Road network level congestion charging method fusing Transform and model predictive control
By combining the collaborative optimization framework of Transformer and MPC, high-quality initial guesses are generated and refined searches are carried out, the real-time and efficient optimization of large-scale urban road network congestion charging strategies is solved, and computing efficiency improvement and control performance optimization is achieved.
Patent Information
- Application Number
- CN202510609566.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-13
- Publication Date
- 2025-08-26
AI Technical Summary
The existing technology is difficult to achieve real-time and efficient optimization of large-scale urban road network congestion charging strategies, traditional MPC methods are insufficient in computing timeliness, LSTM-enhanced MPC models have accumulated timing prediction errors, SLSQP algorithm solution time exceeds the real-time requirements, and the existing methods cannot effectively cope with the complex characteristics of multi-region coupling and dynamic propagation of congestion.
A collaborative optimization framework combined with Transformer and model prediction control (MPC) is adopted, and high-quality initial guess values are generated using Transformer, and MPC combined with the inner point method for refined searches, a lightweight MFD optimization model is built, and the paths are filtered through the Dijkstra algorithm is reduced to the number of MPC iterations, and real-time and efficient optimization is achieved.
Significantly improve computing efficiency by 40%, while ensuring real-time and excellent control performance, scientifically optimize the traffic efficiency of road networks and reduce the congestion in the central urban area.
Smart Images
Figure CN120544288A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of traffic congestion management, and specifically relates to a road network-level congestion charging method that integrates Transformer and model predictive control. Background Art
[0002] As urbanization continues to advance and transportation demand continues to grow, traffic congestion has become a common problem in cities. Large cities, in particular, frequently face regional congestion during peak commuting hours. Traditional microscopic traffic control methods are difficult to directly apply to large-scale road networks. Macroscopic congestion management strategies are needed to optimize the spatial distribution of traffic flow, alleviate congestion, and improve safety and operational efficiency. The introduction of the Macroscopic Fundamental Diagram (MFD) provides a direction for congestion management in road network-level transportation systems at a lower computational cost.
[0003] A well-defined MFD describes the unimodal, low-scatter relationship between the cumulative number of vehicles and the trip completion rate in a certain area of the road network. Congestion management methods based on MFD focus on macro-optimization at the road network level and can effectively cope with the complex dynamic characteristics of urban transportation systems at a low computational cost. As a typical demand management strategy, congestion charging regulates traffic demand entering the core area of the road network (usually the CBD) by levying additional tolls during peak hours, thereby alleviating congestion. The MFD theory provides a scientific basis for the road network-level congestion charging method: when the cumulative number of vehicles in a certain area exceeds the critical value of the cumulative number of vehicles in the MFD, an additional toll is levied on the area to induce travelers to choose alternative routes (i.e., detour to non-core areas), thereby optimizing the spatial distribution of traffic flow within the entire road network.
[0004] At present, there are congestion charging methods based on MFD in existing papers (such as Ye Li and Mohsen Ramezani, 2022. "Quasi revenue-neutral congestion pricing in cities: Crediting drivers to avoid city centers." Transportation Research Part C: Emerging Technologies, 145, 103932.), which adopts traditional model predictive control (MPC). MPC is an advanced automatic control method, also known as rolling horizon control, which processes dynamic systems through rolling optimization and feedback correction. Its core advantage is that it can simultaneously process complex systems with multiple input and output variables to achieve optimal control of the system. The present invention adopts an MPC optimizer based on the interior-point optimization algorithm, and dynamically generates the optimal charging scheme for each area through state prediction and feedback correction of the macro traffic flow model. In addition, large-scale urban transportation systems have strong time-varying characteristics, and the congestion charging values of each area need to be adjusted according to the real-time congestion status of the road network. In the process of solving toll collection in various areas of large-scale urban road networks, the optimization characteristics of MPC require resolving a finite time domain optimization problem within each control cycle. Due to its high computational complexity and long calculation time, existing research is difficult to meet the needs of real-time and efficient control.
[0005] Transformer, a deep learning architecture based on the self-attention mechanism, significantly improves the performance of natural language processing tasks through its advantages in parallel computing and long sequence modeling. This paper innovatively combines Transformer with MPC to construct a Transformer-MPC collaborative optimization framework. This framework leverages the powerful learning capabilities of Transformer to obtain an initial guess of the optimal congestion charging value for each region. This guess is then used as a "hot start" for the MPC solution, thereby reducing the number of MPC optimization iterations and enabling real-time and efficient optimization of congestion charging strategies for large-scale urban road networks.
[0006] Prior Art 1, Source of Prior Art 1: Ye Li and Mohsen Ramezani, 2022. “Quasirevenue-neutral congestion pricing in cities: Crediting drivers to avoid city centers.” Transportation Research Part C: Emerging Technologies, 145, 103932.
[0007] This article proposes a congestion pricing method that integrates MPC and long short-term memory (LSTM) neural networks. Its core innovation lies in using LSTM to improve MPC's optimization model: LSTM predicts the traffic flow transfer ratio between regions in real time and embeds this prediction into the MFD model (i.e., the MPC optimization model), thereby improving the optimization model's accuracy in capturing dynamic changes in the road network. The specific implementation method involves dividing the road network into multiple regions, each with a well-defined MFD. Using LSTM, the historical temporal dependencies of the traffic flow transfer ratio between regions are learned to estimate the inter-regional transfer flow ratio. The LSTM-predicted traffic flow ratio is integrated into the MFD model and fed into the MPC as an optimization input. The MPC then calculates the congestion pricing value for each region.
[0008] Disadvantages of Existing Technology 1: Disadvantage 1: For the real-time control needs of large-scale urban road networks, traditional MPC methods face significant computational timeliness challenges: their optimization solution time increases exponentially with the problem scale, making it difficult for the control strategy update frequency to keep pace with the dynamic evolution of large-scale road networks, severely restricting the real-time response of congestion management. This computational delay often results in MPC-based control solutions being based on outdated traffic status information, making it impossible to achieve timely response within the "perception-decision-control" closed loop.
[0009] Disadvantage 2: The optimization model input to MPC must ensure high accuracy. Its core principle is to perform rolling optimization based on the model's prediction of future road network conditions. Errors in the optimization model directly impact control performance. In the aforementioned study, although LSTM enhanced the predictive capabilities of the MPC optimization model, its time series prediction errors accumulate as the prediction horizon lengthens. Because MPC's rolling optimization mechanism relies on continuous multi-step prediction results for decision-making, this accumulated error can cause the optimization objective to gradually deviate from actual traffic conditions, ultimately impacting control effectiveness.
[0010] The technical solution of the second existing technology is CN119723912A.
[0011] The road network is partitioned; traffic parameters for each area are collected and the MFD curve for each area is fitted based on the collected data. An optimization model is established with the goal of minimizing the number of vehicles on the road network. An MPC controller with a built-in SLSQP algorithm is used to optimize and solve the optimal boundary control value, thereby controlling the boundary traffic flow and boundary traffic light timing of each area of the road network. Traffic flow is simulated using SUMO, and the MPC controller with a built-in SLSQP algorithm continuously iterates and optimizes the solution from the current time step to the next time step, achieving rolling optimization. The traffic state of each area is modeled using MFD to determine the regional flow dynamics. By establishing a network structure to simulate traffic dynamics, the MPC method is implemented to control the transfer of boundary traffic flow and dynamically control the timing of traffic lights, thereby alleviating traffic congestion.
[0012] Disadvantages of the second prior art
[0013] Disadvantage 1: While the SLSQP algorithm is used for optimization, MPC requires high-dimensional nonlinear optimization within each control cycle. For large-scale networks (with more than two regions), the number of variables and constraints increases exponentially, and the serial computational nature of SLSQP may cause the solution time to exceed real-time requirements.
[0014] Disadvantage 2: The control strategy of this patent is only applicable to a simplified dual-area road network model and cannot effectively cope with the complex characteristics of multi-area coupling and dynamic propagation of congestion in actual urban road networks. It has obvious scenario limitations.
[0015] Disadvantage 3: This patent takes minimizing the number of vehicles on the road network as its optimization goal, which has theoretical limitations. The MFD curve shows that only when the cumulative number of vehicles in the area approaches the critical value can the trip completion rate of the area reach its peak (i.e., the best traffic efficiency, see Figure 3 Simply minimizing the cumulative number of vehicles will lead to an “empty” road network and suboptimal control of MPC. Summary of the Invention
[0016] In order to solve the above technical problems, the present invention provides a road network-level congestion charging method that integrates Transformer and model predictive control, which solves the problem that it is difficult to use MPC to achieve real-time solution and dynamic control of congestion charging strategies for macro road networks.
[0017] The technical solution of the present invention is:
[0018] A road network-level congestion pricing method that integrates Transformer and model predictive control includes the following steps:
[0019] Step 1: Divide the metropolitan road network into regions, each with a well-defined macro-basic map and determine the corresponding critical cumulative vehicle number and maximum cumulative vehicle number based on the regional congestion charging method;
[0020] Step 2: Build a lightweight MFD optimization model for MPC with the goal of minimizing the total travel time of vehicles in the road network;
[0021] Step 3: Integrate the in-transit route selection model into the MFD optimization model, use the improved Dijkstra algorithm to select the path with the lowest travel cost for use in the optimization model, and use the interior point method model predictive controller to solve the optimal toll value for each area;
[0022] Step 4: Obtain historical toll data from the road network and build and train the Transformer model.
[0023] Step 5: Combine the Transformer model with the model predictive controller to build a Transformer-MPC collaborative optimization framework. This framework uses the high-quality initial guess toll value generated by the Transformer model to approximate the MPC optimal solution. This significantly reduces the number of MPC iterations and enables efficient optimization calculation of real-time congestion pricing in large-scale urban road networks.
[0024] Preferably, step 1 divides the road network of a large city into regions, where the divided regions include a central area and a peripheral area.
[0025] Preferably, the divided regions include a single peak, low scattering MFD.
[0026] Preferably, the MPC using the interior point method in step 3 includes the following sub-steps:
[0027] Step 3-1: Conduct system diagnosis based on the current congestion status observations in each area, and then use multi-step prediction to deduce the evolution trend of traffic flow in the road network to solve the optimal toll for each area;
[0028] Step 3-2: The charging mechanism continuously adjusts the traffic volume distribution between regions, and its iterative calculation process can effectively reduce the congestion level in the central urban area.
[0029] Preferably, step 4 is specifically:
[0030] Build and train a Transformer model. Through its multi-head attention mechanism, the model deeply explores the dependency between different toll rates in different areas and the cumulative number of vehicles. It outputs an initial toll guess value that is slightly different from the actual optimal toll value in each area under given traffic conditions, thereby reducing the number of subsequent MPC optimization iterations.
[0031] Preferably, step 5 includes the following sub-steps:
[0032] Step 5-1: Based on the pre-trained Transformer model, the system receives the current cumulative number of vehicles in each area and recent congestion charging data as input features in real time, and generates an initial guess of the optimal charging value for each area within milliseconds through parallel computing.
[0033] Step 5-2: Use the output value of the Transformer model as the initial guess value of the MPC charge MPC uses the interior point method to perform local refinement search and then output the optimal charging value u for each area i This method can significantly reduce the number of MPC iterations and improve computational efficiency;
[0034] Step 5-3: Observe the overall road network traffic status every 5 minutes and adjust the congestion charge value.
[0035] The beneficial effects of the network-level congestion pricing method that integrates Transformer and model predictive control are as follows:
[0036] 1. The present invention significantly improves computational efficiency; using a Transformer neural network to predict the initial guess value of charges in each area, the number of MPC optimization iterations is greatly reduced, and computational efficiency is increased by nearly 40%;
[0037] 2. The performance of the present invention is better; the MPC optimization algorithm based on the interior point method is particularly suitable for high-dimensional complex road network scenarios (such as the multi-region road network in the present invention), maintaining excellent control performance while ensuring real-time performance;
[0038] 3. The control objective of the present invention is more scientific; minimizing the total travel time of vehicles on the road network is the optimization goal, strictly following the MFD theory, and ensuring the optimization of the road network traffic efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 Flowchart of the present invention.
[0040] Figure 2 This is a regional map of the present invention.
[0041] Figure 3 This is the MFD diagram of the present invention.
[0042] Figure 4 This is the Transformer-MPC collaborative optimization framework diagram of the present invention. DETAILED DESCRIPTION
[0043] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0044] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.
[0045] A road network-level congestion pricing method that integrates Transformer and model predictive control includes the following steps:
[0046] Step 1: Divide the metropolitan road network into regions, each with a well-defined macro-basic map and determine the corresponding critical cumulative vehicle number and maximum cumulative vehicle number based on the regional congestion charging method;
[0047] Step 2: Build a lightweight MFD optimization model for MPC with the goal of minimizing the total travel time of vehicles in the road network;
[0048] Step 3: Integrate the in-transit route selection model into the MFD optimization model, use the improved Dijkstra algorithm to select the path with the lowest travel cost for use in the optimization model, and use the interior point method model predictive controller to solve the optimal toll value for each area;
[0049] Step 4: Obtain historical toll data from the road network and build and train the Transformer model.
[0050] Step 5: Combine the Transformer model with the model predictive controller to build a Transformer-MPC collaborative optimization framework. This framework uses the high-quality initial guess toll value generated by the Transformer model to approximate the MPC optimal solution. This significantly reduces the number of MPC iterations and enables efficient optimization calculation of real-time congestion pricing in large-scale urban road networks.
[0051] Step 1 of this implementation plan divides the road network of a large city into regions, and the divided regions include a central area and a peripheral area.
[0052] The divided regions of this embodiment include a single peak, low scattering MFD.
[0053] The MPC using the interior point method in step 3 of this embodiment includes the following sub-steps:
[0054] Step 3-1: Conduct system diagnosis based on the current congestion status observations in each area, and then use multi-step prediction to deduce the evolution trend of traffic flow in the road network to solve the optimal toll for each area;
[0055] Step 3-2: The charging mechanism continuously adjusts the traffic volume distribution between regions, and its iterative calculation process can effectively reduce the congestion level in the central urban area.
[0056] Step 4 of this embodiment is specifically as follows:
[0057] Build and train a Transformer model, and use a multi-head attention mechanism to deeply explore the dependency between different toll rates in each area and the cumulative number of vehicles. Output an initial toll guess value that is slightly different from the optimal toll value in each area under given traffic conditions, thereby reducing the number of subsequent MPC optimization iterations.
[0058] Step 5 of this embodiment includes the following sub-steps:
[0059] Step 5-1: Based on the pre-trained Transformer model, the system receives the current cumulative number of vehicles in each area and recent congestion charging data as input features in real time, and generates an initial guess of the optimal charging value for each area within milliseconds through parallel computing.
[0060] Step 5-2: Use the output value of the Transformer model as the initial guess value of the MPC charge MPC uses the interior point method to perform local refinement search and then output the optimal charging value u for each area i This method can significantly reduce the number of MPC iterations and improve computational efficiency;
[0061] Step 5-3: Observe the overall road network traffic status every 5 minutes and adjust the congestion charge value.
[0062] When this implementation plan is implemented,
[0063] The present invention proposes an innovative method for optimizing dynamic congestion charging for urban road networks. By combining the advantages of Transformer neural networks and MPC, it achieves real-time and efficient control of large-scale road networks. This method constructs an optimization model based on MFD theory and designs a congestion charging strategy based on MPC, with the control goal of minimizing the total travel time of the road network to alleviate traffic congestion. Its core innovation lies in: first, using a pre-trained Transformer model to predict the congestion charging value of each area based on the current state of the road network, providing high-quality initial guess values for MPC optimization; then using the interior point method of MPC to fine-tune the charging value to obtain the optimal charging value for each area. This Transformer-MPC collaborative optimization strategy fully utilizes the Transformer's ability to process large-scale long-sequence data, while retaining the theoretical advantages of MPC in constrained optimization, so that the congestion charging strategy of the macro road network can be solved in real time and the road network can be dynamically controlled.
[0064] The proposed method for optimizing dynamic congestion charging in urban road networks by integrating Transformer and model predictive control includes the following steps:
[0065] S1: Divide the road network of a large city into regions, with each region having a well-defined macro basic map and determining the corresponding critical cumulative number of vehicles and the maximum cumulative number of vehicles;
[0066] S2: Build an MFD-based optimization model with the goal of minimizing the total travel time of vehicles in the road network;
[0067] S3: Based on the current congestion level of each area, the model predictive controller using the interior point method is used to solve the optimal charging value for each area;
[0068] S4: Obtain historical toll data from the road network for multiple times to build and train the Transformer model;
[0069] S5: Combining the Transformer model with a model predictive controller (MPC) to construct a Transformer-MPC collaborative optimization framework. Using the high-quality initial guesses of toll charges generated by the Transformer model to approximate the MPC optimal solution, this framework significantly reduces the number of MPC iterations and enables efficient optimization of real-time congestion pricing in large-scale urban road networks.
[0070] Furthermore, the present invention divides the complex urban road network into multiple regional units with uniform congestion characteristics, such as Figure 2 As shown, the central region 7 is identified as the core region (ie, CBD), and the other regions 1 to 6 are peripheral regions. Each region has a well-defined single peak and low scattering MFD.
[0071] MFD reflects the deterministic relationship between the cumulative number of vehicles Ni(t) in a region and the road network trip completion rate Gi(Ni(t)), providing an objective quantitative indicator for the congestion level in each region.
[0072] like Figure 3 As shown in Figure 2, the horizontal axis of the MFD represents the cumulative number of vehicles in a certain area of the road network, and the vertical axis reflects the trip completion rate. When the trip completion rate reaches its maximum value, the corresponding horizontal axis value is the critical cumulative number of vehicles, Ncri, at which the road network operates at its optimal efficiency.
[0073] The single-peak, low-scattering MFD characteristics ensure that the relationship is highly reliable, avoiding the misjudgment of congestion levels due to MFD hysteresis (i.e., the macroscopic traffic flow exhibits different state evolution paths during the formation and dissipation of congestion).
[0074] In addition, after dividing into multiple areas, a region-based congestion charging method can be introduced to provide support for the method to judge the real-time congestion level, and achieve adaptive congestion charging optimization control based on real-time traffic status perception.
[0075] Furthermore, based on the MFD theoretical framework, the present invention constructs a lightweight optimization model for MPC. This model has two major features: (1) it only needs to track key macro variables such as the inflow and outflow of traffic in each area to simulate the evolution of the cumulative number of vehicles; (2) it embeds an in-transit path selection model to consider the impact of charging policies on travelers' path decisions. In addition, MPC optimizes the entire road network based on this model, with the goal of reducing the total travel time of all vehicles in the road network. Therefore, the established macro traffic flow model is as follows:
[0076]
[0077] Among them, N ii (t) represents the cumulative number of vehicles from the starting area i to the destination area i at time t, N ij (t) represents the cumulative number of vehicles from the starting area i to the destination area j at time t, so the cumulative number of vehicles in area i is in represents the set of road network division areas,
[0078] Q ii (t) represents the traffic demand from area i to destination area i at time t, Q ij (t) represents the traffic demand from area i to destination area j at time t. Similarly, Q i (t) = Q ii (t)+Q ij (t).
[0079] The outflow of vehicles in area i that complete the journey within the area at time t is expressed as M i i i (t) means,
[0080] The external outflow of vehicles in area i whose final destination is area j through the adjacent area h is expressed as M i h j (t) indicates.
[0081] The internal and external outflow of area i can be determined by the following formulas:
[0082]
[0083] Among them, ρ ih j (t) is the ratio of routes from region i to destination region j via region h, estimated by the Dijkstra algorithm considering real-time inter-region travel costs. ii (t) represents the cumulative number of vehicles from area i to destination area i at time t, N ij (t) represents the cumulative number of vehicles from area i to destination area j at time t. i (N i (t)) is the trip completion rate of area i, which is related to the number of vehicles N in the area. i (t). In addition, region i and region j are not the same region, and region h is the adjacent region of region i. is the set of adjacent regions of region i.
[0084] The present invention integrates the in-transit path selection behavior model - the k-shortest path algorithm based on real-time congestion status (i.e., the improved Dijkstra algorithm) into the optimization model, allowing travelers to adjust their driving routes in real time according to the dynamic changes in the road network congestion status.
[0085] In the path selection behavior model, the improved Dijkstra algorithm selects routes based on travel costs. The travel cost of a path between any two areas is the sum of the travel time and the congestion costs of all areas passed through during the transit period.
[0086] Combining the MFD-based optimization model and the congestion charging mechanism, the objective function J of the present invention is to minimize the total travel time of all vehicles in the road network, which is determined by the following formula:
[0087]
[0088] subject to,
[0089]
[0090] The above formula realizes the conversion from the continuous time t model (Equations (1)-(4)) to the discrete time step k form (Equations (6)-(7)). Both forms represent the macroscopic flow conservation relationship of region i. p and K c are the prediction and control time domains of MPC, T c is the sampling time, and T is the time step duration predicted by the model. i jam is the maximum cumulative number of vehicles that area i can accommodate. Define u i min 、u i maxare the upper and lower limit constraints of congestion charging in area i, which are used to avoid drastic fluctuations in charging values between adjacent control time steps and excessive charging.
[0091] Furthermore, the core of the real-time congestion charging proposed by the present invention is to regulate the traffic flow entering the congested area to improve the efficiency of road network operation by levying dynamic tolls on vehicles crossing the boundaries of the congested area. The optimal charging scheme for each area is calculated based on the MPC framework, which improves the traffic efficiency of the central urban area while maintaining the stability of the traffic flow in the surrounding areas. The MPC control system of the present invention adopts the interior-point method as the core optimization algorithm, and realizes dynamic congestion management through the rolling time domain control framework. Its working process is as follows: first, based on the current congestion status observation of each area, a system diagnosis is performed, and then the evolution trend of traffic flow in the road network is deduced through multi-step prediction, and finally the optimal toll for each area is solved. This charging mechanism continuously adjusts the traffic volume distribution between regions, and its iterative calculation process can effectively reduce the congestion level in the central urban area.
[0092] The iterative calculation process of MPC is a closed-loop control with rolling optimization and feedback correction. Its core lies in generating control actions by solving the optimization target J online, i.e., the real-time charging value u of each area in this patent. i Furthermore, according to the current road network measurement, the cumulative number of vehicles N in each area ii (k) and N ij (k), solve the charging value u of each area through the optimization model i (k), at this time the MPC system evolves to time k+1, and then repeats the calculation of the above steps, and calculates u of the previous time step according to the calculation of the subsequent time step. i (k) Optimize.
[0093] Furthermore, the present invention constructs and trains a Transformer model based on the time-varying cumulative number of vehicles during multiple historical morning rush hours and the optimal charging data for each area.
[0094] The Transformer model is trained to deeply explore the dependency between different toll rates in each area and the cumulative number of vehicles through a multi-head attention mechanism. Ultimately, the model outputs an initial guess value that is slightly different from the actual optimal toll value in each area under given traffic conditions, thereby reducing the number of subsequent MPC optimization iterations.
[0095] Furthermore, the present invention builds a Transformer-MPC collaborative optimization framework, achieving a breakthrough in computational efficiency through the organic combination of deep learning methods and traditional optimization control. The specific implementation includes three core technical links:
[0096] ① Initial guess solution generation: Based on the pre-trained Transformer model, the system receives the current cumulative number of vehicles in each area and recent congestion charging data as input features in real time, and generates the initial guess value of the optimal charging value in each area within milliseconds through parallel computing.
[0097] ②MPC hot start optimization: Transformer output value (i.e. the initial guess value of the optimal charge) is used as the initial solution of the MPC optimization input, and then the interior point method is used to perform local refined search, thereby significantly reducing the number of MPC iterations. The core of MPC lies in its rolling time domain control mechanism, which enables it to have dynamic adjustment capabilities. Specifically, at each sampling moment (i.e., every time interval), MPC needs to perform: 1) based on the current road network status and optimization model, deduce the dynamic evolution of traffic flow in various regions within the future limited time domain; 2) with the goal of minimizing the total travel time of vehicles on the road network, solve the optimal congestion charging sequence in the time domain; 3) only apply the optimal charge value at the current moment, and re-roll the optimization in the next sampling period based on the actual congestion status of the road network.
[0098] ③ Dynamic rolling execution: Congestion charging is adjusted for the entire road network every 5 minutes as a sampling cycle, such as Figure 4 As shown in the figure, the entire process of congestion status collection in each area of the road network → Transformer inference → MPC optimization → policy issuance is completed in each cycle, while maintaining real-time performance and compressing the calculation time of large-scale road network charging.
Claims
1. A road network-level congestion pricing method that integrates Transformer and model predictive control, characterized by: The following steps are involved: Step 1: Divide the metropolitan road network into regions, each with a well-defined macro-basic map and determine the corresponding critical cumulative vehicle number and maximum cumulative vehicle number based on the regional congestion charging method; Step 2: Build a lightweight MFD optimization model for MPC with the goal of minimizing the total travel time of vehicles in the road network; Step 3: Integrate the route selection behavior model into the MPC lightweight MFD optimization model and input it into the model predictive controller using the interior point method to solve the optimal toll value for each area; Step 4: Obtain historical toll data from the road network and build and train the Transformer model. Step 5: Combine the Transformer model with the model predictive controller to build a Transformer-MPC collaborative optimization framework. This framework uses the high-quality initial guess toll value generated by the Transformer model to approximate the MPC optimal solution. This significantly reduces the number of MPC iterations and enables efficient optimization calculation of real-time congestion pricing in large-scale urban road networks.
2. The road network level congestion charging method integrating Transformer and model predictive control according to claim 1 is characterized in that: The step 1 divides the road network of the large city into regions, where the divided regions include a central area and a peripheral area.
3. The road network-level congestion charging method integrating Transformer and model predictive control according to claim 2 is characterized in that: The partitioned region includes a single-peak, low-scattering MFD.
4. The road network-level congestion pricing method integrating Transformer and model predictive control according to claim 1 is characterized in that: The interior point method in step 3 includes the following sub-steps: Step 3-1: Perform system diagnosis based on the current congestion status observations in each area, and then use multi-step prediction to deduce the evolution trend of traffic flow in the road network to solve the optimal toll for each area; Step 3-2: The charging mechanism continuously adjusts the traffic volume distribution between regions, and its iterative calculation process can effectively reduce the congestion level in the central urban area.
5. The road network level congestion pricing method integrating Transformer and model predictive control according to claim 1 is characterized in that: The step 4 is specifically as follows: Through the multi-head attention mechanism, the Transformer model is trained to deeply explore the mapping relationship between different rates in each area and the changes in the cumulative number of vehicles, and output an initial guess value with a small difference from the optimal charging value in each area under given traffic conditions, so as to reduce the number of subsequent MPC optimization iterations.
6. The road network level congestion charging method integrating Transformer and model predictive control according to claim 1 is characterized in that: The step 5 includes the following sub-steps: Step 5-1: Based on the pre-trained Transformer model, the system receives the current cumulative number of vehicles in each area and recent congestion charging data as input features in real time, and generates an initial guess of the optimal charging value for each area within milliseconds through parallel computing. Step 5-2: Initial guess of the optimal charge predicted by the Transformer model As the initial solution of MPC optimization (this solution is close to the true optimal solution u i ), and then use the interior point method to solve the optimal charging value u of each area in real time i , thus significantly reducing the number of MPC iterations; Step 5-3: Perform congestion charging adjustments on the entire road network every 5 minutes as a sampling period.