linear Variable Traffic signal variables Screen the control variables to reduce the bias; use the Pearson correlation coefficient to eliminate variables with strong correlation, delete variables that are not conducive to estimation (non-confounding variables), and use the borutaSHAP method as the basis for judging variable elimination; this step is to preliminarily screen the control variables (including environmental attributes, traffic status and road attribute variables), and use the Pearson correlation coefficient and borutaSHAP method to eliminate irrelevant or interfering variables, aiming to reduce the bias of the model and ensure the scientificity and explanatory power of the estimation results;

[0129]

[0130] Table 1 Preliminary selection of control variables

[0131] S6. Based on the causal relationship between the variables, a causal relationship diagram is constructed; as follows Figure 4 As shown;

[0132] S7. Construct a Doubly Robust Learning (DRL) causal inference model and proceed to S8. This step uses the DRL method combined with control variables to construct a causal inference model and calculate CATE, thereby accurately estimating the causal effect of the accident, solving the model bias problem, and improving the robustness and explanatory power of causal inference.

[0133]

[0134] in, Therefore, the estimated value of the accident causal effect CATE is:

[0135]

[0136] S8. Based on the causal relationship diagram constructed in S6, Doubly Robust Learning (DRL) causal inference is performed on the speed impact values ​​under different dur, and based on the estimated value of the speed impact on different roads under the influence of the accident, the speed of each road under different dur is calculated. Through the real-time changes in traffic flow and speed, the travel time T1 of the vehicles affected by the accident under different dur is obtained to prepare for step three;

[0137] Step 2 S8 includes:

[0138] S8-1. Calculate the propensity score under different conditional variable combinations:

[0139]

[0140] S8-2. Construct inverse propensity weighting (IPW) to solve the problem of sample imbalance, that is, use propensity scores to solve the problem of uneven sample distribution (IPW):

[0141]

[0142] S8-3. At this time, the ATE estimate of IPW is (the robustness of the estimate is further improved by the DRL method):

[0143]

[0144] Since the sensitivity of IPW estimates is too high, an additional The regression model forms the DRL estimate:

[0145]

[0146] in, Using classification machine learning for estimation, Using regression machine learning for estimation:

[0147]

[0148] in, is the travel time between the ith intersection and the jth intersection under dur;

[0149] By comparing the travel time of the "treatment group" (with some intervention) and the travel time of the "control group" (without intervention), a specific CATE value is obtained.

[0150] Step 3: Construct a dynamic traffic assignment model, such as Figure 3 shown.

[0151] Step three specifically includes:

[0152] Based on the results of causal inference, the system constructs dynamic traffic allocation to optimize the global traffic flow distribution; the objective function of the model is set to minimize the total travel time or total travel cost of the entire network, especially in emergency situations to ensure the balance and rapid evacuation of traffic; the decision variable is defined as the traffic flow on each path, which is subject to multiple constraints, including flow conservation constraints, road capacity constraints, accident impact constraints and path uniqueness constraints;

[0153] S1. Dynamic traffic distribution (calculate traffic distribution between all intersections)

[0154] For a given road network, E is the set of the i-th and j-th intersections, the i-th and j-th intersections must be two adjacent intersections, P is the p(i,j) set, p(i,j) represents the road section from node i to node j, and TT represents the total travel time of all traffic demands; the optimal dynamic traffic assignment model based on the system of individual vehicles with path constraints can be expressed as: minTT = ∑ (i,j)T1(i,j), that is, the total travel time of vehicles in the area affected by the accident is minimized; the dynamic traffic assignment model is solved using mathematical programming software GAMS (The General Algebraic Modeling System), and the output path set Ω is prepared for step three S2; (the goal of S1 is to minimize the overall traffic time TT, which belongs to global optimization);

[0155] Objective function: min TT = ∑ (i,j) T1(i,j)

[0156] Add constraints:

[0157] Flow conservation constraint: the incoming flow of each road section is equal to the outgoing flow; q is the flow, and p represents the road section between the i-th and j-th intersections;

[0158]

[0159] Road capacity constraint: the traffic flow of each road segment does not exceed its maximum capacity; C ij is the maximum capacity of road traffic flow

[0160]

[0161] Accident impact constraint: the capacity of the road segment affected by the accident is reduced; γ1 represents the parameter affecting the capacity of the road segment;

[0162] Road sections affected by the accident

[0163] Rescue impact constraint: The capacity of the road segment affected by the emergency rescue path is greatly reduced; γ2 represents the parameter that affects the capacity of the road segment.

[0164] Road sections affected by the accident

[0165] Non-negativity constraint: All traffic flow variables must be non-negative;

[0166]

[0167] S2. Generate a traffic induction plan (based on the results of S1, further generate an executable traffic induction plan) Design a traffic induction plan, where Z is the square sum of the difference between the actual service traffic demand of all paths and the expected distribution traffic demand. The smaller Z is, the better the induction plan is, and the distribution demand meets the actual service traffic; minimize the square sum of the difference between the actual service traffic demand of all paths and the expected distribution traffic demand, that is, is the objective function, and Z is minimized. For a specific path, the traffic demand it actually serves consists of four parts, namely, the number of vehicles that choose path i among vehicles without on-board induction terminals installed, the number of vehicles that comply with the induction plan and do not choose emergency rescue path i among vehicles with on-board induction terminals installed, the number of vehicles that do not comply with the induction plan but do not choose emergency rescue path i among vehicles with on-board induction terminals installed, and the number of vehicles that do not comply with the induction plan and choose emergency rescue path i among vehicles with on-board induction terminals installed; GAMS software is used to solve the above problem and generate an induction plan for implementation (the goal of S2 is to minimize the difference between the traffic induction plan and the actual traffic demand, which belongs to the local optimization and implementation level. Finally, a set of regional induction plans after an accident is obtained, which can effectively and quickly solve the congestion caused by the accident);

[0168] Objective function:

[0169] Constraints:

[0170]

[0171]

[0172]

[0173]

[0174]

[0175]

[0176] Where: Z represents the sum of squares of the difference between the actual service traffic demand of each path and the expected allocated traffic demand;

[0177] i and j represent the path numbers;

[0178] Ω represents the path set;

[0179] It represents the number of vehicles that choose path i among the vehicles without vehicle guidance terminals installed;

[0180] represents the number of vehicles installed with on-board guidance terminals that comply with the guidance plan and do not choose emergency rescue path i;

[0181] It indicates the number of vehicles installed with vehicle-mounted guidance terminals that did not comply with the guidance plan but did not choose emergency rescue path i;

[0182] The number of vehicles that are equipped with vehicle-mounted guidance terminals but do not comply with the guidance plan and choose emergency rescue path i;

[0183] represents the expected distribution of traffic demand for path i;

[0184] P represents the total traffic demand between the OD point pairs;

[0185] θ represents the coverage rate of vehicle-mounted induced terminals;

[0186] α i It represents the proportion of vehicles that choose route i in the total demand under the condition of no induced information;

[0187] γ i represents the proportion of induced path i in all the induced information issued; η represents the induced compliance rate;

[0188] β i It represents the proportion of the expected traffic demand of path i in the total demand; a represents the proportion of vehicles that choose emergency rescue path i; k represents a route in the Ω path set.

Claims

1. A traffic management method combining causal inference and dynamic path induction, characterized in that: The steps include: Step 1. Obtain the topological structure information of the urban road network G from public channels, including the basic data of nodes (i, j), edges, road capacity c, number of lanes n and length L, in preparation for step 2; Step 2: Construct a spatiotemporal impact inference model based on causal inference; Step 3: Construct a dynamic traffic assignment model.

2. A traffic management method combining causal inference and dynamic path induction according to claim 1, characterized in that: The second step comprises: S1. Obtain the accident impact area, define the initial position s and the end position t, represent the initial position s as the source node, and represent the end position t as the target node; obtain the GIS geographic information data in the evaluation road network area through networking, obtain the dynamic traffic flow information of the road network based on the historical road network information, including the corresponding vehicle flow traffic status information on all road sections in the accident impact area, and use the obtained data and information to model the traffic network in the evaluation road network area; S2. Preprocess the vehicle GPS trajectory data in the evaluation road network area, extract the vehicle density and vehicle flow information of each road section, and calculate the speed information of each road section including the current speed v of each road section within the time interval t ij (t), prepare for calculating the travel time T for S3 (i,j) v ij : The average speed on the road segment (i, j) (unit: km / h or m / s); q ij : The flow rate on the road segment (i, j) (unit: vehicles / hour); ρ ij : Traffic density on road segment (i, j) (unit: vehicles / hour); S3. Based on the road network constructed in S1 and the speed information in S2, calculate the travel time T of the vehicle on each road section (i,j) , as the initial road resistance of each road section; L (i,j) : Indicates the length of the road section; T (i,j) : Indicates the travel time of the vehicles affected by the accident within the road section; S4. Definitions is the travel time of the road section connected by the i-th and j-th intersections after the accident dur, R (i,j) ∈(0,1) indicates whether an accident occurs, where dur indicates the time (min) after the accident occurs, and the value corresponding to dur is used as an update cycle; multiple control variables X are introduced to eliminate the influence of heterogeneity, and the conditional average treatment effect (CATE) is calculated; S5. Pre-select a series of variables as control variables. Control variables are a type of variable that affects the type, location, and time of an accident. According to the type of variable, it can be divided into attribute variables. Traffic state variables Road Linear Variable Traffic signal variables Screen the control variables to reduce the bias; use the Pearson correlation coefficient to eliminate the variables with strong correlation, delete the variables that are not conducive to estimation (non-confounding variables), and use the borutaSHAP method as the basis for judging the elimination of variables; S6. Construct a causal relationship diagram based on the causal relationship between the variables; S7. Build a Doubly Robust Learning (DRL) causal inference model and proceed to S8; in, Therefore, the estimated value of the accident causal effect CATE is: S8. Based on the causal relationship diagram constructed in S6, Doubly Robust Learning (DRL) causal inference is performed on the speed impact values ​​under different dur, and based on the estimated value of the speed impact on different roads under the influence of the accident, the speed of each road under different dur is calculated. Through the real-time changes in traffic flow and speed, the travel time T1 of the vehicles affected by the accident under different dur is obtained to prepare for step three.

3. A traffic management method combining causal inference and dynamic path induction according to claim 2, characterized in that: Step 2 S4 includes: S4-1. Calculate the time impact effect based on the parameters in S4: ATE is the average treatment effect of the entire sample, reflecting the overall impact of the accident on the travel time; S4-2. In order to solve the problem of heterogeneity, multiple control variables are usually introduced represents the control variable for the ith and jth intersections under dur; therefore, the estimated result is the conditional average treatment effect (CATE):

4. A traffic management method combining causal inference and dynamic path induction according to claim 2, characterized in that: Step 2 S8 includes: S8-1. Calculate the propensity score under different conditional variable combinations: S8-2. Construct inverse propensity weighting (IPW) to solve the problem of sample imbalance: S8-3. At this time, the ATE of IPW is estimated to be: Since the sensitivity of IPW estimates is too high, an additional The regression model forms the DRL estimate: in, Using classification machine learning for estimation, Using regression machine learning for estimation: in, is the travel time between the i-th intersection and the j-th intersection under dur.

5. A traffic management method combining causal inference and dynamic path induction according to claim 1, characterized in that: The step three comprises: Based on the results of causal inference, the system constructs dynamic traffic allocation to optimize the global traffic flow distribution; the objective function of the model is set to minimize the total travel time or total travel cost of the entire network, especially in emergency situations to ensure the balance and rapid evacuation of traffic; the decision variable is defined as the traffic flow on each path, which is subject to multiple constraints, including flow conservation constraints, road capacity constraints, accident impact constraints and path uniqueness constraints; S1. Dynamic Traffic Assignment For a given road network, E is the set of the i-th and j-th intersections, the i-th and j-th intersections must be two adjacent intersections, P is the set of p(i,j), p(i,j) represents the road section from node i to node j, and TT represents the total travel time of all traffic demands; the optimal dynamic traffic assignment model based on the system of individual vehicles with path constraints can be expressed as: min TT = ∑ (i,j) T1(i,j), that is, the total travel time of vehicles in the area affected by the accident is minimized; the dynamic traffic assignment model is solved using mathematical programming software GAMS (The General Algebraic Modeling System), and the output path set Ω is prepared for step three S2; Objective function: min TT = ∑ (i,j) T1(i,j) Add constraints: Flow conservation constraint: the incoming flow of each road section is equal to the outgoing flow; q is the flow, and p represents the road section between the i-th and j-th intersections; Road capacity constraint: the traffic flow of each road segment does not exceed its maximum capacity; C ij is the maximum capacity of road traffic flow Accident impact constraint: the capacity of the road segment affected by the accident is reduced; γ1 represents the parameter affecting the capacity of the road segment; Rescue impact constraint: The capacity of the road segment affected by the emergency rescue path is greatly reduced; γ2 represents the parameter that affects the capacity of the road segment; Road sections affected by the accident Non-negativity constraint: All traffic flow variables must be non-negative; S2. Generate traffic guidance plan Design a traffic induction scheme, where Z is the sum of the squares of the difference between the actual service traffic demand of all paths and the expected allocated traffic demand. The smaller Z is, the better the induction scheme is, and the allocated demand meets the actual service traffic; minimize the sum of the squares of the difference between the actual service traffic demand of all paths and the expected allocated traffic demand, that is, is the objective function, and Z is minimized. For a specific path, the traffic demand it actually serves consists of four parts, namely, the number of vehicles that choose path i among vehicles without vehicle-mounted induction terminals, the number of vehicles that comply with the induction plan and do not choose emergency rescue path i among vehicles with vehicle-mounted induction terminals, the number of vehicles that do not comply with the induction plan but do not choose emergency rescue path i among vehicles with vehicle-mounted induction terminals, and the number of vehicles that do not comply with the induction plan and choose emergency rescue path i among vehicles with vehicle-mounted induction terminals. GAMS software is used to solve the above problem and generate an induction plan for implementation. Objective function: Constraints: Where: Z represents the sum of squares of the difference between the actual service traffic demand of each path and the expected allocated traffic demand; i and j represent the path numbers; Ω represents the path set; It represents the number of vehicles that choose path i among the vehicles without vehicle guidance terminals installed; represents the number of vehicles installed with on-board guidance terminals that comply with the guidance plan and do not choose emergency rescue path i; It indicates the number of vehicles installed with vehicle-mounted guidance terminals that did not comply with the guidance plan but did not choose emergency rescue path i; The number of vehicles that are equipped with vehicle-mounted guidance terminals but do not comply with the guidance plan and choose emergency rescue path i; represents the expected distribution of traffic demand for path i; P represents the total traffic demand between the OD point pairs; θ represents the coverage rate of vehicle-mounted induced terminals; α i It represents the proportion of vehicles that choose route i in the total demand under the condition of no induced information; γ i Indicates the proportion of induced path i in all the induced information sent; η represents the induced compliance rate; β i It represents the proportion of the expected traffic demand of path i in the total demand; a represents the proportion of vehicles that choose emergency rescue route i k represents a route in the Ω path set.

Images