A Regional Road Network Optimization Layout Decision-Making Method Based on Traffic Flow Simulation and AI Prediction

Abstract

This invention relates to the field of deep learning technology, specifically to a regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction. The method acquires time-series data from checkpoint monitoring, floating car trajectories, and signal control, generates a multi-dimensional traffic flow feature map according to the regional road network topology, and inputs it into a deep learning spatiotemporal prediction model to obtain a road network prediction state sequence containing traffic flow, speed, queuing trends, and prediction confidence. Based on the prediction confidence, the prediction state is dynamically written into a traffic flow simulation module, the simulation is run, and the state residual between the simulation state and the prediction state is calculated. Based on the state residual, a prediction state calibration factor and a simulation parameter correction factor are generated to update the prediction model input state and the simulation boundary input, respectively. This invention can reduce the input offset caused by the separation of prediction and simulation states, enabling the evaluation of road network layout schemes to be based on traceable and calibrable dynamic traffic states.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction. Background Technology

[0002] Current regional road network layout decisions are typically based on data such as traffic surveys, historical cross-sectional traffic flow, road grades, intersection channelization, and signal timing schemes to establish a traffic demand matrix. This matrix is ​​then input into traffic assignment models or traffic flow simulation software to evaluate the operation of planned, reconstructed, or newly added road networks. These schemes usually divide the research scope according to administrative districts, traffic zones, or major corridors, using historical peak hourly traffic flow, average daily traffic volume, or artificially corrected growth coefficients to represent future traffic demand. Simulation scenarios are then set based on parameters such as road capacity, saturation flow rate, turning ratio, and signal cycle. After the simulation, indicators such as average delay, saturation, queue length, segment speed, and service level are used to determine whether the road network layout scheme meets traffic requirements. While these conventional techniques can describe traffic operation under certain conditions, the input data often comes from static surveys or historical averages. The time-varying characteristics of future traffic demand, the regional traffic transfer process, and the propagation relationship of congestion on adjacent road segments usually rely on manual experience correction, making it difficult to maintain consistency with real-time or near-real-time traffic flow evolution.

[0003] In traffic flow prediction, existing technologies employ multi-source time-series data, including checkpoint monitoring data, floating car trajectory data, signal control data, and navigation speed data, to construct prediction samples. Deep learning models such as recurrent neural networks, graph neural networks, and temporal convolutional networks are then used to predict road segment flow, speed, or congestion indices. These approaches typically abstract the road network as a graph structure, using road segments or intersections as nodes and road connections as edges. A continuous sequence of traffic states is input in the time dimension, allowing the model to learn periodic fluctuations, short-term disturbances, and relationships between adjacent road segments. Some approaches also impute missing data, remove outlier data, or assign different weights based on data sources. While these prediction methods can output traffic conditions for several future time slices, the prediction results typically remain at the data layer, primarily serving congestion warnings or short-term scheduling. If the predicted flow is directly used as input for traffic simulation, the spatiotemporal relationships, confidence differences, and the impact of missing data within the prediction model are often not synchronously transmitted to the simulation environment.

[0004] In traffic flow simulation and road network layout optimization, existing technologies typically use simulation modules as evaluation tools, converting different road network layout schemes into inputs such as road connectivity, number of lanes, capacity, signal control parameters, and traffic demand. Then, macroscopic, mesoscopic, or microscopic simulation models are used to extrapolate vehicle operation processes. For multi-objective optimization, a common approach is to set indicators such as traffic efficiency, construction constraints, congestion risk, or network connectivity, and use heuristic search, genetic algorithms, particle swarm optimization, or reinforcement learning methods to find optimal results among candidate schemes. Since the computation process of the simulation module depends on boundary inputs and traffic demand settings, the evaluation results of candidate schemes are significantly affected by the input state. Existing optimization processes often run the prediction model, simulation model, and optimization model separately. Prediction results are usually not corrected based on the simulation state after entering the simulation, and feedback from queuing overflow, downstream congestion, and road segment saturation generated in the simulation rarely flows back to the prediction state update process.

[0005] The main technical problem with the aforementioned existing technologies lies in the lack of an iterative state assimilation mechanism between deep learning traffic flow prediction results and traffic flow simulation states. While prediction models can provide future flow, speed, or queuing trends based on multi-source time-series data, the predicted states may not satisfy capacity constraints, vehicle conservation constraints, signal phase constraints, and congestion propagation constraints in the simulation module. Simulation modules can extrapolate congestion propagation according to traffic operation rules, but they typically rely on static or semi-static inputs written only once, failing to continuously absorb the prediction model's spatiotemporal evolution judgments of future operating conditions. This results in different states for the same road network in the prediction and simulation spaces, manifesting as acceptable traffic conditions in the prediction results, while simulations show queue overflow or downstream congestion, or stable simulation results that fail to cover future changes in traffic demand. This inconsistency directly affects the reliability of the road network layout evaluation state set, causing regional road network optimization layout decisions to be based on inputs that deviate from the actual evolution process. Summary of the Invention

[0006] The purpose of this invention is to provide a regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction, which can effectively solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction includes: acquiring time-series data of checkpoint monitoring, floating car trajectories and signal control, and generating multi-dimensional traffic flow feature maps according to the regional road network topology; The multi-dimensional traffic flow feature map is input into a deep learning spatiotemporal prediction model to obtain a road network prediction state sequence that includes traffic flow, speed, queuing trend and prediction confidence. The predicted road network state sequence is dynamically written into the traffic flow simulation module according to the predicted confidence level, and the simulation is run. Calculate the state residual between the simulated state and the predicted state sequence of the road network, and generate a predicted state calibration factor and a simulation parameter correction factor based on the state residual; The input state of the deep learning spatiotemporal prediction model is updated with the predicted state calibration factor, and the boundary input of the traffic flow simulation module is updated with the simulation parameter correction factor, forming a road network layout evaluation state set after state assimilation.

[0008] Preferably, the step of generating a multi-dimensional traffic flow feature map based on the regional road network topology includes: encoding road segments, intersections, and traffic zone boundaries in the regional road network as graph nodes, connecting edges, and boundary constraints, respectively. Extract cross-sectional flow and headway from checkpoint monitoring data; extract road segment speed and travel time from floating car trajectories; extract phase sequence and period offset from signal control timing. According to a unified time slice, various features are written into the corresponding graph nodes or connecting edges, and source identifiers, missing masks and time granularity identifiers are configured for each type of feature to obtain a multi-dimensional traffic flow feature map consistent with the regional road network topology.

[0009] Preferably, the multi-dimensional traffic flow feature map is input into a deep learning spatiotemporal prediction model to obtain a road network prediction state sequence, including: constructing a sliding graph sequence using continuous time slices in the multi-dimensional traffic flow feature map, extracting the periodic component and short-term disturbance component of road segment flow using a temporal convolutional layer, and generating congestion propagation weights using a graph attention layer based on the connection relationship, turning relationship and signal phase relationship of adjacent road segments; The missing mask and source identifier are embedded in the prediction process, and the traffic prediction value, speed prediction value, queuing trend code and confidence interval parameter are output respectively. The confidence interval parameter is then converted into prediction confidence.

[0010] Preferably, dynamically writing the road network predicted state sequence into the traffic flow simulation module according to the predicted confidence level includes: dividing the road network predicted state sequence into high confidence state, state to be corrected and uncertain state according to road segment and time slice; For high-confidence states, the corresponding inflow rate, speed distribution, and turning ratio are written into the boundary input of the traffic flow simulation module; For the state to be corrected, the predicted value is fused with the historical state of the same period according to the reliability determined by the source identifier and the missing mask, and then written. For uncertain states, a candidate boundary input set is generated under the conservation constraint of adjacent road segments; The signal phase sequence is synchronously written into the same simulation clock as the aforementioned boundary input.

[0011] Preferably, the configuration of the source identifier, missing mask, and time granularity identifier includes: establishing cross-source aligned records for the traffic flow, floating car speed, and signal phase status of the same road segment within each time slice; Calculate the time delay identifier of each data source between adjacent time slices, and calculate the mutual verification consistency identifier between cross-sectional flow and trajectory velocity within the same time slice; When any source is missing or its granularity is below a uniform time slice, a missing mask is written to the corresponding graph node or connecting edge, and a source reliability is generated based on the time delay identifier, mutual verification consistency identifier, and historical continuity identifier. The source reliability varies with multi-dimensional traffic flow characteristics. Figure One And input the deep learning spatiotemporal prediction model.

[0012] Preferably, the confidence interval parameter is converted into a prediction confidence level, including setting a mean prediction branch, an upper bound prediction branch, and a lower bound prediction branch at the output of the deep learning spatiotemporal prediction model; Uncertainty features are constructed using the source identifier, missing mask, congestion propagation weights generated by the graph attention layer, and prediction bias of consecutive time slices. The uncertainty features are written into the upper and lower bound prediction branches respectively to form the state interval width of each road segment in each future time slice. The prediction confidence is generated based on the width of the state interval and the overlap between adjacent road segments, and the prediction confidence is bound to the traffic flow prediction, speed prediction and queuing trend code for output.

[0013] Preferably, the state residual between the calculated simulation state and the road network predicted state sequence includes: the traffic flow simulation module outputs the simulation flow, simulation speed, queue length and saturation by road segment and intersection after running the simulation; Align the simulated flow rate with the predicted flow rate, the simulated speed with the predicted speed, and the queue length with the queue trend code to generate a state residual matrix. Based on the prediction confidence and the duration of the state residual matrix, the state residual matrix is ​​divided into structural capacity residual, data uncertainty residual, and boundary input residual; Different residual backflow paths are configured for the structural capacity residual, data uncertainty residual, and boundary input residual, respectively, and the residual backflow paths point to the generation process of the prediction state calibration factor or the simulation parameter correction factor.

[0014] Preferably, the process of generating the simulation parameter correction factor includes: when the residual return path points to the simulation parameter correction factor, extracting the inflow rate correction, the steering ratio correction, and the signal phase offset correction based on the boundary input residual; The inflow rate correction is limited to the vehicle conservation relationship between adjacent road segments, the steering ratio correction is limited to the reachable steering relationship at the intersection, and the signal phase offset correction is limited to the existing signal cycle sequence. Write the defined correction values ​​into the boundary input table for the next simulation, and retain the parameter version identifiers and time slice indexes before and after the correction.

[0015] Preferably, the process of generating the predicted state calibration factor includes: when the residual backflow path points to the predicted state calibration factor, establishing a calibration sample based on the source identifier, the missing mask, and the state residual matrix; For structural capacity residuals, the congestion propagation weight of the corresponding road segment is bound to the downstream saturation and written into the calibration sample; For uncertain residuals in the data, the reliability of the corresponding source is bound to the confidence interval parameter and written into the calibration sample; The graph edge weights, node state intervals, and missing feature imputation values ​​in the next round of input states of the deep learning spatiotemporal prediction model are updated using calibration samples.

[0016] Preferably, the process of forming the state-assimilated road network layout evaluation state set includes: recording the predicted state calibration factor, simulation parameter correction factor, simulation state, and state residual matrix at the end of each iteration. When the state residual matrix changes in two consecutive rounds meet the preset convergence flag or reach the preset number of iterations, a state pair is established by selecting the most recent simulation state and the calibrated road network prediction state sequence. The state pairs are grouped according to traffic development scenarios, time slices, and road network area indexes, and written into the road network layout evaluation state set. A confidence level identifier, residual type identifier, and parameter version identifier are added to each state pair.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By generating a multi-dimensional traffic flow feature map consistent with the regional road network topology from time-series data of checkpoint monitoring, floating car trajectories, and signal control, and outputting a road network prediction state sequence containing flow rate, speed, queuing trends, and prediction confidence by a deep learning spatiotemporal prediction model, the input state entering the simulation environment retains information on temporal changes, spatial correlations, and data reliability. After the prediction state sequence is written into the traffic flow simulation module according to confidence level, the difference between the simulation state and the prediction state is converted into state residuals, which are then used to generate prediction state calibration factors and simulation parameter correction factors. The prediction state calibration factors are written back to the input state of the deep learning spatiotemporal prediction model, and the simulation parameter correction factors are written back to the boundary input of the traffic flow simulation module, ensuring that the prediction space and the simulation space maintain the state correspondence under the same time slice, the same road segment, and the same traffic constraints in continuous iterations. The resulting road network layout evaluation state set no longer depends on a single prediction or fixed simulation input, but is generated based on the dynamic road network state after state residual correction, which can reduce congestion propagation misjudgments, boundary input offsets, and evaluation state distortions caused by the separation of prediction state and simulation state.

[0018] 2. In terms of secondary technical effects, multi-source time-series data are input into the deep learning spatiotemporal prediction model through source identification, missing masking, time granularity identification, and source reliability. This enables the flow rate at checkpoints, floating car speeds, and signal phase states to form a traceable data representation within a unified time slice, reducing the interference of inconsistent data granularity from different sources on the prediction input. Temporal convolutional layers and graph attention layers handle the temporal fluctuations of traffic flow and the propagation relationships between adjacent road segments, respectively. The confidence interval parameter and the overlap between adjacent road segment intervals together form the prediction confidence, enabling the simulation writing process to distinguish between high-confidence states, states requiring correction, and uncertain states. State residuals are further divided into structural capacity residuals, data uncertainty residuals, and boundary input residuals, and each enters a different return path, ensuring that model input calibration and simulation boundary correction have corresponding data sources and state types. After continuous iteration, state pairs are written into the evaluation state set according to traffic development scenarios, time slices, and road network area indices, with added confidence identifiers, residual type identifiers, and parameter version identifiers. This facilitates subsequent calculation and evaluation of road network layout schemes under the same state benchmark. Attached Figure Description

[0019] Figure 1 This is a closed-loop flowchart of the regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction of the present invention. Figure 2 This is a flowchart illustrating the construction process of the multi-dimensional traffic flow feature map of the present invention. Figure 3 This is a flowchart of the deep learning spatiotemporal prediction and prediction confidence generation process of the present invention; Figure 4This is a flowchart of the traffic flow simulation state assimilation and road network layout evaluation state set generation process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please refer to Figure 1 This embodiment provides a regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction. The method includes acquiring time-series data from checkpoint monitoring, floating car trajectories, and signal control; generating a multi-dimensional traffic flow feature map according to the regional road network topology; inputting the multi-dimensional traffic flow feature map into a deep learning spatiotemporal prediction model to obtain a road network prediction state sequence containing flow rate, speed, queuing trend, and prediction confidence; dynamically writing the road network prediction state sequence into a traffic flow simulation module according to the prediction confidence and running the simulation; calculating the state residual between the simulation state and the road network prediction state sequence; generating a prediction state calibration factor and a simulation parameter correction factor based on the state residual; updating the input state of the deep learning spatiotemporal prediction model with the prediction state calibration factor; and updating the boundary input of the traffic flow simulation module with the simulation parameter correction factor, forming a road network layout evaluation state set after state assimilation. Specifically, checkpoint monitoring... Measurement data is used to express the vehicle passage status of the regional road network section, floating car trajectories are used to express the vehicle speed and travel time on the road segment, and signal control timing data is used to express the intersection release status. The above data are not input separately in the form of independent tables, but are uniformly mapped to the regional road network topology according to the road segment connection relationship, intersection turning relationship and traffic zone boundary relationship, so that the traffic status within the same time slice has a clear spatial reference basis. Under the constraints of this topology, the deep learning spatiotemporal prediction model outputs the road network prediction state sequence for future multiple time slices. The traffic flow simulation module determines the writing method according to the prediction confidence. After the simulation is completed, the simulation state and the prediction state are aligned and compared. The state residuals are then fed back to the prediction model and the simulation boundary input respectively. This embodiment makes the prediction state, simulation state and layout evaluation state within the same spatiotemporal index system, avoiding the state offset caused by the separation of prediction results and simulation input.

[0022] In this embodiment, reference Figure 2 The multi-dimensional traffic flow feature map uses road segments or intersections in the regional road network as index objects, a unified time slice as the time index, and data source, traffic status, and data reliability as feature dimensions. The multi-dimensional traffic flow feature map can be represented by the following formula: ; in, This represents the feature vector of a node or connecting edge v in time slice t. Indicates the flow rate at the checkpoint cross-section. Indicates the headway of the train. Indicates the floating car speed. Indicates travel time. Indicates signal phase sequence encoding, Indicates signal period offset. , , These represent the missing masks for checkpoint monitoring, floating car trajectory, and signal control timing data, respectively. , , These represent the time granularity identifiers of the corresponding data sources. The superscripts k, f, and s correspond to the time sequence data of checkpoint monitoring, floating car trajectory, and signal control, respectively. For example, when a certain road segment has checkpoint cross-sectional flow and signal phase sequences at time slice t but lacks floating car trajectories... The missing identifier value is taken. When reading the feature vector, the deep learning spatiotemporal prediction model does not treat the missing speed as a true low-speed state, but instead sends the corresponding missing information into the confidence calculation process. In this embodiment, traffic flow state, data source state, and missing state are encoded synchronously to avoid the formation of traffic state in the fusion stage of multi-source data. The fields of the multi-dimensional traffic flow feature map and their writing positions are shown in Table 1.

[0023] Table 1. Fields and Input Locations of Multi-Dimensional Traffic Flow Feature Map ; Table 1 above illustrates the field writing method for the multi-dimensional traffic flow feature map. Checkpoint monitoring features are mainly written to the road segment connection edges, floating car trajectory features are mainly written to the road segment connection edges, and signal control timing features are mainly written to the intersection nodes. Source identifier, missing mask, time granularity identifier, and source reliability are written along with the corresponding traffic state, so that the subsequent deep learning spatiotemporal prediction model can distinguish traffic state values, data source attributes, and data availability attributes. In this embodiment, the same feature map can simultaneously support prediction calculation, confidence generation, and simulation boundary input selection.

[0024] In a preferred embodiment, the step of generating a multi-dimensional traffic flow feature map based on the regional road network topology includes encoding road segments, intersections, and traffic zone boundaries in the regional road network as graph nodes, connecting edges, and boundary constraints, respectively; extracting cross-sectional flow and headway from checkpoint monitoring data; extracting road segment speed and travel time from floating car trajectories; extracting phase sequence and period offset from signal control timing; writing various features into corresponding graph nodes or connecting edges according to a unified time slice; and configuring source identifier, missing mask, and time granularity identifier for each type of feature to obtain a multi-dimensional traffic flow feature map consistent with the regional road network topology. Specifically, road segments can be used as connecting edges to store cross-sectional flow, speed, travel time, and vehicle conservation relationships; intersections can be used as graph nodes to store signal phase sequences, turning relationships, and queueing information. The traffic zone boundary can be used as a boundary constraint to store the permissible directions of inflow and outflow. When the floating car trajectory crosses multiple road segments, the trajectory is first divided into time slice records corresponding to the road segments through road segment matching, and then each segment record is written to the corresponding connecting edge. When the checkpoint monitoring position is at the end of the road segment, the cross-sectional flow is written to the entrance or exit index of the corresponding road segment. When the signal phase is associated with multiple turning directions, the phase code is written to the intersection node and distributed to the adjacent connecting edge through the turning relationship. When the sampling granularity is inconsistent within the same time slice, the low-granularity data retains the original source identifier and indicates its coverage area through the time granularity identifier. In this embodiment, the traffic status in the regional road network no longer depends on loose table splicing, but forms a computational object with topological constraints and temporal constraints.

[0025] Furthermore, the configuration of the source identifier, missing mask, and time granularity identifier includes establishing cross-source aligned records for the checkpoint cross-sectional flow, floating car speed, and signal phase status of the same road segment within each time slice; calculating the time delay identifier of each data source between adjacent time slices; calculating the mutual verification consistency identifier between cross-sectional flow and trajectory speed within the same time slice; when any source is missing or its granularity is lower than the unified time slice, writing a missing mask to the corresponding graph node or connecting edge; and generating source reliability based on the time delay identifier, mutual verification consistency identifier, and historical continuity identifier. The source reliability varies with multi-dimensional traffic flow characteristics. Figure One And input the deep learning spatiotemporal prediction model; the source reliability can be calculated using the following formula: ; in, This indicates the source reliability of source r on index object v and time slice t. Indicates the normalized time delay indicator. Indicates a cross-source mutual verification consistency identifier. Indicates historical continuity. Indicates the time granularity matching identifier. , , , This represents the non-negative weights obtained during training. This represents a function that maps the calculation result to a defined interval. For example, when Take the smaller value and , , When both values ​​are relatively high, the source reliability is close to the high range, indicating that the corresponding source can directly participate in the predicted state output. When the time delay indicator is high and the mutual verification consistency indicator is low, the source reliability decreases, indicating that the corresponding source mainly enters the uncertainty calculation rather than directly replacing other sources. The data fusion process in this embodiment has computable source constraints, reducing the prediction input deviation caused by the equal processing of data with different sampling granularities.

[0026] In a preferred embodiment, reference Figure 3 The multi-dimensional traffic flow feature map is input into a deep learning spatiotemporal prediction model to obtain a road network prediction state sequence. This includes constructing a sliding graph sequence using consecutive time slices from the multi-dimensional traffic flow feature map, extracting the periodic and short-term disturbance components of road segment flow using a temporal convolutional layer, generating congestion propagation weights based on the connection, turning, and signal phase relationships of adjacent road segments using a graph attention layer, embedding the missing mask and source identifier into the prediction process, and outputting traffic flow prediction values, speed prediction values, queuing trend codes, and confidence interval parameters, respectively. The confidence interval parameters are then converted into prediction confidence levels. Specifically, the sliding graph sequence uses feature maps from several consecutive time slices as input windows and the road network state of future time slices as output objects. The temporal convolutional layer performs congestion propagation along the time dimension. Traffic flow changes along a road segment are convolutionally calculated to encode periodic peaks, short-term surges, and short-term declines into different temporal features. The graph attention layer reads the connection and turning relationships between adjacent road segments along the spatial dimension and determines the weight of the propagation of the state of a certain upstream road segment to the downstream road segment by combining the signal phase sequence. The missing mask is used to mask the direct numerical influence of the missing source after entering the model. The source identifier is used to retain the difference in contribution of different data sources to the predicted state after entering the model. The model output is set with traffic flow prediction branch, speed prediction branch, queuing trend branch, and interval parameter branch. Each branch shares the spatiotemporal encoding result but outputs different state variables. In this embodiment, the traffic flow prediction result includes both state values ​​and state confidence boundaries, which facilitates the subsequent simulation module to select the input writing method according to the confidence level.

[0027] In a preferred embodiment, the confidence interval parameter is converted into a prediction confidence score. This includes setting a mean prediction branch, an upper bound prediction branch, and a lower bound prediction branch at the output of the deep learning spatiotemporal prediction model; constructing uncertainty features using the source identifier, missing mask, congestion propagation weights generated by the graph attention layer, and prediction bias in consecutive time slices; writing the uncertainty features into the upper bound prediction branch and the lower bound prediction branch respectively to form the state interval width of each road segment in each future time slice; and generating a prediction confidence score based on the overlap between the state interval width and adjacent road segment intervals. The prediction confidence score can be calculated using the following formula: ; in, This represents the prediction confidence of indexed object v in time slice t. Indicates the width of the normalized state interval. Indicates the degree of overlap between adjacent road segments. Indicates the prediction bias of continuous time slices. , , This indicates non-negative weights. For example, when the prediction interval width is small, the overlap between adjacent road segment intervals is high, and the prediction deviation of consecutive time slices is small, the denominator is close to the smaller interval, the prediction confidence enters the higher interval, and the corresponding prediction state can be used as the main source of simulation boundary input. When the prediction interval width is large or the overlap between adjacent road segment intervals is low, the prediction confidence decreases, and the corresponding prediction state needs to be generated together with the historical state of the same period or vehicle conservation constraints to generate simulation input. In this embodiment, the simulation input is no longer determined solely based on the magnitude of the prediction value, but is written in a hierarchical manner based on the uncertainty structure of the prediction state.

[0028] Furthermore, the predicted road network state sequence is dynamically written into the traffic flow simulation module according to the predicted confidence level. This includes dividing the predicted road network state sequence into high-confidence states, states to be corrected, and uncertain states according to road segments and time slices. For high-confidence states, the corresponding inflow rate, speed distribution, and turning ratio are written into the boundary input of the traffic flow simulation module. For states to be corrected, the predicted values ​​are fused with historical states of the same period based on the reliability determined by the source identifier and missing mask before being written. For uncertain states, a candidate boundary input set is generated under the conservation constraints of adjacent road segments, and the signal phase sequence is synchronously written into the same simulation clock along with the above boundary inputs. Specifically, the inflow rate corresponding to the high-confidence state can be directly used as the road segment entrance boundary condition, and the speed distribution can be... As the initial distribution of vehicle operating speed status, the steering ratio can be written into the steering input table according to the intersection steering relationship. The predicted value corresponding to the state to be corrected does not directly cover the historical state. Instead, the predicted value, historical same-period value and adjacent road segment status are synthesized with the source reliability as the fusion weight. For uncertain states, no single boundary value is set. Instead, a candidate boundary input set that conforms to the upstream inflow, downstream outflow and intersection reachable steering relationship is generated. The simulation module reads the traffic demand input and signal phase input at the same simulation clock during operation to avoid the misalignment between traffic demand changes and signal release status. In this embodiment, the predicted states with different confidence levels have different constraint paths when entering the simulation module, reducing the simulation distortion caused by directly writing low-confidence predicted states.

[0029] In a preferred embodiment, reference Figure 4 After reading the dynamically written boundary input, the traffic flow simulation module updates the simulation state according to the sequence of vehicle entry into the road segment, driving on the road segment, turning at the intersection, and phase release. After the simulation ends, it outputs the simulation flow rate, simulation speed, queue length, and saturation according to the road segment and intersection. The calculation of the state residual between the simulation state and the road network predicted state sequence includes aligning the simulated flow rate with the predicted flow rate, the simulated speed with the predicted speed, and the queue length with the queue trend code, respectively, to generate a state residual matrix. The state residual matrix can be generated using the following formula: ; in, This represents the state residual of indexed object v at time slice t. , , These represent the quantified values ​​of predicted traffic flow, predicted speed, and predicted queuing trend, respectively. , , These represent the quantized values ​​of simulated flow rate, simulated speed, and simulated queue length, respectively. , , The non-negative weights of different state variables are represented. For example, when the predicted flow rate is close to the simulated flow rate but the predicted speed rate is significantly different from the simulated speed rate, the state residual mainly comes from the speed term. This indicates that the predicted demand and the simulated throughput are similar, but the operating state is affected by capacity or queuing constraints. When the queuing term is significantly different, it indicates that the predicted state does not fully express the queuing propagation process. The residual calculation in this embodiment can distinguish between flow rate differences, speed differences, and queuing differences, providing a numerical basis for subsequent calibration path selection.

[0030] Furthermore, based on the prediction confidence level and the duration of the state residual matrix, the state residual matrix is ​​divided into structural capacity residuals, data uncertainty residuals, and boundary input residuals. Different residual backflow paths are configured for the structural capacity residuals, data uncertainty residuals, and boundary input residuals, respectively. These residual backflow paths point to the generation process of the prediction state calibration factor or simulation parameter correction factor. Specifically, when the prediction confidence level of a certain road segment is high and the state residuals are concentrated in the speed and queuing terms within a continuous time slice, while the downstream saturation is in a high-occupancy state, the corresponding residuals are classified as structural capacity residuals and enter the prediction state calibration factor generation process. When the prediction confidence of a segment is low and the residuals appear simultaneously with the missing mask or low source reliability, the corresponding residuals are classified as data uncertainty residuals and enter the prediction state calibration factor generation process. When the residuals are mainly concentrated at the simulation input boundary position and are related to the inflow rate, turning ratio or signal phase offset, the corresponding residuals are classified as boundary input residuals and enter the simulation parameter correction factor generation process. The residual classification is not based on a single numerical value, but simultaneously reads the confidence, residual duration, downstream saturation and source reliability. In this embodiment, the simulation feedback can be converted into different types of reflowable information to avoid all deviations being mistakenly regarded as prediction errors or simulation errors.

[0031] In a preferred embodiment, the generation process of the simulation parameter correction factor includes, when the residual return path points to the simulation parameter correction factor, extracting the inflow rate correction, steering ratio correction, and signal phase offset correction based on the boundary input residual; limiting the inflow rate correction to the vehicle conservation relationship between adjacent road segments; limiting the steering ratio correction to the reachable steering relationship at the intersection; and limiting the signal phase offset correction to the existing signal cycle sequence; writing the limited corrections into the boundary input table for the next round of simulation; and retaining the parameter version identifier and time slice index before and after correction. The iterative update of the boundary input and prediction input can be expressed by the following formula:

[0032] ; in, This represents the boundary input vector of the indexed object v at time slice t in the nth simulation round. This represents the simulation parameter correction factor. Indicates the boundary correction step size. This indicates projecting the boundary input onto the constraint set. The operation, This includes vehicle conservation relationships, intersection accessibility and turning relationships, and existing signal cycle sequences. This represents the input state of the nth round deep learning spatiotemporal prediction model. Indicates the predicted state calibration factor. This indicates the predicted input calibration step size. For example, when the corrected steering ratio exceeds the reachable steering relationship at the intersection, the projection operation will write the corresponding steering ratio back into the allowed steering set. When the signal phase offset correction exceeds the existing signal cycle sequence, the projection operation retains the writable part within the existing cycle. In this embodiment, the simulation boundary correction is always limited by traffic constraints and will not produce an input state that violates vehicle conservation or phase sequence due to residual backflow.

[0033] Furthermore, during the simulation parameter correction process, the inflow rate correction amount originates from the continuous difference between the predicted flow and the simulated flow at the boundary entrance; the turning ratio correction amount originates from the matching result between the queuing differences and reachable turning relationships at each exit direction of the intersection; and the signal phase offset correction amount originates from the time misalignment between the simulated queue dissipation time and the signal phase sequence. After each round of updates, the boundary input table records the inflow rate before correction, the inflow rate after correction, the turning ratio before correction, the turning ratio after correction, the phase offset before correction, the phase offset after correction, the parameter version identifier, and the time slice index. The parameter version identifier is used to indicate which boundary input version a certain simulation state originates from, and the time slice index is used to indicate which simulation moment the correction amount is applied to. When the boundary input table is read in the next round of simulation, only the input corresponding to the current version identifier is read, and historical correction amounts are not called across versions. If the boundary input residuals of the same road segment repeatedly appear in consecutive rounds, the correction amount of the most recent round is retained and the historical correction amount is used as a candidate record. The boundary input correction process in this embodiment has version traceability, which makes it easy to correspond the simulation state with the input parameters that generated the state.

[0034] In a preferred embodiment, the process of generating the predicted state calibration factor includes: when the residual backflow path points to the predicted state calibration factor, establishing calibration samples based on the source identifier, missing mask, and state residual matrix; for structural capacity residuals, binding the congestion propagation weights of the corresponding road segments with downstream saturation and writing them into the calibration samples; for data uncertainty residuals, binding the reliability of the corresponding source with confidence interval parameters and writing them into the calibration samples; and using the calibration samples to update the graph edge weights, node state intervals, and missing feature imputation values ​​in the next round of input states of the deep learning spatiotemporal prediction model. Specifically, the calibration samples corresponding to the structural capacity residuals are used to adjust the edge weights propagating from upstream road segments to downstream road segments in the graph attention layer, so that the downstream... Road segments with high saturation and persistent queuing exhibit stronger congestion propagation representation in the next round of prediction. Calibration samples corresponding to uncertain data residuals are used to adjust the confidence interval width and missing feature imputation values ​​in the input state, ensuring that areas with low source reliability or concentrated missing mask values ​​are not mistakenly written as stable states. Node state intervals are used to limit the acceptable range of traffic flow, speed, and queuing trends within future time slices. Missing feature imputation values ​​are generated jointly by the historical continuity of the same road segment, the state of adjacent road segments, and source reliability, rather than using a global average value. The input state of the prediction model in this embodiment can absorb capacity constraints and data uncertainty information from simulation feedback, thereby making the next round of prediction state closer to the evolution state constrained by traffic constraints.

[0035] Furthermore, during the calibration sample writing process, the structural capacity residual and the data uncertainty residual of the same road segment are not mixed in the same calibration field. The structural capacity residual is written into the edge weight calibration field, downstream saturation field, and queuing persistence field, while the data uncertainty residual is written into the source reliability field, confidence interval width field, and missing mask field. If structural capacity residual and data uncertainty residual exist simultaneously in a certain time slice, they are written into the node calibration sample and the connecting edge calibration sample respectively according to the location where the residual is formed. When the deep learning spatiotemporal prediction model reads the calibration sample in the next round, the graph edge weight update prioritizes reading the field corresponding to the structural capacity residual, the node state interval update prioritizes reading the field corresponding to the data uncertainty residual, and the missing feature filling value reads the adjacent road segment state and source reliability at the same time. The calibration sample retains the same time slice index and road network area index as the state residual matrix to avoid the calibration information being written into the wrong road segment or the wrong time period. The prediction calibration process in this embodiment has a clear data landing point, which can prevent the simulation residual from forming unconstrained diffusion in the model input state.

[0036] In a preferred embodiment, a state-assimilated road network layout evaluation state set is formed. This includes recording the predicted state calibration factor, simulation parameter correction factor, simulation state, and state residual matrix at the end of each iteration. When the changes in the state residual matrix of two consecutive iterations satisfy a preset convergence flag or reach a preset number of iterations, a state pair is established by selecting the most recent simulation state and the calibrated road network predicted state sequence. The state pairs are grouped according to traffic development scenario, time slice, and road network area index, and written into the road network layout evaluation state set. A confidence flag, residual type flag, and parameter version flag are added to each state pair. The road network layout evaluation state set can be represented by the following formula: ; in, This represents the set of evaluation states for the road network layout after state assimilation. This represents the predicted state of road network region a after calibration at time slice t. This indicates the state of road network region a in the most recent simulation at time slice t. Indicates the regional confidence level. Indicates the residual type identifier. Indicates the parameter version identifier. Represents the set of road network areas. This represents a set of time slices. For example, when the state residual matrix of a certain region changes into a stable range in two consecutive iterations, a state pair is established for the corresponding predicted state and simulation state and written into the state set. When the preset number of iterations is reached but the residual is still not stable, the state pair retains the residual type identifier for subsequent layout evaluation. The state set used for road network layout scheme evaluation in this embodiment includes predicted state, simulation state, confidence source, residual type and input version, which can support scheme comparison under the same state benchmark.

[0037] In this embodiment, traffic development scenarios can be composed of historical traffic demand, planned traffic demand, and traffic demand during special periods. Each traffic development scenario corresponds to a set of road network prediction state sequences and a set of traffic flow simulation inputs. The road network area index can be divided according to traffic zones, major channels, or intersection groups. The time slice index is used to connect the prediction state, simulation state, and signal phase state. After the state pairs are established, the road network layout conclusions are not directly output, but are used as the basis for subsequent multi-objective road network layout calculations. Traffic efficiency evaluation reads the simulation speed, simulation flow, and queue length. Construction cost constraints read the change records corresponding to the road network layout variables. Road network resilience evaluation reads the alternative traffic states when key road sections are restricted. If a candidate layout scheme causes the residual type of a certain area in the state set to change from boundary input residual to structural capacity residual, it indicates that the candidate layout scheme has changed the traffic flow distribution but still has capacity constraints. Subsequent iterations need to continue reading the calibration samples corresponding to the structural capacity residuals. In this embodiment, the road network optimization layout decision no longer directly relies on uncalibrated prediction results or isolated simulation outputs, but uses the evaluation state with state source and residual identifier as the calculation basis.

[0038] In one extended application embodiment, the deep learning spatiotemporal prediction model, traffic flow simulation module, and road network layout evaluation state set operate in a closed-loop data flow. The data flow begins with multi-source time-series data entering a multi-dimensional traffic flow feature map, then passes through a sliding graph sequence to enter the prediction model to form a road network prediction state sequence. This sequence is then written to the simulation module under prediction confidence control. The simulation state output by the simulation module is decomposed into a state residual matrix to form a prediction state calibration factor and a simulation parameter correction factor. The prediction state calibration factor affects the graph edge weights, node state intervals, and missing feature imputation values ​​in the next round of prediction model input states. The simulation parameter correction factor... The inflow rate, turning ratio, and signal phase offset in the next round of simulation boundary input are applied to the network layout evaluation state set after several rounds of closed-loop calculation. If a certain source in the input data is continuously missing in multiple time slices, the missing mask will cause the confidence level to decrease and write the corresponding area into a state to be corrected or uncertain into the path. If the simulation module repeatedly outputs queuing spillover caused by downstream saturation, the structural capacity residual will adjust the congestion propagation weight through calibration samples. In this embodiment, a repeatable data closed loop is formed between the prediction model and the simulation module, so that the future traffic demand evolution and simulated traffic constraints can be expressed in the same evaluation state.

[0039] In another extended application embodiment, candidate layout schemes in the regional road network are first converted into topology change records before entering the evaluation. The topology change records include the addition and deletion of connecting edges, adjustment of intersection turning relationships, adjustment of traffic zone boundary inflow direction, and changes in the capacity attributes of key road segments. After the above topology change records are written into the multi-dimensional traffic flow feature map, the corresponding road segment connection relationships and turning relationships are updated synchronously. The deep learning spatiotemporal prediction model reads the updated topology relationships to form a new road network prediction state sequence. The traffic flow simulation module reads the same topology relationships to form a new simulation operation scenario. The state residual matrix is ​​still calculated according to the same index object and time slice. If the candidate layout scheme changes the road segment connection relationships but does not change the signal phase relationships, the signal phase sequence maintains the original version identifier in the simulation clock. If the candidate layout scheme changes the traffic zone boundary inflow direction, the boundary input table synchronously generates a new inflow rate correction space. In this embodiment, the layout scheme changes can simultaneously affect the predicted topology and the simulation topology, avoiding the inconsistency between the prediction model still predicting according to the old road network and the simulation module running according to the new road network.

[0040] In a further embodiment, when the road network layout evaluation state set is used to compare schemes under different traffic development scenarios, each candidate layout scheme corresponds to the same set of traffic development scenarios, the same set of time slices, and the same set of road network area indexes. The confidence index recorded in the state set is used to determine the data support level of the evaluation results, the residual type index is used to determine the source of deviation in the evaluation results, and the parameter version index is used to trace the changes in the simulation boundary input. When comparing schemes, states from different sources are not directly mixed. Instead, states with the same residual type and the same parameter version are read and compared within the same traffic development scenario. If two candidate schemes are similar in traffic status but have different residual types, the candidate scheme with a residual type closer to the correctable range of the boundary input is retained first and the iteration continues. If a candidate scheme generates structural capacity residuals in multiple scenarios, the area corresponding to the layout scheme is marked as the area where topology change records need to be regenerated. The evaluation state set in this embodiment can provide the homogeneity constraints required for scheme comparison, so that the calculation evaluation of candidate layout schemes has a consistent data basis.

[0041] In a fully operational embodiment, the regional road network optimization layout decision-making process is executed under a unified simulation clock. During the data access phase, time-series data from checkpoint monitoring, floating car trajectories, and signal control are converted into multi-dimensional traffic flow feature maps. In the prediction phase, a deep learning spatiotemporal prediction model outputs a road network prediction state sequence containing traffic flow, speed, queuing trends, and prediction confidence. During the simulation phase, based on the prediction confidence, the data is categorized into high-confidence states, states requiring correction, and uncertain states. In the residual phase, the simulated traffic flow, simulated speed, and simulated queue length are aligned with the predicted traffic flow, predicted speed, and predicted queuing trends. Finally, in the backflow phase, structural capacity residuals and data uncertainty residuals are written into the pre-calculated data. The system measures the calibration factor of the state and writes the boundary input residual into the simulation parameter correction factor. During the iteration phase, it updates the prediction model input state and the simulation boundary input respectively. During the evaluation phase, it generates a road network layout evaluation state set with confidence level identifier, residual type identifier, and parameter version identifier. The whole process revolves around the consistency between the prediction state and the simulation state. It does not introduce hardware structures or external business rules that are unrelated to the regional road network traffic flow calculation. In this embodiment, multi-source data processing, deep learning prediction, traffic flow simulation, and layout evaluation state generation are organized into the same closed-loop calculation process, so that the basic state on which the regional road network optimization layout decision is based has the technical attributes of traceability, calibrability, and reusability.

Claims

1. A regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction, characterized in that, include: Acquire time-series data from checkpoint monitoring, floating car trajectories, and signal control, and generate multi-dimensional traffic flow feature maps based on the regional road network topology; The multi-dimensional traffic flow feature map is input into a deep learning spatiotemporal prediction model to obtain a road network prediction state sequence that includes traffic flow, speed, queuing trend and prediction confidence. The predicted road network state sequence is dynamically written into the traffic flow simulation module according to the predicted confidence level, and the simulation is run. Calculate the state residual between the simulated state and the predicted state sequence of the road network, and generate a predicted state calibration factor and a simulation parameter correction factor based on the state residual; The input state of the deep learning spatiotemporal prediction model is updated with the predicted state calibration factor, and the boundary input of the traffic flow simulation module is updated with the simulation parameter correction factor, forming a road network layout evaluation state set after state assimilation.

2. The regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction according to claim 1, characterized in that, The process of generating a multi-dimensional traffic flow feature map based on the regional road network topology includes: encoding road segments, intersections, and traffic zone boundaries in the regional road network as graph nodes, connecting edges, and boundary constraints, respectively. Extract cross-sectional flow and headway from checkpoint monitoring data; extract road segment speed and travel time from floating car trajectories; extract phase sequence and period offset from signal control timing. According to a unified time slice, various features are written into the corresponding graph nodes or connecting edges, and source identifiers, missing masks and time granularity identifiers are configured for each type of feature to obtain a multi-dimensional traffic flow feature map consistent with the regional road network topology.

3. The regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction according to claim 2, characterized in that, The multi-dimensional traffic flow feature map is input into a deep learning spatiotemporal prediction model to obtain a road network prediction state sequence, including: constructing a sliding graph sequence with continuous time slices in the multi-dimensional traffic flow feature map, extracting the periodic component and short-term disturbance component of road segment flow using a temporal convolutional layer, and generating congestion propagation weights based on the connection relationship, turning relationship and signal phase relationship of adjacent road segments using a graph attention layer. The missing mask and source identifier are embedded in the prediction process, and the traffic prediction value, speed prediction value, queuing trend code and confidence interval parameter are output respectively. The confidence interval parameter is then converted into prediction confidence.

4. The regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction according to claim 3, characterized in that, Dynamically writing the road network prediction state sequence into the traffic flow simulation module according to the prediction confidence level includes: dividing the road network prediction state sequence into high confidence state, state to be corrected and uncertain state according to road segment and time slice; For the high confidence state, the corresponding inflow rate, speed distribution, and turning ratio are written into the boundary input of the traffic flow simulation module; For the state to be corrected, the predicted value is fused with the historical state of the same period according to the reliability determined by the source identifier and the missing mask, and then written. For the aforementioned uncertain state, a candidate boundary input set is generated under the conservation constraint of adjacent road segments; The signal phase sequence is synchronously written into the same simulation clock as the boundary input.

5. The regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction according to claim 2, characterized in that, The configuration of the source identifier, missing mask, and time granularity identifier includes: establishing cross-source alignment records for the traffic flow, floating car speed, and signal phase status of the same road segment within each time slice; Calculate the time delay identifier of each data source between adjacent time slices, and calculate the mutual verification consistency identifier between cross-sectional flow and trajectory velocity within the same time slice; When any source is missing or its granularity is lower than the unified time slice, the missing mask is written into the corresponding graph node or connecting edge, and the source reliability is generated based on the time delay identifier, mutual verification consistency identifier and historical continuity identifier. The source reliability is input into the deep learning spatiotemporal prediction model along with the multi-dimensional traffic flow feature map.

6. The regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction according to claim 3, characterized in that, The confidence interval parameter is converted into prediction confidence, including setting a mean prediction branch, an upper bound prediction branch, and a lower bound prediction branch at the output of the deep learning spatiotemporal prediction model. Uncertainty features are constructed using the source identifier, missing mask, congestion propagation weights generated by the graph attention layer, and prediction bias of consecutive time slices. The uncertainty features are written into the upper bound prediction branch and the lower bound prediction branch respectively to form the state interval width of each road segment in each future time slice. The prediction confidence level is generated based on the width of the state interval and the overlap between adjacent road segments, and the prediction confidence level is bound to the traffic flow prediction value, speed prediction value and queuing trend code for output.

7. The regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction according to claim 4, characterized in that, The state residual between the calculated simulation state and the road network predicted state sequence includes: the traffic flow simulation module outputs the simulation flow, simulation speed, queue length and saturation by road segment and intersection after running the simulation; Align the simulated flow rate with the predicted flow rate, the simulated speed with the predicted speed, and the queue length with the queue trend code to generate a state residual matrix; Based on the prediction confidence level and the duration of the state residual matrix, the state residual matrix is ​​divided into structural capacity residual, data uncertainty residual, and boundary input residual; Different residual backflow paths are configured for the structural capacity residual, data uncertainty residual, and boundary input residual, respectively, and the residual backflow paths point to the generation process of the prediction state calibration factor or the simulation parameter correction factor.

8. The regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction according to claim 7, characterized in that, The process of generating the simulation parameter correction factor includes: when the residual return path points to the simulation parameter correction factor, extracting the inflow rate correction, the steering ratio correction, and the signal phase offset correction based on the boundary input residual; The inflow rate correction is limited to the vehicle conservation relationship between adjacent road segments, the steering ratio correction is limited to the reachable steering relationship at the intersection, and the signal phase offset correction is limited to the existing signal cycle sequence. Write the defined correction values ​​into the boundary input table for the next simulation, and retain the parameter version identifiers and time slice indexes before and after the correction.

9. The regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction according to claim 7, characterized in that, The process of generating the predicted state calibration factor includes: when the residual backflow path points to the predicted state calibration factor, establishing a calibration sample based on the source identifier, missing mask and state residual matrix; For the structural capacity residual, the congestion propagation weight of the corresponding road segment is bound to the downstream saturation and written into the calibration sample; For the uncertain residuals in the data, the reliability of the corresponding source is bound to the confidence interval parameter and written into the calibration sample; The calibration samples are used to update the graph edge weights, node state intervals, and missing feature imputation values ​​in the next round of input states of the deep learning spatiotemporal prediction model.

10. The regional road network optimization layout decision-making method based on traffic flow simulation and AI prediction according to claim 9, characterized in that, The process of forming a state-assimilated road network layout evaluation state set includes: recording the predicted state calibration factor, simulation parameter correction factor, simulation state, and state residual matrix at the end of each iteration. When the state residual matrix changes in two consecutive rounds meet the preset convergence flag or reach the preset number of iterations, a state pair is established by selecting the most recent simulation state and the calibrated road network prediction state sequence. The state pairs are grouped according to traffic development scenarios, time slices, and road network area indexes, and written into the road network layout evaluation state set. A confidence index, residual type index, and parameter version index are added to each state pair.

Images