A dynamic prediction method for road network traffic distribution fusing data driving and traffic distribution

Abstract

The application discloses a kind of fusion data-driven and traffic distribution road network traffic distribution dynamic prediction method, comprising: according to the historical traffic flow data of road section of road network and the traffic cell of pre-classification, the road section flow data set of each historical statistical period is constructed;According to the road section flow data set of each historical statistical period, respectively, the origin-destination matrix estimation is carried out, and the historical origin-destination matrix corresponding to each historical statistical period is obtained;According to the time series of each historical origin-destination matrix, the road network origin-destination matrix of target period is predicted by time series prediction model;According to the road network origin-destination matrix of target period, the predicted traffic is distributed to each road section of road network by user equilibrium traffic distribution method, and the road network traffic distribution prediction result of target period is obtained.The application can be used in urban road traffic management, dynamic path induction and network vehicle prospective path planning etc.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation, and in particular relates to a method for dynamic prediction of road network traffic distribution that integrates data-driven and traffic assignment. Background Technology

[0002] In recent years, with the continuous deepening of urbanization and the increasing number of motor vehicles in my country, urban road traffic congestion has become a prominent problem restricting the high-quality development of megacities. Driven by the national strategy of building a strong transportation network and the construction of smart cities, utilizing massive amounts of road segment traffic flow data collected by roadside radar, video, and other sensing devices, combined with data-driven methods and traffic engineering models to improve the ability to perceive the operational status of road networks, has become an important development direction for intelligent traffic management. Currently, cities such as Beijing are accelerating the construction of intelligent transportation infrastructure and vehicle-road cooperative systems, placing higher demands on short-cycle, interpretable, and easily deployable road network traffic flow distribution prediction. Road network traffic flow distribution prediction is a crucial foundation for dynamic path guidance, signal coordination and control, traffic organization optimization, and forward-looking path planning for connected vehicles. Existing research methods are mainly divided into two categories: one is end-to-end traffic flow prediction methods based on deep learning, which directly use historical road segment traffic flow data to predict future traffic flow; the other is traffic assignment methods based on the traditional four-stage method, which obtains the road network traffic flow distribution based on OD demand input through an assignment model. Both types of methods have certain value in their respective applicable scenarios.

[0003] However, existing technologies still have the following shortcomings in practical applications. First, although deep learning-based end-to-end traffic forecasting methods perform well in terms of prediction accuracy, they are essentially "black box" models, making it difficult to reveal the intrinsic mechanism between the evolution of travel demand and changes in road network traffic distribution. This results in insufficient interpretability and an inability to be organically integrated with classic traffic engineering theories such as maximum entropy estimation and user equilibrium allocation, thus limiting the credibility and interpretability of the prediction results in traffic management decisions. Second, traffic assignment in the traditional four-stage method typically relies on static or quasi-static OD demand as input, making it difficult to effectively reflect the dynamic evolution characteristics of travel demand under scenarios such as peak-hour transitions and the impact of unforeseen events. This leads to significant deficiencies in the timeliness and dynamic adaptability of road network traffic forecasting for the next time period, making it difficult to respond promptly to real-time changes in urban traffic. Third, with the rapid development of vehicle-road cooperative and intelligent connected technologies, traffic management systems not only need to predict the overall road network status in future periods, but also need to provide forward-looking route suggestions for connected vehicles and provide management departments with global guidance schemes. However, the existing methods are not yet perfect in the integrated processing chain of "historical traffic collection - OD demand back-calculation - OD demand prediction - balanced traffic allocation". There is a lack of a systematic technical solution that organically combines data-driven methods with traffic engineering allocation models, which restricts the improvement of the ability of intelligent transportation systems in dynamic decision support. Summary of the Invention

[0004] This invention aims to address the technical problems of insufficient interpretability and difficulty in reflecting dynamic changes in travel demand in existing road network traffic forecasting methods. It proposes a dynamic road network OD demand and traffic distribution forecasting method that integrates data-driven approaches and traffic assignment. Based on estimating and predicting OD travel demand using historical road network traffic data, this method performs traffic assignment on road network traffic, enhancing the interpretability of road network traffic changes and providing strong support for dynamic route guidance in intelligent transportation systems and traffic management in the context of intelligent connected vehicles.

[0005] To achieve the above objectives, this invention provides a method for dynamic prediction of road network traffic distribution that integrates data-driven and traffic assignment approaches, comprising: Based on historical traffic flow data collected by roadside detectors on each road segment of the road network, and pre-divided traffic zones, a traffic flow dataset for each historical statistical period is constructed. Based on the traffic flow dataset of the road segment, OD matrix estimation is performed to obtain the historical OD matrix corresponding to each historical statistical period. Based on the time series formed by the historical OD matrix, the road network OD matrix for the target time period is predicted using a time series prediction model; Based on the road network OD matrix for the target time period, the predicted traffic volume is allocated to each road segment of the road network using a user-balanced traffic allocation method, thereby obtaining the predicted road network traffic distribution for the target time period.

[0006] Preferably, the process of obtaining the historical OD matrix corresponding to each historical statistical period includes: For each historical statistical period, the OD matrix is ​​back-derived using the road segment detection flow as a constraint and the maximum entropy OD estimation model. User-balanced traffic allocation is then performed on the back-derived OD matrix. The OD matrix is ​​iteratively corrected based on the relative error between the allocated road segment flow and the detection flow until the preset accuracy requirement is met. The OD matrix finally back-derived for the corresponding period is then used as the historical OD matrix for that period.

[0007] Preferably, the process of performing user-balanced traffic allocation on the back-derived OD matrix, and iteratively correcting the OD matrix based on the relative error between the allocated road segment traffic and the detected traffic until the preset accuracy requirements are met includes: The outer layer iterative method is used to back-engineer the OD matrix. In each round of outer layer iteration, the OD matrix obtained by back-engineering is used as the demand input. The inner layer iterative method is used to perform user balanced traffic allocation to obtain the balanced traffic allocation of each road segment. The relative error between the balanced distribution of traffic on each road segment and the detected traffic on each road segment is used as the outer convergence criterion. If the preset convergence threshold is not reached, the prior OD quantity is updated based on the OD matrix obtained by the current back-inference, and the OD matrix back-inference and inner user balanced distribution are re-executed until the relative error of all road segments meets the preset accuracy requirements.

[0008] Preferably, the process of performing user-balanced traffic allocation in an inner-layer iterative manner includes: Initialize the flow rate of each road segment to zero. Based on the current flow rate of each road segment, update the impedance of each road segment using the impedance function that changes with the flow rate over the travel time of the road segment. Using the current impedance of each road segment as the weight, perform shortest path search for each OD pair, allocate all travel volume of each OD pair to the current shortest path, and accumulate to obtain the auxiliary flow of each road segment; Determine the optimal step size that minimizes the total travel cost of the system, and update the traffic flow of each road segment according to the optimal step size; Calculate the relative gap index used to measure the deviation of the current traffic allocation from the user balance state. If the relative gap index is less than or equal to the preset inner layer convergence threshold, stop the inner layer iteration and output the current traffic of each road segment as the balanced allocation traffic; otherwise, return to the update impedance step to continue the iteration.

[0009] Preferably, the impedance function of the road segment travel time as a function of flow rate is adopted as a function relationship in which the travel time increases monotonically as the flow rate increases. The function relationship is determined by the road segment free-flow travel time, road segment flow rate, road segment capacity, and preset calibration parameters.

[0010] Preferably, the maximum entropy OD estimation model takes the detected traffic flow of the road segment as a constraint, and on the set of all feasible OD matrices that satisfy the traffic flow constraint, it takes the maximization of entropy as the objective function and solves the travel volume of each OD by using the Lagrange multiplier method.

[0011] Preferably, in the maximum entropy OD estimation model, the travel volume of each OD is determined by the prior OD quantity, the Lagrange multiplier corresponding to each road segment, and the diversion coefficient of each OD pair on each road segment. Specifically, the travel volume of each OD is equal to the prior OD quantity multiplied by the exponential form of a negative exponential function of the product of the Lagrange multiplier and the diversion coefficient.

[0012] Preferably, the time-series prediction model is a long short-term memory network model, which takes the OD matrix of a number of consecutive historical statistical periods as the input sequence, and outputs the road network OD matrix of the target period after encoding by the long short-term memory network. The Long Short-Term Memory Network model learns the temporal evolution of the OD matrix during the offline training phase, and completes single-step prediction in the online prediction phase using the most recent preset number of historical OD matrices as input.

[0013] Preferably, the long short-term memory network model updates its state at each time step through a forget gate, an input gate, and an output gate. The forget gate determines the retention ratio of the cell state at the previous time step, the input gate controls the degree to which the current candidate state is written into the cell state, the cell state is updated according to the outputs of the forget gate and the input gate, and the output gate adjusts the output of the current hidden state.

[0014] Preferably, the user-balanced traffic allocation method is as follows: Using the travel volume of each OD in the road network OD matrix during the target time period as the allocation demand, a successive average iterative algorithm is used to solve the mathematical programming model with user balance as the objective. In each iteration, an all-in-one auxiliary allocation is first performed to determine the descent direction, and then a one-dimensional search is used to determine the optimal step size and update the road segment flow. The iteration stops when the relative gap index is less than or equal to the preset convergence threshold, and the balanced flow of each road segment is output as the road network flow distribution prediction result for the target time period.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: First, this invention constructs a complete processing chain from historical road segment traffic to historical OD matrix, then to predicted OD matrix, and finally to road network traffic distribution through user equivalence allocation. It combines the maximum entropy estimation model with the user equivalence allocation principle, so that the traffic prediction results can clearly reflect the impact mechanism of changes in travel demand on road network traffic distribution. This overcomes the problem of insufficient interpretability of existing end-to-end deep learning methods and significantly improves the credibility of prediction results in traffic management decisions.

[0016] Secondly, by estimating the OD matrix for each historical statistical period and then using a time-series prediction model to capture the temporal evolution of OD demand, this invention can effectively address the dynamic changes in travel demand under scenarios such as peak-hour switching and the impact of sudden events. It overcomes the shortcomings of the traditional four-stage method, which relies on static OD input and suffers from insufficient dynamic adaptability, and improves the timeliness and accuracy of road network traffic prediction for the next period.

[0017] Third, this invention organically combines data-driven OD prediction with traffic assignment based on the principle of user equilibrium, forming an integrated technical solution of "historical traffic flow - OD back-calculation - OD prediction - equilibrium assignment". It can directly serve forward-looking path planning and global traffic guidance in the vehicle-road cooperative environment, and provides a complete and engineering-deployable decision support framework for intelligent traffic management systems, filling the gap in the integration of the overall processing chain in existing methods.

[0018] Furthermore, this invention introduces an inner and outer nested iterative correction mechanism in the historical OD estimation stage, which effectively ensures the reliability of the OD back-calculation results and makes full use of the historical traffic flow data collected by existing roadside detectors. It does not require additional deployment of dedicated survey equipment, making it easy to promote and apply in actual traffic management systems. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the iterative solution process for historical OD estimation of the road network according to an embodiment of the present invention. Figure 3 This is a flowchart of the Frank-Wolfe algorithm for user-balanced traffic allocation in an embodiment of the present invention. Figure 4 This is a schematic diagram of the LSTM road network OD prediction model according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the topology of the Sioux-Falls transportation network according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] like Figure 1 As shown, this embodiment provides a method for dynamic prediction of road network traffic distribution that integrates data-driven and traffic assignment approaches, including: Based on historical traffic flow data collected by roadside detectors on each road segment of the road network, and pre-divided traffic zones, a traffic flow dataset for each historical statistical period is constructed. Based on the road segment traffic data set, OD matrix estimation is performed to obtain the historical OD matrix corresponding to each historical statistical period; Based on the time series formed by the historical OD matrix, the road network OD matrix for the target time period is predicted using a time series prediction model; Based on the road network OD matrix for the target time period, the predicted traffic volume is allocated to each road segment of the road network using a user-balanced traffic assignment method, thereby obtaining the predicted road network traffic distribution results for the target time period.

[0023] Furthermore, the data collection and traffic zone division involved in this embodiment include: (1) Road network topology transformation and traffic zone division; Define the road network boundaries within the research scope and perform a topological transformation on the road network: define intersections and dead-end points in the road network as basic point units, and assign a unique number to each point unit. A directed connection segment between point units is defined as a directed edge.

[0024] The above-mentioned basic point units are used as traffic zones. Traffic zones are a well-known concept in traffic engineering, referring to several traffic sub-zones divided into traffic networks according to certain principles. Each traffic zone can serve as a starting point for a trip. (Origin) can also be used as the destination of a trip. (Destination), which corresponds to the OD matrix and ,in and All are traffic zone numbers. A single trip has exactly two endpoints (origin and destination), corresponding to a pair of origin-destination (OD) pairs, and there is at least one connecting path between the OD pairs.

[0025] Since all nodes in the road network are considered traffic zones, both ends of a road segment are traffic zones. Therefore, each road segment can be represented by the traffic zone number at its starting point and the traffic zone number at its ending point in the direction of travel. - (For example, "1-2" indicates a directed road segment from traffic zone 1 to traffic zone 2).

[0026] (2) Collection of historical traffic flow data; Extensive roadside sensing equipment, including radar and video sensors, has detected and recorded a wealth of traffic flow data. Based on detectors deployed across various road segments within the road network, historical traffic flow data is collected for each segment and aggregated statistically at specific time intervals (e.g., 5 minutes, 10 minutes, 15 minutes, or 30 minutes, the specific intervals to be determined based on sensor performance, the scale of the road network within the study area, and operational requirements). During the period The traffic is recorded as This will create a historical traffic volume database for road sections within the research area.

[0027] The data format of the historical traffic volume database includes: road segment start number, road segment end number, road segment length, and road segment traffic flow in each statistical period.

[0028] (3) Selection of historical flow data for prediction; For the target time period of the proposed traffic forecast Extracting data from the historical traffic volume database for the forecast period ahead. Historical traffic data for each time period is used as the basis for subsequent OD estimation and traffic forecasting.

[0029] Furthermore, the process of obtaining the historical OD matrix corresponding to each historical statistical period includes: For each historical statistical period, the OD matrix is ​​back-derived using the road segment detection flow as a constraint and the maximum entropy OD estimation model. User-balanced traffic allocation is then performed on the back-derived OD matrix. The OD matrix is ​​iteratively corrected based on the relative error between the allocated road segment flow and the detection flow until the preset accuracy requirement is met. The OD matrix finally back-derived for the corresponding period is then used as the historical OD matrix for that period.

[0030] Furthermore, the process of performing user-balanced traffic allocation on the back-derived OD matrix, and iteratively correcting the OD matrix based on the relative error between the allocated road segment traffic and the detected traffic until the preset accuracy requirements are met includes: The outer layer iterative method is used to back-engineer the OD matrix. In each round of outer layer iteration, the OD matrix obtained by back-engineering is used as the demand input. The inner layer iterative method is used to perform user balanced traffic allocation to obtain the balanced traffic allocation of each road segment. The relative error between the balanced distribution of traffic on each road segment and the detected traffic on each road segment is used as the outer convergence criterion. If the preset convergence threshold is not reached, the prior OD quantity is updated based on the OD matrix obtained by the current back-inference, and the OD matrix back-inference and inner user balanced distribution are re-executed until the relative error of all road segments meets the preset accuracy requirements.

[0031] Furthermore, the process of performing user-balanced traffic allocation in an inner-layer iterative manner includes: Initialize the flow rate of each road segment to zero. Based on the current flow rate of each road segment, update the impedance of each road segment using the impedance function that changes with the flow rate over the travel time of the road segment. Using the current impedance of each road segment as the weight, perform shortest path search for each OD pair, allocate all travel volume of each OD pair to the current shortest path, and accumulate to obtain the auxiliary flow of each road segment; Determine the optimal step size that minimizes the total travel cost of the system, and update the traffic flow of each road segment based on the optimal step size; Calculate the relative gap index used to measure the deviation of the current traffic allocation from the user balance state. If the relative gap index is less than or equal to the preset inner layer convergence threshold, stop the inner layer iteration and output the current traffic of each segment as the balanced allocation traffic; otherwise, return to the update impedance step to continue the iteration.

[0032] Furthermore, the impedance function for the change of travel time with flow rate adopts a function relationship in which travel time increases monotonically as flow rate increases. The function relationship is determined by the free-flow travel time of the road segment, the flow rate of the road segment, the traffic capacity of the road segment, and preset calibration parameters.

[0033] Furthermore, the maximum entropy OD estimation model takes the detected traffic flow of road segments as a constraint. On the set of all feasible OD matrices that satisfy the traffic flow constraint of road segments, it takes the maximization of entropy as the objective function and solves the traffic volume of each OD by using the Lagrange multiplier method.

[0034] Furthermore, this embodiment includes the following steps for historical OD estimation: (1) Selection of OD matrix estimation model; Among various OD matrix estimation models, this embodiment selects the maximum entropy model. This model assumes that vehicle travel behavior is random, and each OD distribution matrix corresponds to an existence probability. The OD distribution matrix that is closest to the actual situation is the matrix with the highest existence probability.

[0035] Road sections within the study area The relationship between the detected traffic volume and the origin-destination (OD) travel volume can be expressed as: in, Road sections detected by the detector Traffic on the internet; For traffic community To the traffic community Travel volume; For OD On the road section The diversion coefficient indicates the route through which the origin-destination (OD) traffic passes. The allocation weights; This refers to the set of road segments within the research area.

[0036] Assume the trip volume of each OD Let entropy be a random variable, and let entropy be the set of all feasible OD matrices that satisfy the road segment flow constraints. When the maximum entropy is satisfied, the maximum entropy principle is followed, and the maximum entropy model is constructed as follows: in, This is the set of OD matrices for all road segments that satisfy the traffic flow constraints. OD matrix The probability of existence; the constraint condition indicates that the sum of the allocated flow rates for each road segment equals the detected flow rate.

[0037] Using the Lagrange multiplier method and the Stirling approximation method, the expression for OD trip volume is obtained as follows: in, For the prior OD quantity, an initial constant can be used instead; For the road section The corresponding Lagrange multipliers.

[0038] (2) Iterative solution and accuracy verification; Solving the above model yields an OD matrix derived from road segment traffic flow. However, due to prior OD values... There is uncertainty, and the diversion coefficient OD travel volume The interdependence necessitates nested iterative solutions and corrections to the estimation results, with the iterative solution process as follows: Figure 2 As shown.

[0039] The All-or-Nothing (AON) allocation method assigns values ​​to each OD pair for a given segment impedance. Search for the shortest path and collect all trips for that OD pair. Assign the route to the shortest path; unselected paths are assigned zero points. This method does not consider the feedback effect of road segment traffic flow on travel time, is computationally simple, and can be used for initial path determination or auxiliary traffic flow calculation in iterative traffic assignment algorithms.

[0040] (1) Diversion coefficient Initialization. In the first outer iteration ( When the free-flow travel time of each road segment is used, For impedance, the all-or-nothing allocation method described above is used to determine the shortest path for each OD pair, and the initial splitting coefficient is calculated accordingly: if the road segment Located at OD pair On the shortest path, then ,otherwise In subsequent outer iterations ( When the user balance allocation result of the previous round is used, the balanced flow is calculated in reverse according to formula (8). .

[0041] (2) Inner-layer user equilibrium allocation. For a given OD matrix, the User Equilibrium (UE) traffic assignment method is used to allocate traffic to the road network. User equilibrium allocation is based on Wardrop's first principle: in equilibrium, the same OD pair... All paths used in between have the same and minimum path travel impedance. (i.e., the impedance of each segment along the path) The sum of all available paths, and there are no unused paths with lower path impedance. Unlike the all-or-nothing allocation method, user-balanced allocation considers the feedback effect of segment traffic on travel time; the diversion coefficient... The result of the balanced distribution is determined according to equation (8), that is... ,satisfy .

[0042] The balanced allocation of inner-layer users is solved iteratively using the Frank-Wolfe algorithm (successive averaging method), and the algorithm flow is as follows: Figure 3 As shown, the main steps are as follows: Step 1: Initialization. Set the inner iteration count. Initialize the flow of each road segment ; Step 2: Update segment impedance. Based on the current segment flow rate. Calculate the travel time for each road segment according to formula (6). ; Step 3: All-or-nothing auxiliary allocation. Following the all-or-nothing allocation method described above, ... For the impedance of the road segment, for each OD pair Perform shortest path search to find all O(d) values. Assign to the current shortest path to obtain auxiliary flow. ; Step 4: Step Size Search. Determine the optimal step size. This causes the Beckmann objective function to decrease, and the bisection method or the golden section method are commonly used to solve it. Step 5: Traffic Update. Update the traffic flow of the road segment using the convex combination: Step 6: Inner layer convergence determination. Calculate the relative gap according to equation (7). ,like Then stop the inner iteration; otherwise, let Return to Step 2.

[0043] The road segment impedance function uses the Bureau of Public Roads (BPR) function to characterize the impact of traffic flow on travel time: in, For road section Free-flow travel time, For the length of the road segment, Free flow velocity ( and A consistent length-time unit system must be used. For road section Traffic capacity; , For calibration parameters, the values ​​are usually taken as follows: , .

[0044] The convergence of the balanced allocation of inner-layer users is determined using the relative gap index: in, This represents the total travel time of the system. For OD Shortest path travel time under current impedance; The inner layer equilibrium convergence threshold is usually set to [value missing]. In step two, during inner layer allocation, the formula is as follows: For the current estimated OD amount; in step four, when assigning the predicted OD matrix, Pick .

[0045] After the inner iteration converges, the balanced flow of each road segment is determined. Inverse calculation of OD splitting coefficient: in, For OD On the road section Distribute traffic on top.

[0046] (3) Verification of the accuracy of outer layer OD estimation. After the inner layer is evenly distributed, let The following formula is used to verify whether the relative error between the balanced distribution flow rate and the detector-detected flow rate meets the preset accuracy requirements: in, For the outermost layer In the next iteration, the inner layer evenly distributes the converged road segments. flow; To estimate the convergence threshold for the outer OD, a threshold is typically set to... .

[0047] If the preset accuracy requirement is not met, then the number of outer iterations is reduced. Using the OD matrix obtained from the current solution Replace the prior OD quantity in the model Updated with the results of the previous round of balanced distribution. Repeat the OD matrix solution and inner traffic assignment until equation (9) is satisfied.

[0048] (3) Solving for the historical OD matrix across multiple time periods; For the traffic flow data of each road segment for each historical statistical period obtained from the summary statistics, the above-mentioned OD estimation model is established and iteratively solved to obtain the OD distribution matrix corresponding to each historical period. A total of [number missing] results are obtained. A historical OD distribution matrix, denoted as This serves as the input for subsequent modeling.

[0049] Furthermore, in the maximum entropy OD estimation model, the travel volume of each OD is determined by the prior OD quantity, the Lagrange multiplier corresponding to each road segment, and the diversion coefficient of each OD pair on each road segment. Specifically, the travel volume of each OD is equal to the prior OD quantity multiplied by the exponential form of the negative exponential function of the product of the Lagrange multiplier and the diversion coefficient.

[0050] Furthermore, the time-series prediction model is a long short-term memory network model. The long short-term memory network model takes the OD matrix of a number of consecutive historical statistical periods as the input sequence, and outputs the road network OD matrix of the target period after encoding by the long short-term memory network. The Long Short-Term Memory (LSTM) network model learns the temporal evolution of the OD matrix during the offline training phase, and completes single-step prediction in the online prediction phase using the most recent preset number of historical OD matrices as input.

[0051] Furthermore, the Long Short-Term Memory (LSTM) network model updates its state at each time step through a forget gate, an input gate, and an output gate. The forget gate determines the retention ratio of the cell state from the previous time step, the input gate controls the degree to which the current candidate state is written into the cell state, the cell state is updated based on the outputs of the forget gate and the input gate, and the output gate adjusts the output of the current hidden state.

[0052] Furthermore, this embodiment specifically includes the following for constructing the LSTM prediction model: Based on the obtained historical OD matrix time series, a Long Short-Term Memory (LSTM) prediction model is constructed to learn the temporal evolution pattern of travel demand and output the road network OD distribution matrix for the target time period.

[0053] (1) Overall Model Structure. LSTM is a recurrent neural network with a gating mechanism, which can selectively retain or forget historical information, thereby effectively alleviating the gradient vanishing and gradient explosion problems of traditional RNNs in long sequence training, and is suitable for characterizing long-term dependencies in OD (Original Dependency) requirements. For example... Figure 4 As shown, the model is continuous The OD matrix of historical time periods is used as the time-series input, and after encoding by an LSTM network, the OD matrix of the target prediction time period is output. The whole process is divided into two stages: offline training and online prediction. The former uses long-term historical data to learn the evolution pattern of OD, while the latter uses the most recent historical data. A historical OD matrix is ​​used as input to complete a single-step prediction.

[0054] (2) Gating mechanism and state update. The core of the LSTM cell unit lies in three gating structures—the forgetting gate, the input gate, and the output gate—as well as the cell state that carries long-term memory. and candidate states .set up This is the hidden state from the previous moment. Given the current input (the OD vector corresponding to a certain historical time period in this embodiment), the formulas for updating each gate variable and the state are as follows: In the formula, , , , The weight matrix to be trained. , , , For bias terms; For the Sigmoid function, the output is... The gate coefficient within the interval; It is the hyperbolic tangent function; Represents the Hadamard product (element-by-element multiplication). Forget gate. Determines the cell state at the previous moment Retention ratio, input gate Control candidate state Write level, output gate Adjust the current hidden state Its external output.

[0055] (3) Offline model training. In the offline stage, a longer period of time (such as the past) is used in the historical traffic volume database. sky, Traffic flow on road segments is calculated, and the OD matrix is ​​estimated for each time period to construct an OD time series training sample set. Input-output sample pairs are constructed using a sliding window approach: each sample is displayed in a continuous window. The model takes the OD matrix of the previous historical time period as input and the OD matrix of the next time period as output. The mean squared error (MSE) is used as the loss function, and the network parameters are updated using optimization algorithms such as Adam until the validation set loss converges. The model is then saved for use in the online prediction stage.

[0056] (4) Online prediction configuration. During online prediction, the input sequence length is set to... The number of historical time periods selected is consistent with the number of historical traffic data used for prediction, and the output is an OD matrix for one target prediction time period. If it is necessary to cover a longer future time span, the prediction can be rolled in segments according to the time statistical interval; to ensure prediction accuracy, the number of time periods output in a single continuous output should not be too large.

[0057] (2) Generation of the OD matrix for prediction period; The result Historical OD distribution matrix As a time-series input, it is applied to the trained LSTM model to obtain the target time period to be predicted. Road network OD distribution matrix .

[0058] Furthermore, the user-balanced traffic allocation method is as follows: Using the travel volume of each OD in the road network OD matrix during the target time period as the allocation demand, a successive average iterative algorithm is used to solve the mathematical programming model with user balance as the objective. In each iteration, an all-in-one or all-out auxiliary allocation is first performed to determine the descent direction, and then a one-dimensional search is used to determine the optimal step size and update the road segment flow. The iteration stops when the relative gap index is less than or equal to the preset convergence threshold. The balanced flow of each road segment is output as the road network flow distribution prediction result for the target time period.

[0059] Furthermore, the steps involved in traffic distribution prediction in this embodiment include: (1) User-balanced traffic assignment based on predicted OD matrix; The obtained prediction period OD matrix By applying the user-balanced traffic assignment method, the predicted traffic volume is allocated to the road network to obtain the road network traffic distribution for the predicted time period.

[0060] User equilibrium allocation is based on Wardrop's first principle: when the transportation network reaches equilibrium, for any OD pair... All selected paths have the same and minimum travel impedance. ,Right now in, For OD The set of all feasible paths between them; For road section The impedance function varies with the flow rate of the road segment. change.

[0061] (2) Impedance function of road segment; The travel time for each road segment is determined by the BPR impedance function, as shown in equation (6). Free-flow travel time ,in For the length of the road segment, Free flow velocity (must be equal to) (Units consistent) The traffic capacity of a road segment can be determined based on parameters such as road grade and number of lanes; , These are calibration parameters.

[0062] (3) User balanced allocation model; The problem of user-balanced traffic assignment can be expressed as the following variational inequality: Find the traffic flow vector for each road segment. , so that: in, The set of feasible road segment flows that satisfy OD demand constraints, i.e., the flow of existing paths. , making ( )and ( ); For path-segment association variables, when path Included road sections If the value is 1, then the value is 0; otherwise, the value is 0.

[0063] Equivalently, the user equilibrium allocation can be solved using the following mathematical programming model: in, The Beckmann objective function; For OD The set of feasible paths; For OD In the path The path traffic on each path equals the total number of trips for that OD pair. .

[0064] (4) Solve using the Frank-Wolfe algorithm; This embodiment uses the Frank-Wolfe algorithm (successive averaging method) to iteratively solve the above user equilibrium allocation problem. The algorithm flow is as follows: Figure 3 As shown.

[0065] The specific steps are as follows: Step 1: Initialization. Set the iteration count. Traffic flow on each road section ; Step 2: Update the segment impedance. Calculate the current segment impedance according to equation (6). ; Step 3: All-or-nothing auxiliary allocation. Following the all-or-nothing allocation method described above, perform a shortest path search using the current impedance to allocate all OD values. Assign to the shortest path to obtain auxiliary flow. ; Step 4: Step Size Search. Determine the optimal step size. Make Beckmann objective function Minimum; Step 5: Traffic flow update. Update the traffic flow of the road segment according to formula (5); Step 6: Convergence Judgment. Calculate the relative gap according to equation (7). ,in Take the predicted OD amount ;like Then output balanced flow; otherwise, let Return to Step 2.

[0066] (5) Predict flow output; After the Frank-Wolfe algorithm converges, the balanced flow of each road segment is... This refers to predicted flow: The predicted flow rates for each road segment That is, the proposed prediction time period. The road network traffic distribution prediction results are presented. Compared with the all-or-nothing allocation method mentioned above, the user equilibrium allocation can reflect the impact of road segment congestion on route selection. The prediction results are closer to actual travel behavior and can provide decision support for route planning of intelligent connected vehicles and global route guidance for traffic management departments.

[0067] This embodiment is based on the maximum entropy model, Wardrop principle and Beckmann model in the process of OD demand estimation and user-balanced traffic allocation. It organically links the processing chain of "historical traffic flow → historical OD → predicted OD → traffic allocation", so that the causal relationship between changes in road network traffic and the evolution of travel demand is clear and explainable. At the same time, it uses LSTM network to capture the time-series evolution pattern of OD, and achieves a balance between prediction accuracy and dynamic adaptability.

[0068] This embodiment uses a maximum entropy model to infer the OD matrix for each time period from the historical traffic flow of road segments, and then uses an LSTM network to capture the temporal evolution of OD demand, which can reflect the dynamic changes in traffic demand and is superior to the static OD input allocation method in the traditional four-stage method.

[0069] This embodiment employs a nested iterative correction mechanism for historical OD estimation. The inner layer uses a relative gap ε_UE to ensure the convergence of user balanced allocation, while the outer layer uses the relative error of traffic ε_OD to check and update the prior OD value until the relative error between the balanced allocated traffic and the detected traffic meets the preset accuracy requirements (e.g., within 2%), thus ensuring the reliability of the OD back-calculation results.

[0070] This embodiment adopts the user equilibrium allocation method, uses the BPR impedance function to characterize the road segment congestion effect, and uses the Frank-Wolfe algorithm to solve the Wardrop equilibrium state. The predicted road network traffic distribution can reflect the pedestrian's response behavior to path impedance, which is closer to the actual traffic operation state.

[0071] The method in this embodiment makes full use of historical traffic data collected by existing roadside detectors (radar, video, etc.) without the need to deploy additional dedicated OD survey equipment; each step has clear logic and the parameters have clear meanings, which facilitates integration and deployment in intelligent transportation systems and supports dynamic path guidance and traffic management in the context of intelligent connected vehicles and vehicle-road cooperation.

[0072] In the above embodiments, the technical solution is further optimized. There are several alternative technical means in each step of this embodiment. Without changing the purpose of the invention, which is to "estimate OD from historical traffic flow, predict future OD, and then obtain the road network traffic distribution through traffic assignment", the following alternative solutions can be adopted: (1) In response to the replacement of traffic zone division and data collection, in addition to the division method of "road network node is a zone", traffic zones can also be aggregated and divided according to macro sub-zones, traffic analysis zones (TAZ) or functional zones; in addition to radar and video detectors, historical traffic data sources can also come from floating cars (FCD), electronic toll collection (ETC) or integrated multi-source data, and the statistical time interval can still be 5min, 10min, 15min or 30min, etc., according to business needs.

[0073] (2) For alternatives to the OD estimation model, in addition to the maximum entropy model, the generalized least squares (GLS) OD estimation method, Bayesian OD estimation method, balancing method, or a combination model of entropy maximization and information minimization can be used to back-calculate the OD matrix under the traffic flow constraint of road segment detection; the inner and outer layer iterative correction framework (outer layer updates prior OD, inner layer updates diversion coefficient of traffic assignment) can still be used.

[0074] (3) Regarding the substitution of traffic assignment algorithms and equilibrium types, in addition to the Frank-Wolfe algorithm, user equilibrium assignment can also adopt MSA (Successive Average Method), projection algorithm, path capacity constraint algorithm, or other iterative solution methods based on variational inequalities. In some application scenarios, user equilibrium (UE) can also be replaced with system optimal (SO) assignment to complete traffic allocation with the goal of minimizing the total system travel time. The BPR impedance function can also be replaced with other monotonically increasing impedance functions (such as Akcelik function, cone queuing function, etc.).

[0075] (4) For the replacement of OD prediction models, in addition to LSTM, deep learning time series models such as gated recurrent unit (GRU), one-dimensional temporal convolutional network (TCN), Transformer encoder, Temporal Fusion Transformer (TFT) can be used, or traditional time series methods such as vector autoregression (VAR) and state space model can be used to model the historical OD matrix sequence and output the target time period OD matrix; the input and output interface (n historical OD matrices → 1 predicted OD matrix) remains unchanged.

[0076] As a preferred implementation method, such as Figure 1 As shown, taking the Sioux-Falls standard test road network as an example, the complete execution process of each step is explained in sequence from step one to step four.

[0077] Example 1: Sioux-Falls network traffic distribution prediction; The research object of this embodiment is as follows: Figure 5 As shown, the road network comprises 24 traffic zones and 76 directional road segments, with roadside detectors covering all segments. Historical traffic flow is summarized at 30-minute intervals; in this embodiment, traffic flow is detected. , allocate traffic and traffic capacity All measurements were taken at 30-minute statistical intervals (unit: veh / 30min). This embodiment uses Sioux-Falls standard test road network data, with road segment lengths... The unit is mile, and the free-flow velocity is taken as... (approximately) ),and With consistent units, calculate the free-flow travel time according to formula (16). (Unit: h); Traffic capacity Taken from the road network basic database; BPR impedance function parameters are taken from... , ; Convergence threshold for outer layer OD estimation Inner layer user equilibrium convergence threshold The target time period for prediction is 10:00-10:30, and the number of historical time periods input for online prediction is [number missing]. .

[0078] Step 1: Data collection and traffic zone delineation; (1) Road network topology transformation and traffic zone division; Define the road network boundaries within the study area, and number the intersections and dead-end points as follows: These are designated as traffic zones. Nodes in the Sioux-Falls road network are numbered sequentially from "1" to "24"; directed road segments are indicated by "start-end," such as "1-2" representing a segment traveling from traffic zone 1 to traffic zone 2. Free-flow travel time for each segment is calculated based on segment length (miles) and free-flow speed (mph). (2) Collection of historical traffic flow data; Using roadside radar, video, and other detectors, historical traffic flow data for each road segment is collected and aggregated at 30-minute intervals to form a historical traffic volume database. Each record includes the road segment start number, road segment end number, and road segment length. and the flow rate detected in each time period The format is shown in Table 1.

[0079] Table 1 (3) Selection of historical flow data for prediction; The traffic flow of all road segments during the four time periods of 8:00-8:30, 8:30-9:00, 9:00-9:30, and 9:30-10:00 is extracted from the historical traffic volume database to form four sets of road segment traffic flow vectors, which are used for OD estimation in step two and LSTM online prediction in step three.

[0080] Step 2: Historical OD estimation; like Figure 2 As shown, this step performs a nested iteration of "maximum entropy back-calculation of OD matrix - user balanced allocation - accuracy verification" for each historical statistical period until the balanced allocation of traffic matches the detected traffic. The following uses the 8:00-8:30 period as an example to illustrate the complete process.

[0081] (1) Selection of OD matrix estimation model; The detected traffic flow and OD (Original Demand) traffic volume on each road segment satisfy the traffic conservation relationship: Detection traffic on Equal to all OD pairs Travel volume Through the diversion coefficient The weighted sum. Under this constraint, the maximum entropy model maximizes the existence probability of the OD matrix. After solving using the Lagrange multiplier method, the trip volume of each OD is expressed as: in, For prior OD quantity; For the road section The corresponding Lagrange multipliers; For OD On the road section The current splitting coefficient is initialized according to the shortest path indicator. or After iteration, the balanced flow rate is calculated in reverse according to equation (24).

[0082] In the first iteration of the outer layer, with free-flow impedance Perform all-or-nothing assignment: For each OD pair, search for the shortest path; the splitting coefficient of the shortest path segment is set to 1, otherwise it is set to 0; prior OD quantity. Using a uniform constant, numerical optimization methods such as BFGS are employed to solve the problem. This yields the initial OD matrix. In the second round and subsequent outer iterations, the OD matrix from the previous round is used to replace... Updated with the results of the previous round of balanced distribution. Solve the above equation again.

[0083] (2) Iterative solution and balanced traffic allocation for inner-layer users; Each round of outer-layer iterations yields an OD matrix, which must then be used by the inner-layer user-balanced allocation to determine the traffic flow for each segment. User-balanced allocation means that, in a balanced state, each OD pair... All selected paths have the same and minimum path travel impedance. Travelers cannot reduce travel costs by unilaterally changing their routes.

[0084] The inner allocation uses the Beckmann mathematical programming model as its objective to minimize the total travel cost of the system. The constraints are that each OD pair has conservation of trip volume, conservation of segment flow and path flow, and non-negativity of segment flow; the relationship between segment flow and path flow is the same as in formula (14) of the invention, i.e. ( ). For road section Traffic on the internet; The segment impedance, which varies with flow rate, is determined by the BPR function: This embodiment takes , The BPR function characterizes the nonlinear feedback effect of increased travel time as traffic volume increases, and is the key difference between balanced user allocation and all-or-nothing allocation.

[0085] Inner layer allocation uses the Frank-Wolfe algorithm (see...) Figure 3 The solution is obtained through iterative steps, and the specific execution is as follows.

[0086] First, initialization. Set the inner iteration count... Set the traffic flow of all 76 road sections to [a certain value]. .

[0087] Second, update the segment impedance. This includes updating the current flow rate. Substitute into the BPR function to calculate the travel time for each road segment. .

[0088] Third, all-or-nothing auxiliary allocation. Using the current impedance as the weight, for each OD pair... Perform shortest path search (such as Dijkstra's algorithm) on the road network to calculate all OD values. The traffic flow is assigned to the current shortest path and accumulated to obtain the auxiliary path traffic flow: in, For OD The shortest path under the current impedance; This indicates the segments included in the shortest path. Otherwise, take 0. Auxiliary flow This represents the flow rate that each road segment should carry when all travelers take the shortest route under the current impedance.

[0089] Fourth, optimal step size search. (In auxiliary flow...) With current traffic Searching between the steps minimizes the Beckmann objective function : This embodiment uses a bisection method or the golden section method to search on a one-dimensional interval. Since the form of the BPR function is determined, the objective function... It can be parsed and has high search efficiency.

[0090] Fifth, traffic flow update. Update the traffic flow for each road segment using the searched step size: Sixth, convergence determination. Calculate the relative gap index to measure the distance between the current solution and the Wardrop equilibrium state: in, This represents the total travel time of the system. For OD The shortest path travel time under the current impedance. If (This embodiment takes) If the inner iteration terminates, the output will be a balanced flow. Otherwise, Then return to step two and continue iterating.

[0091] After the inner layer converges, record And calculate the diversion coefficient in reverse: in, For OD On the road section Distribute traffic on top.

[0092] (3) Outer layer accuracy verification and iteration; After the inner layer allocation is completed, the relative error between the balanced allocation flow and the detected flow is checked segment by segment: This embodiment takes If all road segments satisfy the above formula, then the OD estimation for that time period is complete; otherwise, proceed to the next round of outer iteration and return to (1) to continue solving.

[0093] (4) Solving for the historical OD matrix over multiple time periods; Traffic flow was detected on road segments during three historical time periods: 8:30–9:00, 9:00–9:30, and 9:30–10:00. The process (1)–(3) was repeated to obtain four historical OD distribution matrices. , , , .

[0094] Step 3: Road network OD prediction; This step utilizes a Long Short-Term Memory (LSTM) network to learn the temporal evolution of demand patterns, starting from the most recent... The historical OD matrix predicts the OD distribution for the target time period. LSTM is a recurrent neural network with a gating mechanism that can selectively retain or forget historical information, effectively alleviating the gradient vanishing and gradient exploding problems of traditional RNNs in long sequence training, and is suitable for characterizing the long-term dependency relationship of OD requirements.

[0095] (1) Construction of LSTM prediction model; like Figure 4 As shown, the LSTM prediction model in this embodiment uses continuous... The OD matrix of a historical time period is used as the time-series input. After being encoded time-by-time by an LSTM network, it is mapped by the output layer to the OD matrix of the target prediction time period. The model is divided into two stages: offline training and online prediction. The offline stage uses the long-term historical OD sequence to learn the evolution pattern; the online stage uses the most recent OD matrix obtained in step two. Using historical OD matrices as input, a single-step prediction is completed.

[0096] The core of the LSTM cell unit lies in three gating structures—the forgetting gate, the input gate, and the output gate—as well as the cell state that carries long-term memory. and candidate states In this embodiment, time Input For the first The vector formed by flattening the OD matrix of each historical period by rows; This is the hidden state from the previous time step, carrying encoded historical OD evolution information. The update formulas for each gating variable and state are as follows: in, , , , The weight matrix to be trained. , , , For bias terms; This represents the concatenation of the hidden state and the current input; For the Sigmoid function, the output is... The gate coefficient within the interval; It is the hyperbolic tangent function; Represents the Hadamard product (element-by-element multiplication). Forget gate. Determines the cell state at the previous moment Which information should be retained; input gate Control the current candidate state What new information is written into the cell state; output gate Which information in the cellular state is regulated to the hidden state? It is exported in the form of [the following].

[0097] During the offline training phase, historical traffic volume data from the Sioux-Falls road network over a longer period (e.g., the past few days) is used to estimate the OD matrix time series according to step two. Each... The OD matrix is ​​flattened by rows into a 576-dimensional vector (including intra-regional trips). (generally small or zero), to slide window ( Constructing training samples: The input for each sample is an OD vector sequence of four consecutive time periods. The output is the OD vector for the next time period. The training and validation sets are divided chronologically, and the mean squared error (MSE) is used to measure the deviation between the predicted and actual values. in, This represents the number of training samples; This is the predicted OD vector output by the LSTM network; Let be the corresponding true OD vector. The Adam optimization algorithm is used to minimize the above loss function, and then updated. , , , The output layer weights are calculated, and the model is saved for online prediction when the validation set loss no longer decreases.

[0098] (2) Generation of OD matrix for prediction period In the online prediction phase, the four historical OD distribution matrices obtained in step two are used... , , , Flattened into vectors , , , The inputs are sequentially fed into the trained LSTM network. At each time step, the network encodes temporal information according to the update relationships between the forget gate, input gate, output gate, and cell state and hidden state. Finally, the output layer outputs the encoded temporal information. Mapped to the predicted OD vector, recovered as The matrix is ​​used to obtain the road network OD distribution matrix for the target time period of 10:00-10:30. .

[0099] Step 4: Flow distribution prediction; This step will The travel volume of each OD pair is used as the allocation demand. User-balanced traffic assignment is performed across the entire Sioux-Falls network to obtain the predicted traffic flow for each road segment. The allocation principle is the same as the inner layer allocation in step two: under Wardrop equilibrium, all used paths between the same OD pair have the same and minimum path travel impedance. The solution process also uses the Beckmann model as the objective, the BPR function to characterize the road segment impedance, and the Frank-Wolfe algorithm to iteratively approximate the equilibrium solution. The difference from step two is that the OD demand here comes from the LSTM prediction results. It is based on the results derived from historical traffic, rather than on the results derived from historical traffic.

[0100] (1) User-balanced traffic assignment based on predicted OD matrix; from Extract each OD pair Predicted travel volume As part of the total allocation demand. For all 76 road sections, according to , and Establish the BPR impedance function (parameters) , This describes the feedback effect of traffic flow on travel time. Based on the road network topology, the feasible path set for each OD pair is pre-determined, and the relationship between paths and road segments is established. User equilibrium allocation uses the Beckmann mathematical programming model as its objective; under Wardrop equilibrium, all used paths between the same OD pair have the same and minimum path travel impedance. .

[0101] (2) Solve using the Frank-Wolfe algorithm; The solution process of the Frank-Wolfe algorithm is as follows: Figure 3 As shown. This embodiment is executed across the entire Sioux-Falls network according to the following procedure: Step 1: Initialization. Let... Traffic flow across all road sections ; Step 2: Update the segment impedance. [The following is a partial translation:] ... Substitute into the BPR function and calculate. ; Step 3: Shortest Path Auxiliary Allocation. Using the current impedance as the weight, perform a shortest path search for each OD pair, and allocate all... Assigned to the shortest path, and accumulated according to the auxiliary flow formula. ; Step 4: Optimal step size search. Within the interval... The up search finds the step size that minimizes the Beckmann objective function. ; Step 5: Traffic Update. (Press...) Update traffic flow on the road section; Step 6: Convergence Judgment. Calculate the relative gap according to equation (23). ,in Pick ;like ,make Return to step two; otherwise, determine that user balance has been achieved.

[0102] The meaning of each step in the Frank-Wolfe algorithm is as follows: Auxiliary allocation gives the system's descent direction under the current impedance; step size search determines the optimal advance distance along this direction; flow rate update gradually brings the current solution closer to the equilibrium point; relative gap... This measures the degree of deviation of the current solution from the Wardrop equilibrium. When Down to In the following case, there is no path impedance between the OD pairs. Smaller, unused paths allow the road network to reach a state of user equilibrium.

[0103] (3) Predict flow output; After the Frank-Wolfe algorithm converges, the balanced flow across all road segments becomes the predicted flow. The traffic flow distribution for all 76 road segments during the 10:00-10:30 time period is obtained. Simultaneously, the travel time for each road segment can be calculated based on the balanced traffic flow. This provides predictive impedance information for path induction.

[0104] The obtained road network traffic flow is the predicted road network traffic flow distribution for the proposed prediction time period, which can be used to support intelligent route guidance and traffic management decisions during that period. Thus, Example 1 fully implements the methods of steps one through four of this example.

[0105] The integrated prediction process of historical flow-driven OD back-dive—time-series prediction—balanced allocation designed in this embodiment uses the historical flow of road segments collected by roadside detectors as the sole (or primary) input, and executes the following sequentially: back-dive the historical OD matrix according to the statistical time period, predict the OD matrix of the target time period using the historical OD sequence as input, and obtain the road network flow distribution by balancing traffic allocation through the predicted OD matrix. This "detected flow..." Historical OD Predicting OD The "balanced flow" closed-loop processing chain constitutes the overall technical architecture of this embodiment, which is different from the end-to-end prediction of road segment flow or the allocation method based on static OD input.

[0106] The maximum entropy OD estimation and inner / outer layer nested iterative correction mechanism designed in this embodiment establishes an OD back-calculation model based on the maximum entropy principle under the constraint of road segment detection flow; and addresses the diversion coefficient... OD travel volume To address the issue of mutual coupling, a nested iterative design is implemented, involving outer-layer OD estimation and inner-layer user balancing allocation: the outer layer balances the relative error between allocated traffic and detected traffic. To convergence criteria and update prior OD values The inner layer uses the Frank-Wolfe algorithm to solve for UE allocation and uses relative gaps. The convergence criterion is used to iteratively update the splitting coefficients. Continue until the accuracy requirements are met.

[0107] The LSTM dynamic OD prediction method based on historical OD matrix time series designed in this embodiment solves the problem time-by-time in step two. Historical OD distribution matrix We construct time-series input samples, use an LSTM network to learn the temporal evolution of OD demand, and output the OD matrix for the target prediction period. Using the OD matrix (rather than road segment traffic flow) as the prediction object and as the demand input for subsequent traffic assignment can dynamically adapt to the evolution of travel demand.

[0108] This embodiment is based on user-balanced traffic assignment using the predicted OD matrix and the Frank-Wolfe solution, and uses the predicted OD matrix output by LSTM. As the total demand for allocation, the user equalization allocation problem is expressed using the Beckmann mathematical programming model, combining the BPR impedance function and Wardrop's first principle. The Frank-Wolfe algorithm (all-or-nothing auxiliary allocation—step-size search—flow update—relative gap convergence determination) is then used to solve for the equalized flow of each segment. By mapping the predicted OD to the road network traffic distribution using UE allocation (rather than all-or-nothing allocation) and outputting traffic predictions, the prediction accuracy is greatly improved.

[0109] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic prediction of road network traffic distribution that integrates data-driven and traffic assignment approaches, characterized in that, include: Based on historical traffic flow data collected by roadside detectors on each road segment of the road network, and pre-divided traffic zones, a traffic flow dataset for each historical statistical period is constructed. Based on the traffic flow dataset of the road segment, OD matrix estimation is performed to obtain the historical OD matrix corresponding to each historical statistical period. Based on the time series formed by the historical OD matrix, the road network OD matrix for the target time period is predicted using a time series prediction model; Based on the road network OD matrix for the target time period, the predicted traffic volume is allocated to each road segment of the road network using a user-balanced traffic allocation method, thereby obtaining the predicted road network traffic distribution for the target time period.

2. The method according to claim 1, characterized in that, The process of obtaining the historical OD matrix corresponding to each historical statistical period includes: For each historical statistical period, the OD matrix is ​​back-derived using the road segment detection flow as a constraint and the maximum entropy OD estimation model. User-balanced traffic allocation is then performed on the back-derived OD matrix. The OD matrix is ​​iteratively corrected based on the relative error between the allocated road segment flow and the detection flow until the preset accuracy requirement is met. The OD matrix finally back-derived for the corresponding period is then used as the historical OD matrix for that period.

3. The method according to claim 2, characterized in that, The process of performing user-balanced traffic allocation on the back-derived OD matrix and iteratively correcting the OD matrix based on the relative error between the allocated road segment traffic and the detected traffic until the preset accuracy requirements are met includes: The outer layer iterative method is used to back-engineer the OD matrix. In each round of outer layer iteration, the OD matrix obtained by back-engineering is used as the demand input. The inner layer iterative method is used to perform user balanced traffic allocation to obtain the balanced traffic allocation of each road segment. The relative error between the balanced distribution of traffic on each road segment and the detected traffic on each road segment is used as the outer convergence criterion. If the preset convergence threshold is not reached, the prior OD quantity is updated based on the OD matrix obtained by the current back-inference, and the OD matrix back-inference and inner user balanced distribution are re-executed until the relative error of all road segments meets the preset accuracy requirements.

4. The method according to claim 3, characterized in that, The process of performing user-balanced traffic allocation using an inner-layer iterative approach includes: Initialize the flow rate of each road segment to zero. Based on the current flow rate of each road segment, update the impedance of each road segment using the impedance function that changes with the flow rate over the travel time of the road segment. Using the current impedance of each road segment as the weight, perform shortest path search for each OD pair, allocate all travel volume of each OD pair to the current shortest path, and accumulate to obtain the auxiliary flow of each road segment; Determine the optimal step size that minimizes the total travel cost of the system, and update the traffic flow of each road segment according to the optimal step size; Calculate the relative gap index used to measure the deviation of the current traffic allocation from the user balance state. If the relative gap index is less than or equal to the preset inner layer convergence threshold, stop the inner layer iteration and output the current traffic of each road segment as the balanced allocation traffic; otherwise, return to the update impedance step to continue the iteration.

5. The method according to claim 4, characterized in that, The impedance function for the change of travel time of the road segment with flow rate adopts a function relationship in which the travel time increases monotonically as the flow rate increases. The function relationship is determined by the free-flow travel time of the road segment, the flow rate of the road segment, the traffic capacity of the road segment, and preset calibration parameters.

6. The method according to claim 2, characterized in that, The maximum entropy OD estimation model takes the detected traffic flow of road segments as a constraint. On the set of all feasible OD matrices that satisfy the traffic flow constraint of road segments, it takes the maximization of entropy as the objective function and solves the traffic volume of each OD by using the Lagrange multiplier method.

7. The method according to claim 6, characterized in that, In the maximum entropy OD estimation model, the trip volume of each OD is determined by the prior OD quantity, the Lagrange multiplier corresponding to each road segment, and the diversion coefficient of each OD pair on each road segment. Specifically, the trip volume of each OD is equal to the prior OD quantity multiplied by the exponential form of the negative exponential function of the product of the Lagrange multiplier and the diversion coefficient.

8. The method according to claim 1, characterized in that, The time-series prediction model is a long short-term memory network model. The long short-term memory network model takes the OD matrix of a number of consecutive historical statistical periods as the input sequence, and outputs the road network OD matrix of the target period after encoding by the long short-term memory network. The Long Short-Term Memory Network model learns the temporal evolution of the OD matrix during the offline training phase, and completes single-step prediction in the online prediction phase using the most recent preset number of historical OD matrices as input.

9. The method according to claim 8, characterized in that, The Long Short-Term Memory (LSTM) network model updates its state at each time step through a forget gate, an input gate, and an output gate. The forget gate determines the retention ratio of the cell state from the previous time step, the input gate controls the degree to which the current candidate state is written into the cell state, the cell state is updated according to the outputs of the forget gate and the input gate, and the output gate adjusts the output of the current hidden state.

10. The method according to claim 1, characterized in that, The user-balanced traffic allocation method is as follows: Using the travel volume of each OD in the road network OD matrix during the target time period as the allocation demand, a successive average iterative algorithm is used to solve the mathematical programming model with user balance as the objective. In each iteration, an all-in-one auxiliary allocation is first performed to determine the descent direction, and then a one-dimensional search is used to determine the optimal step size and update the road segment flow. The iteration stops when the relative gap index is less than or equal to the preset convergence threshold, and the balanced flow of each road segment is output as the road network flow distribution prediction result for the target time period.

Images