Congestion tracing and strategy generation method and device based on path backstepping

Abstract

The embodiment of the application provides a congestion tracing and strategy generation method and device based on path backstepping, and relates to the technical field of road traffic operation control. The method comprises the following steps: generating an OD prior matrix based on historical travel data, time characteristics and road network state; loading the simulation results to the mesoscopic cell road network according to the proportion of the candidate path; iteratively correcting the OD based on the observation and simulation deviation and the OD prior, and outputting the corrected OD and road section data; determining the congestion cause, supply-demand gap and contribution object according to the corrected OD and road section data; generating a candidate strategy, executing the candidate strategy after pre-playing and safety checking, and updating the parameters in the direction according to the pre-playing and actual measurement deviation. The scheme of the application forms a closed-loop control which is traceable, controllable and iterative.

Description

Technical Field

[0001] This application relates to the field of road traffic operation management technology, specifically to a method and apparatus for congestion source tracing and strategy generation based on path backpropagation. Background Technology

[0002] In road traffic management, multi-source data such as detectors, video capture, floating cars, and mobile phone signaling can reflect the traffic operation status of the road network at different spatiotemporal scales. Existing technologies adopt a modular and independent architecture to realize functions such as traffic demand forecasting, road segment state estimation, congestion identification, and signal control: traffic demand forecasting extrapolates future travel demand based on historical traffic sequences; route allocation maps origin-destination (OD) travel demand to the road network topology; congestion diagnosis locates low-speed traffic or queuing sections of the road network; and control optimization adjusts signal timing, driving guidance routes, ramp release rates, entrance flow, or lane organization schemes according to real-time traffic conditions.

[0003] However, the aforementioned stages employ different data sources, time scales, and evaluation criteria, lacking a unified physical state benchmark and a complete causal transmission chain. The predicted OD demand struggles to reasonably explain the cross-sectional flow and operating speed of on-site detectors; identified congested road sections are not necessarily network activation bottlenecks, but may merely be derivative results of congestion propagation; control algorithms generate regulatory actions solely based on the current state of low-speed road sections, failing to trace the source OD, propagation path, and the root causes of supply-demand mismatch corresponding to congestion. This results in a disconnect between traffic prediction, congestion diagnosis, and control, hindering the formation of a coordinated and systematic management capability.

[0004] Existing technologies suffer from the following main technical shortcomings: First, OD demand prediction is disconnected from the actual road network conditions. Predicting traffic demand solely based on historical OD or traffic flow time series data makes it difficult to simultaneously meet the multi-dimensional observation constraints of multiple detection sections, such as traffic flow, speed, and queue length. Correcting OD solely through traffic flow differences can easily misinterpret detection noise, speed fluctuations, or state judgment errors as demand changes, further exacerbating prediction bias. Second, there is a lack of implementable mapping mechanisms between mobile phone signaling and traffic cell OD. Mobile phone signaling is mostly user dwell and movement records at the base station or grid level, which cannot be directly used as vehicle OD data between traffic cells. Without stable anchor points and calibration rules, it is impossible to accurately distinguish between local activities, commuting, and non-commuting trips, making it difficult to support road network-level traffic demand extrapolation. Third, the path allocation process is incomplete, and the state mapping relationship is unclear. Some solutions only output candidate paths or static road obstructions, without specifying the path allocation algorithm, path cost composition, and parameter update mechanism. This fails to fully depict the transformation process of OD demand into road segment traffic flow, queue length, and lane occupancy status, resulting in a lack of clear physical correspondence between macroscopic demand and microscopic road network conditions. Fourth, congestion identification, cause analysis, and control actions are disconnected. Identified low-speed road sections may be the result of upstream queue propagation, rather than the source of congestion. Even when bottleneck sections are located, it's difficult to distinguish between causes such as demand overload, insufficient signal supply, lane mismatch, ramp merging conflicts, inconsistent speed limits, or sudden capacity drops. Because the causes of congestion, contributing factors, and control gaps cannot be clearly identified, the type and magnitude of control actions lack quantitative basis. Fifth, strategy generation and execution feedback lack a controllable closed-loop iterative mechanism. There are inherent model errors between simulation and actual execution results. Existing technologies use a single comprehensive score to correct all parameters, failing to accurately pinpoint whether the deviation originates from OD prediction, path selection, road capacity, or congestion threshold. This results in a lack of directionality in parameter updates, hindering targeted iterative optimization of the control system. Summary of the Invention

[0005] At least one embodiment of this application provides a method and apparatus for congestion tracing and strategy generation based on path backpropagation, which is used to solve the problem that existing road traffic control schemes cannot build an integrated system of prediction, estimation, diagnosis and control, and are difficult to form a traceable, controllable and iterative closed-loop control capability.

[0006] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for congestion tracing and strategy generation based on path backpropagation, including: In the current control cycle, based on historical travel data, time characteristics, and the road network status of the previous control cycle, a priori matrix of origin and destination points (OD) corresponding to traffic zones and the prediction time domain is generated. Based on the non-zero OD in the OD prior matrix, candidate paths are generated, the path ratio of the candidate paths is determined according to a preset selection model, the traffic of each path is determined based on the path ratio, and the path traffic is loaded into the mesoscale cellular road network to determine the mesoscale simulation results. The observation data of the detection section and the mesoscopic simulation results are projected onto the same macroscopic traffic flow basic map. The OD is iteratively corrected with the deviation between observation and simulation and the OD prior matrix as constraints. After each correction, the road segment status is updated and the residual is checked until the residual converges or reaches the preset iteration limit. The corrected OD and the corrected road segment data are then output. Based on the corrected OD and the corrected road segment data, determine the causes of congestion, the supply-demand gap, and the objects that contribute the most to congestion. Based on the causes of congestion, the supply-demand gap, and the contributing objects, candidate control strategies are generated. The candidate control strategies are then pre-rendered and safety-verified in the meso-level cellular road network. After the verification is passed, the strategies are sent to the execution terminal, and the corresponding parameter groups are updated according to the deviation between the pre-rendering and the actual measurement.

[0007] Optionally, after the current control cycle ends, the method further includes: Save the corrected OD, the path ratio, the corrected road segment data, and the strategy execution record; In the next control cycle, the corrected OD and the path ratio of the current control cycle are used to replace the road network state of the previous control cycle to generate the OD prior matrix of the next control cycle. If the observation data of the next control cycle conflicts with the corrected OD, the observation data of the next control cycle shall be used for recalibration.

[0008] Optionally, in the current control cycle, after generating the origin-destination (OD) prior matrix corresponding to traffic zones and the prediction time domain based on historical travel data, time characteristics, and the road network status of the previous control cycle, the method further includes: Obtain operator base station location records, map the base station coverage area to the traffic cell, perform jitter removal on short-term jumps of the same terminal between adjacent base stations, and retain records exceeding a preset displacement threshold as valid traffic cell transfers; For the effective traffic community transfer and stop points, based on a preset time period, residential anchor points and employment anchor points are determined respectively; Based on the spatial relationship and round-trip periodicity between the residential anchor point and the employment anchor point, the effective traffic zone transfer is divided into commuting, local activities and non-commuting. The effective traffic cell transfers are expanded by deduplicating devices and using travel conversion coefficients to determine the composition ratio of commuting, local activities, and non-commuting for each traffic cell pair in the OD prior matrix; the composition ratio is used to determine the path ratio of the candidate paths.

[0009] Optionally, in the current control cycle, based on historical travel data, time characteristics, and the road network status of the previous control cycle, a priori origin-destination (OD) matrix corresponding to the traffic zone and the prediction time domain is generated, including: Establish spatial indexes based on traffic zones, road sections, and detection sections; align historical OD sequences, time characteristics, and road network status of the previous control cycle with a unified timestamp to determine the aligned input data. Based on the historical OD sequence in the aligned input data, the spatial coding result is determined by aggregating adjacent and upstream / downstream traffic zone information according to adjacency and historical travel connections using the traffic zone map. Based on the spatial coding results, the temporal features and the road network status are taken as exogenous variables in chronological order and fused with the spatial coding results to determine the temporal coding results of each traffic zone. Traffic cells with missing or insufficient historical origins in the time-encoding results are assigned low-confidence markers, and the missing values ​​are replaced with preset smoothing values; Based on the time coding results of each traffic cell, a priori matrix of origin and destination points (OD) corresponding to the traffic cell and the prediction time domain is generated.

[0010] Optionally, based on the non-zero ODs in the OD prior matrix, candidate paths are generated; the path proportions of the candidate paths are determined according to a preset selection model; the traffic flow of each path is determined based on the path proportions; and the path traffic flow is loaded into the mesoscale cellular road network to determine the mesoscale simulation results, including: Based on the non-zero ODs in the OD prior matrix, prohibited road sections and paths with unreachable turns are eliminated to generate candidate paths; Based on the corrected road segment data from the previous control cycle, the generalized cost of the candidate path is calculated. Based on the generalized cost, the path ratio of each candidate path is determined according to a preset selection model; Based on the path ratio and the corresponding OD travel volume, the traffic flow for each path is determined; The traffic flow of each path is loaded into the mesoscopic cellular road network to determine the mesoscopic simulation results; the mesoscopic simulation results include at least one of the following: path ratio, path-segment correlation, segment traffic flow, segment speed, segment density, queue length, and lane occupancy status.

[0011] Optionally, the observation data of the detection section and the mesoscopic simulation results are projected onto the same macroscopic traffic flow basic map. The OD is iteratively corrected using the deviation between observation and simulation and the OD prior matrix as constraints. After each correction, the road segment status is updated and the residuals are checked until the residuals converge or reach the preset iteration limit. The corrected OD and the corrected road segment data are then output, including: The observation density is calculated based on the ratio of observed flow rate to observed velocity in the observed data; Based on the critical density and hysteresis interval of the macroscopic traffic flow basic graph, the branch of the basic graph in which the observed density is located is determined; The observed flow rate, observed speed, observed density, and road segment flow rate, road segment speed, and road segment density from the mesoscopic simulation results are projected onto the same basic graph branch to determine the corrected observation state and simulation state. Based on the deviation between the corrected observed state and the simulated state, flow residuals and velocity residuals are constructed. Using the flow residual, the speed residual, and the OD prior matrix as prior constraints, the OD is iteratively corrected to obtain the corrected OD; after each correction, the road segment status is updated based on the corrected OD to obtain the corrected road segment data, and the flow residual and the speed residual are verified. The iteration stops when the relative change of the corrected OD, the change of the flow residual, and the change of the velocity residual are all less than the corresponding preset thresholds, or when the preset iteration limit is reached. Output the corrected OD and the corrected road segment data.

[0012] Optionally, based on the corrected road segment data, the cause of congestion is determined, including: Based on the corrected road segment data, a directed propagation chain for queuing is established along the road topology direction; Based on the directed propagation chain, when the queuing duration reaches the minimum confirmation duration and the passage speed of the passage direction section recovers to above the preset threshold, the queuing source section is determined. Based on the arrival demand and current effective traffic capacity of the queuing source section, when the arrival demand continuously exceeds the current effective traffic capacity, the signal timing, lane function, speed limit, ramp release rate and event status corresponding to the queuing source section are read. Based on the difference between the arrival demand and the current effective traffic capacity, as well as the read signal timing, lane function, speed limit, ramp release rate and event status, the cause of congestion is determined; the cause of congestion includes at least one of the following: demand overload, insufficient signal supply, lane configuration mismatch, ramp merging conflict, speed limit incoordination, sudden drop in event capacity, and decrease in traffic capacity.

[0013] Optionally, identify the supply-demand gap and the entities that contribute most to congestion, including: Based on the corrected OD, the arrival demand for entering the queuing source section is determined; Based on the supply parameters and confirmed events of the aforementioned meso-cellular road network, the effective traffic capacity is determined; The supply-demand gap is determined based on the difference between the arrival demand and the effective capacity; and; Based on the corrected OD, for the preset quantity combinations that rank high, keeping other conditions unchanged, the corresponding requirements are reduced and a rehearsal is performed, and the combination whose queue length change direction is consistent with the share ranking direction is taken as the contribution object.

[0014] Optionally, based on the congestion causes, the supply-demand gap, and the contributing entities, candidate control strategies are generated. These candidate control strategies are then rehearsed and validated in the meso-level cellular road network. After successful validation, the strategies are sent to the execution terminal, and the corresponding parameter groups are updated based on the deviation between the rehearsal and the actual measurements, including: The action type is determined based on the cause of the congestion, the target object is determined based on the contributing object, and the search range is determined based on the supply and demand gap. Based on the action type, the target object, and the search range, candidate control strategies are generated within preset safety constraints; The candidate control strategy is input into the meso-level cellular road network to determine the path cost after the action is implemented. The path ratio is recalculated based on the path cost after the action is implemented to determine the meso-level simulation result after iteration. Using the corrected OD and the iterative mesoscale simulation results as initial conditions, the candidate control strategy is pre-simulated in the mesoscale cellular road network to determine the pre-simulation results; A security check is performed on the pre-simulation results. Candidate control strategies that fail the check are eliminated. Candidate control strategies that pass the check are sorted according to management objectives, and the optimal solution is selected. If all candidates are eliminated, the current status is retained. The selected optimal solution is compiled into an execution terminal instruction and issued. After execution, the pre-simulation results are compared with the actual test results. The corresponding parameter group is updated according to the source of the deviation. The corresponding parameter group is frozen when the data is abnormal.

[0015] Secondly, embodiments of this application provide a road traffic cooperative control device based on origin and destination points, comprising: The first processing module is used to generate a priori origin-destination (OD) matrix corresponding to traffic zones and the prediction time domain, based on historical travel data, time characteristics, and the road network status of the previous control cycle in the current control cycle. The second processing module is used to generate candidate paths based on the non-zero ODs in the OD prior matrix, determine the path ratio of the candidate paths according to a preset selection model, determine the traffic of each path based on the path ratio, load the path traffic into the mesoscale cellular road network, and determine the mesoscale simulation results. The third processing module is used to project the observation data of the detection section and the mesoscopic simulation results onto the same macroscopic traffic flow basic map, and iteratively correct the OD with the deviation between observation and simulation and the OD prior matrix as constraints. After each correction, the road segment status is updated and the residual is checked until the residual converges or reaches the preset iteration limit, and the corrected OD and the corrected road segment data are output. The fourth processing module is used to output the causes of congestion, the supply-demand gap, and the objects that contribute the most to congestion based on the corrected OD and the corrected road segment data. The fifth processing module is used to generate candidate control strategies based on the congestion causes, the supply-demand gap, and the contributing objects, perform a pre-run and safety verification of the candidate control strategies in the meso-level cellular road network, and send them to the execution terminal after the verification is passed, and update the corresponding parameter group according to the deviation between the pre-run and the actual measurement.

[0016] Thirdly, embodiments of this application provide a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the method described in any of the first aspects.

[0017] Fourthly, embodiments of this application provide a computer program product, including computer instructions that, when executed by a processor, implement the steps of the method described in any of the first aspects.

[0018] Compared with existing technologies, the congestion tracing and strategy generation method and apparatus based on path backpropagation provided in this application generates an origin-destination (OD) prior matrix based on historical travel data, time characteristics, and the road network status of the previous control cycle, thus connecting demand forecasting with the road network situation; it generates candidate paths based on non-zero ODs, determines the path proportions according to a preset selection model, and loads the flow of each path into the meso-level cellular road network to determine the meso-level simulation results, so that ODs are stably mapped to road segment status; it projects the observation data of the detection section and the meso-level simulation results onto the same macro-level traffic flow basic map, and iteratively corrects the ODs with the deviation between observation and simulation and the OD prior matrix as constraints, and verifies the residuals each time until the residuals converge or reach a preset value. After iterating to the upper limit, the corrected OD and corrected road segment data are output, making the corrected OD traceable. Based on the corrected OD and corrected road segment data, the causes of congestion, supply-demand gaps, and the objects that contribute the most to congestion are determined, so that subsequent control strategies can be explained by the diagnostic results. Candidate control strategies are generated according to the causes of congestion, supply-demand gaps, and contributing objects, and are performed in the same meso-level cellular road network for pre-simulation and safety verification. After the verification is passed, the strategies are sent to the execution terminal, and the corresponding parameter groups are updated according to the deviation between the pre-simulation and the actual measurement, so that the closed loop has controllability and iterability. The scheme of this application overcomes the problem that it is difficult to form traceable, controllable, and iterative closed-loop control capabilities in the existing road traffic control separation design scheme. Attached Figure Description

[0019] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is one of the flowcharts illustrating the congestion tracing and strategy generation method based on path backpropagation provided in this application embodiment; Figure 2 The second flowchart illustrates the congestion tracing and strategy generation method based on path backpropagation provided in this application embodiment. Figure 3 A schematic diagram of the basic macroscopic traffic flow map provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the congestion tracing and strategy generation device based on path backpropagation provided in the embodiments of this application. Detailed Implementation

[0020] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, without limiting the number of objects; for example, the first object can be one or more. Furthermore, "or" in this application indicates at least one of the connected objects. For example, "A or B" covers three scenarios: Scenario 1: including A but not B; Scenario 2: including B but not A; Scenario 3: including both A and B. The character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0021] The term "instruction" in this application can be either a direct instruction (or explicit instruction) or an indirect instruction (or implicit instruction). A direct instruction can be understood as one in which the sender explicitly informs the receiver of specific information, the operation to be performed, or the requested result, etc.; an indirect instruction can be understood as one in which the receiver determines the corresponding information based on the instruction sent by the sender, or makes a judgment and determines the operation to be performed or the requested result, etc., based on the judgment result.

[0022] As described in the background section, existing technologies suffer from several problems, including a disconnect between OD prediction and on-site detection status, incomplete path allocation and meso-level road network loading, fragmented congestion bottleneck identification, cause diagnosis, and control actions, and a lack of closed-loop directional updates between strategy pre-drafting and execution feedback. These issues make it difficult to form a complete, physically consistent link from demand estimation to strategy execution within a unified control cycle. To address at least one of these problems, this application provides a method and apparatus for congestion tracing and strategy generation based on path backpropagation. This method can reduce or avoid the occurrence of these problems by sequentially completing OD prior generation, path allocation and meso-level loading within the same meso-level cellular road network, and based on basic... Figure 1 The system incorporates OD inversion, bottleneck tracing and contribution identification, candidate strategy pre-simulation and security verification, and directional parameter updates based on component deviations to achieve physically consistent closed-loop control from demand estimation to strategy execution.

[0023] This application provides a method and apparatus for congestion tracing and strategy generation based on path backpropagation. The method and apparatus are based on the same concept, and since they solve problems using similar principles, their implementations can be mutually referenced; repeated details will not be elaborated further.

[0024] Please refer to Figure 1 This application provides a method for congestion tracing and strategy generation based on path backpropagation, comprising: Step 11: In the current control cycle, based on historical travel data, time characteristics and the road network status of the previous control cycle, generate the origin-destination (OD) prior matrix corresponding to the traffic zone and the prediction time domain. Step 12: Based on the non-zero OD in the OD prior matrix, generate candidate paths, determine the path ratio of the candidate paths according to the preset selection model, determine the traffic of each path based on the path ratio, and load the path traffic into the mesoscale cellular road network to determine the mesoscale simulation results. Step 13: Project the observation data of the detection section and the mesoscopic simulation results onto the same macroscopic traffic flow basic map. Use the deviation between observation and simulation and the OD prior matrix as constraints to iteratively correct the OD. After each correction, update the road segment status and check the residual until the residual converges or reaches the preset iteration limit. Output the corrected OD and the corrected road segment data. Step 14: Based on the corrected OD and the corrected road segment data, determine the causes of congestion, the supply-demand gap, and the objects that contribute the most to congestion; Step 15: Based on the congestion cause, the supply-demand gap, and the contributing object, generate a candidate control strategy, perform a pre-run and safety verification of the candidate control strategy in the meso-level cellular road network, and after the verification is passed, send it to the execution terminal and update the corresponding parameter group according to the deviation between the pre-run and the actual measurement.

[0025] It should be noted that the scheme in this application uses the control cycle as the outer operating unit and the simulation step of the meso-cellular road network as the inner state advancement unit. The meso-cellular road network is a simulation system that uses a discrete model of cellular automata to simulate how vehicles flow, congest, and disperse within a regional road network. Each control cycle executes the following steps sequentially: Step 11 generates the OD prior matrix; Step 12 assigns paths and loads the road network; Step 13 performs OD inversion; Step 14 determines the causes of congestion, the supply-demand gap, and the contributing objects; Step 15 generates candidate control strategies, performs pre-simulation and safety verification, and issues and updates parameters. The corrected OD and path ratio of the previous control cycle serve as the initial constraints for the next control cycle, enabling continuous updates.

[0026] It should also be noted that, unlike schemes that stack multiple algorithms in parallel, this application sets only one OD inversion target. A pre-defined spatiotemporal prediction model is used to provide an OD prior matrix, which does not directly replace the inversion; a meso-level cellular road network is used to map OD to observable road segment states; flow and speed residuals jointly drive OD updates; congestion causes, supply-demand gaps, and contributing objects directly constrain strategy search; the same set of state variables and evaluation quantities are used for both pre-simulation and actual execution. The above closed loop uses a single meso-level cellular road network engine. This engine expands the road topology into a three-level structure of links, lanes, and cells, storing the length of each cell, the number of vehicles it can accommodate, free-flow speed, congestion wave speed, lane connection relationships, and steering constraints, and updating the sending, receiving, and transfer states of vehicle groups in fixed small steps. Signal phase, ramp merging, lane closure, temporary lane occupancy, and speed limit commands are all written into the same engine as supply-side boundaries, thus queue formation, queue tail tracing, and downstream recovery have the same physical caliber.

[0027] In this embodiment, step 11 generates an OD prior matrix for the current control period based on historical travel data, time characteristics, and the road network status of the previous control period, providing demand input for subsequent path allocation; step 12 allocates non-zero ODs in the OD prior matrix to candidate paths, determines the path ratio and flow rate of each path according to a preset selection model, and loads the path flow rate into the meso-level cellular road network to obtain meso-level simulation results that can be aligned with the detection section; step 13 projects the observation data of the detection section and the meso-level simulation results onto the same macro-level traffic flow basic map, iteratively corrects the OD using flow residuals and speed residuals as drivers and the OD prior matrix as constraints, and outputs corrected ODs and corrected road segment data; step 14 determines the causes of congestion, supply-demand gaps, and the objects that contribute the most to congestion based on the corrected ODs and corrected road segment data, providing diagnostic basis for strategy generation; step 15 generates candidate control strategies based on the diagnostic results, performs pre-simulation and safety verification in the meso-level cellular road network, and sends the strategy to the execution terminal after verification, and updates the corresponding parameter group according to the component deviation between the pre-simulation and the actual measurement, forming a closed loop.

[0028] This can also be understood as follows: Step 11 of this application transforms historical travel data, time characteristics, and road network status into an OD prior matrix; Step 12 assigns ODs to candidate paths and loads them into the mesoscopic cellular road network; Step 13 corrects ODs based on on-site traffic flow and speed, outputting corrected ODs and corrected road segment data; Step 14 determines the causes of congestion, supply-demand gaps, and contributing factors; Step 15 generates candidate control strategies, completing pre-running, safety verification, deployment, and parameter updates. The outputs of each step directly serve as inputs for the next step. During continuous operation, the corrected ODs and path ratios of the previous cycle serve as initial constraints for the next cycle; if the observed data of the current cycle conflicts with these constraints, the data of the current cycle is used for recalibration to avoid the permanent accumulation of historical errors.

[0029] In an optional embodiment, after the current control cycle ends, the method further includes: Save the corrected OD, the path ratio, the corrected road segment data, and the strategy execution record; In the next control cycle, the corrected OD and the path ratio of the current control cycle are used to replace the road network state of the previous control cycle to generate the OD prior matrix of the next control cycle. If the observation data of the next control cycle conflicts with the corrected OD, the observation data of the next control cycle shall be used for recalibration.

[0030] In this embodiment, after each control cycle ends, the corrected OD, path ratio, corrected road segment data, and strategy execution records for that cycle are saved as a traceable state snapshot. At the beginning of the next control cycle, the state snapshot of the previous control cycle is used, that is, the corrected OD and path ratio of the previous control cycle are used to replace the road network state of the previous cycle in the historical travel data, as the initial constraint for generating the OD prior matrix of the next cycle, thus realizing continuous recursion between control cycles. When the detection section observation data of the next cycle deviates significantly from the inherited corrected OD, the OD inversion is re-executed based on the observation data of the current cycle, avoiding the cumulative propagation of the estimation error of the previous cycle to subsequent cycles.

[0031] In an optional embodiment, after step 11 described above, the method further includes: Obtain operator base station location records, map the base station coverage area to the traffic cell, perform jitter removal on short-term jumps of the same terminal between adjacent base stations, and retain records exceeding a preset displacement threshold as valid traffic cell transfers; For the effective traffic community transfer and stop points, based on a preset time period, residential anchor points and employment anchor points are determined respectively; Based on the spatial relationship and round-trip periodicity between the residential anchor point and the employment anchor point, the effective traffic zone transfer is divided into commuting, local activities and non-commuting. The effective traffic cell transfers are expanded by deduplicating devices and using travel conversion coefficients to determine the composition ratio of commuting, local activities, and non-commuting for each traffic cell pair in the OD prior matrix; the composition ratio is used to determine the path ratio of the candidate paths.

[0032] In this embodiment, the operator base station positioning records are not directly used as vehicle OD, but are used to form auxiliary constraints of the OD prior matrix according to the following process: 1) Spatial mapping. The operator base station positioning records are obtained, and the base station coverage area or signaling grid is superimposed on the traffic cell polygon. For records that cover multiple traffic cells simultaneously, spatial weights are assigned according to coverage area, road accessibility, and trajectory continuity at adjacent times. When a positioning point can be obtained, the point landing area and road proximity constraints are used first. For short-term round-trip jumps of the same terminal between adjacent base stations, the overlapping coverage, dwell time, and trajectory forward and backward directions are combined. Only when the continuous displacement exceeds a preset threshold is it identified as a valid traffic cell transfer, so as to eliminate false cross-area records caused by base station handover and positioning drift. 2) Anchor point identification. The anonymous terminal trajectory is subjected to dwell point identification. Records that continuously dwell for more than a set time and whose position drift is less than a set threshold are merged into a dwell point, which is used for anchor point judgment after cross-day stability check. Among them, for the identification of stop points for effective traffic zone transfers, residential anchor points and employment anchor points are determined based on preset time periods. Specifically, high-frequency locations that repeatedly appear on multiple effective workdays during the nighttime are identified as residential anchor points; high-frequency locations that form stable stops during the daytime on workdays are identified as employment anchor points. Terminals with unstable multi-center employment or travel patterns retain multiple candidate anchor points and reduce classification confidence. 3) Three-class classification. Records where residential anchor points and employment anchor points are located in different traffic zones and the round trips are periodic are marked as commuter type; records where departure and arrival are both located in the same traffic zone or preset adjacent zones and do not constitute stable residential-employment round trips are marked as local activity type; records with effective cross-regional displacement but not meeting the commuter periodicity are marked as non-commuter type; records that cannot form stable anchor points or have insufficient trajectory continuity are not included in the classification and are only used for total calibration. 4) Expansion and proportion determination. The effective traffic zone transfer sample is expanded according to equipment deduplication and vehicle travel conversion coefficient. The vehicle travel expansion coefficient is calibrated using checkpoints, detectors, or known total OD within the same time window, and non-negativity and total consistency constraints are applied to the three types of components, ensuring that the sum of the three types of demand equals the OD prior corresponding to the OD prior matrix determined in step 11. When data is insufficient, the classification ratio of the most recent valid period is used with a reduced confidence level. This yields the local activity, commuting, and non-commuting components of the current or future OD prior. The three types of components are not additional independent ODs, but rather components within the same OD prior, used for differentiated settings of subsequent path response parameters and policy acceptability.

[0033] It should be noted that the jitter reduction rule uses a sliding time window. For short-term jumps between adjacent base stations of the same terminal, the data is combined with coverage overlap, dwell time, and trajectory direction. Only when the continuous displacement exceeds a preset displacement threshold is it identified as a valid traffic cell transfer, in order to eliminate false cross-region records caused by base station handover, positioning drift, and repeated reporting.

[0034] Optionally, the determination of residential and employment anchor points involves comparing not only frequency of occurrence but also cross-day stability and consistency over time. High-frequency nighttime locations must repeat across multiple valid workdays, while daytime anchor points must exhibit stable dwell times during weekday daytime hours. For terminals with multi-center employment or unstable travel patterns, multiple candidate anchor points are retained, and the classification confidence is reduced. Anchor points are only marked as commuter-type locations when both spatial relationships and round-trip periodicity reach thresholds, avoiding long-term classification based solely on a single morning or evening round trip.

[0035] Optionally, during the expansion process, equipment deduplication is first performed within traffic zones and time windows, and then the stratified expansion coefficient is determined based on operator coverage and travel conversion coefficients. Detectors, floating cars, or existing survey data can be used to correct deviations in specific areas and time periods. If the three components after expansion are inconsistent with the total amount of the OD prior matrix, proportional calibration is performed according to confidence level and non-negativity constraints, and the total amount of the OD prior matrix is ​​retained as the unified demand benchmark before proceeding to step 13 inversion.

[0036] The differences among the three groups are reflected in step 12 regarding route cost sensitivity, probability of compliance with guidance, and ability to adjust departure time. Stable commuters are constrained by fixed arrival times and route habits, resulting in smaller route shifts; non-commuters are more sensitive to recommended detours and departure time adjustments; and local residents have shorter journeys and are less sensitive to minor detours within the region. Specific parameters are calibrated using historical route shifts before and after guidance releases, survey samples, or anonymous trajectory statistics, and are gradually revised in step 15 based on implementation feedback.

[0037] It should be explained that the composition ratio determined here can be used for subsequent path allocation, for setting policy acceptance based on differences in different population groups, and for scaling the three components proportionally according to the composition ratio after correcting the total OD. This avoids mobile signaling generating a separate set of ODs while retaining the population information useful for policy acceptance.

[0038] In a specific implementation method, refer to Figure 2 As shown, Figure 2 This is a schematic diagram illustrating the overall process of the congestion tracing and strategy generation method based on path backpropagation provided in this application embodiment. The method uses a control cycle as the outer operating unit, and each control cycle sequentially goes through the stages shown in the diagram, as detailed below: Multi-source data. Corresponding to the input stage of step 11. Historical travel data, time characteristics, road network status of the previous control period, and operator base station location records are collected and aligned according to a unified timestamp to form aligned input data.

[0039] Prior knowledge and population demographics. This corresponds to step 11 and the processing of operator base station location records. Based on the aligned input data, spatial encoding, temporal encoding, and missing value downgrading are performed to generate an OD prior matrix corresponding to traffic cell pairs and the prediction time domain. Simultaneously, based on operator base station location records, through shaking, anchor point identification, three-class classification, and expansion, the composition ratio of commuting, local activities, and non-commuting activities for each traffic cell pair in the OD prior matrix is ​​determined, for subsequent path allocation and strategy generation based on population-specific differences.

[0040] Path allocation and mesoscopic loading. Corresponding to step 12. Based on the non-zero ODs in the OD prior matrix, candidate paths are generated. The path ratio and traffic flow of each path are determined according to the preset selection model. The traffic flow of each path is loaded into the mesoscopic cellular road network, and the mesoscopic simulation results are determined. The mesoscopic simulation results include, but are not limited to, path ratio, path-segment relationship, segment traffic flow, segment speed, segment density, queue length, and lane occupancy status.

[0041] Physically constrained OD inversion. Corresponding to step 13. Project the observed data of the detection section and the mesoscopic simulation results onto the same macroscopic traffic flow base map. After branch projection of the base map, construct the flow residual and velocity residual. Use the OD prior matrix as a prior constraint to iteratively correct the OD, and output the corrected OD and the corrected road segment data.

[0042] Determine the causes, contributions, and gaps. (Corresponding to step 14) Based on the corrected road segment data, establish a queuing directed propagation chain along the road topology to determine the queuing source section and the cause of congestion; determine the supply-demand gap based on the corrected OD; and determine the contributing object with the greatest contribution to congestion based on the corrected OD and path-segment correlation, through share ranking and short-term pre-simulation verification.

[0043] Determine the strategy. This corresponds to the strategy generation stage in step 15. Based on the congestion causes, determine the action type; based on the contributing objects, determine the target; based on the supply-demand gap, determine the search range; generate candidate control strategies within preset safety constraints; input the candidate control strategies into the meso-cellular road network; and through path cost updates and alternating manager-traveler iterations, determine the road segment status after rerouting iterations.

[0044] Strategy Pre-Drill. This corresponds to the pre-drill and deployment stage in step 15. Using the corrected OD and the corrected road segment data as initial conditions, the candidate control strategy is pre-drilled in the meso-level cellular road network. The pre-drill results are then subjected to safety verification. The candidate control strategies that pass the verification are sorted according to management objectives, and the optimal solution is selected, compiled into an execution terminal command, and deployed.

[0045] Execute feedback and proceed to the next control cycle. This corresponds to the parameter update in step 15 and the transition to the next control cycle. After execution, compare the pre-simulation results with the measured results, and update the corresponding parameter group according to the source of the deviation. Save the corrected OD, the path ratio, the corrected road segment data, and the strategy execution record. In the next control cycle, replace the road network state of the previous control cycle with the corrected OD and the path ratio to generate the OD prior matrix for the next control cycle. If the observed data of the current cycle conflicts with it, recalibrate with the observed data of the current cycle, and then enter the cycle of the next control cycle.

[0046] In an optional embodiment, step 11 described above includes: Establish spatial indexes based on traffic zones, road sections, and detection sections; align historical OD sequences, time characteristics, and road network status of the previous control cycle with a unified timestamp to determine the aligned input data. Based on the historical OD sequence in the aligned input data, the spatial coding result is determined by aggregating adjacent and upstream / downstream traffic zone information according to adjacency and historical travel connections using the traffic zone map. Based on the spatial coding results, the temporal features and the road network status are taken as exogenous variables in chronological order and fused with the spatial coding results to determine the temporal coding results of each traffic zone. Traffic cells with missing or insufficient historical origins in the time-encoding results are assigned low-confidence markers, and the missing values ​​are replaced with preset smoothing values; Based on the time coding results of each traffic cell, a priori matrix of origin and destination points (OD) corresponding to the traffic cell and the prediction time domain is generated.

[0047] In this embodiment, three types of spatial indexes are established based on traffic zones, road segments, and detection sections. Historical windows and input data for the current control cycle are aligned using a unified timestamp to determine the aligned input data. The input data includes: historical OD sequences, time characteristics, and the road network status of the previous control cycle. The time characteristics include time, day of the week, holidays, and weather or large-scale event tags; the road network status includes at least traffic flow and speed, and may also include queue length, occupancy rate, event status, and policy execution records from the previous control cycle.

[0048] Based on the historical OD sequences in the aligned input data, spatial coding results are determined by aggregating adjacent and upstream / downstream traffic cell information according to adjacency and historical travel connections using a traffic cell map. Edges in the traffic cell map are established based on geographical adjacency, historical travel connections, or shared reachable paths. For each time period within a historical window, adjacent traffic cells, upstream / downstream cells in the same corridor, and cell pairs with strong historical connections are aggregated according to adjacency and historical travel connections. When describing remote connections, multi-hop aggregation is performed, with upper limits set on the number of hops and weights to prevent remote noise from overshadowing local demand changes. Spatial coding results are saved according to specific origin-destination traffic cell pairs, without compressing all ODs into a single network state.

[0049] Based on the spatial coding results, temporal features and road network status are treated as exogenous variables in chronological order and fused with the spatial coding results to determine the temporal coding results for each traffic zone. The temporal coding is recursively applied according to the historical window sequence, utilizing pre-defined gated cyclic units or spatiotemporal attention mechanisms, while preserving stable patterns such as weekday cycles and morning / evening peak hours, as well as short-term changes in recent cycles. Temporal features and road network status are concatenated or gatedly fused with the spatial coding results during the recursive process. The prediction time domain includes a pre-defined number of control cycles; the model's hidden state is used only for internal calculations, and the external output is an OD prior matrix that corresponds one-to-one with the traffic zone pairs and the prediction time domain. Traffic zones with missing or insufficiently covered historical ODs in the temporal coding results are assigned low-confidence labels, and downgraded priors are formed using the most recent date of the same type, smoothed values ​​of adjacent time periods, or zonal growth coefficients. Low-confidence labels do not delete the corresponding ODs; instead, in step 13, the prior uncertainty range is increased, the update amplitude of unobservable directions is reduced, or the corresponding ODs are frozen to ensure that the inversion has definite boundaries. Furthermore, based on the time coding results of each traffic cell, a priori matrix of origin-destination points (OD) corresponding to the traffic cell and the prediction time domain is generated.

[0050] Optionally, this application can directly utilize a prediction model to receive the aforementioned historical OD sequence, time characteristics, and road network status from the previous control period as input, and directly output the OD prior matrix. This prediction model can employ graph convolution combined with gated recurrent networks, spatiotemporal attention networks, or other network structures; the form of the network structure is not specifically limited here. Since the real-time detector only observes flow and velocity at a limited cross-section, the same set of cross-section observations can be generated by multiple OD and path combinations, making real-time OD inversion an underdetermined problem. The OD prior matrix output by the prediction model is used to limit the reasonable scale, direction of change, and prediction time domain range of each traffic sub-zone pair, reducing the drift of unobservable ODs. The OD prior matrix is ​​not the final OD and still undergoes on-site physical state correction after entering step 13.

[0051] Optionally, this application can use historical offline inversion samples that have been quality-screened to train the aforementioned prediction model. Once online, this prediction model is not retrained for each control cycle. Only during holidays, severe weather, large-scale events, or long-term changes in control schemes that cause data distribution shifts, small-step updates are performed on the output calibration layer or a small number of parameters using the most recent high-quality samples, while retaining the previous version for rollback.

[0052] Furthermore, step 12 above includes: Based on the non-zero ODs in the OD prior matrix, prohibited road sections and paths with unreachable turns are eliminated to generate candidate paths; Based on the corrected road segment data from the previous control cycle, the generalized cost of the candidate path is calculated. Based on the generalized cost, the path ratio of each candidate path is determined according to a preset selection model; Based on the path ratio and the corresponding OD travel volume, the traffic flow for each path is determined; The traffic flow of each path is loaded into the mesoscopic cellular road network to determine the mesoscopic simulation results; the mesoscopic simulation results include at least one of the following: path ratio, path-segment correlation, segment traffic flow, segment speed, segment density, queue length, and lane occupancy status.

[0053] Optionally, the input data for step 12 of this application includes, but is not limited to: the OD prior matrix determined in step 11, the composition ratio determined above, static road network data, and the segment cost of the previous control cycle. Here, the static road network data includes, but is not limited to: road network topology, turning connection relationships, road capacity, free flow time, and signal control status. The segment cost of the previous control cycle is the segment travel time, queuing delay, and signal waiting time determined after loading the mesoscopic cellular road network in step 12 of the previous control cycle. Among them, the road network topology, turning connection relationships, road capacity, and free flow time are static basic data, which are only updated when the road network is modified or facilities are changed; the signal control status is a dynamic input, which is read from the signal controller at the beginning of each control cycle; the segment cost of the previous control cycle is a closed-loop recursive quantity, which is directly provided by the mesoscopic simulation results of the previous cycle and does not require external input.

[0054] In this embodiment, for each non-zero OD, a candidate path set is constructed using the K-shortest path, constrained K-shortest path, or equivalent feasible path generation algorithm, excluding paths with restrictions, road closures, unreachable turns, and those exceeding detour limits. Free-flow time and traffic restriction constraints are used when candidate paths are initially generated; after closed-loop operation, candidate paths that are still feasible in the previous control cycle are retained, and additions or deletions are only made when a path becomes unreachable, detour rates exceed limits, costs remain consistently high, or new effective alternative paths emerge. Emergency reallocation triggered by road interruptions or major events only affects the affected ODs and does not recalculate irrelevant paths across the entire network. At the beginning of the control cycle, the generalized cost of each candidate path is calculated based on the corrected road segment data from the previous control cycle. The generalized cost includes at least road segment travel time, queuing delay, and signal waiting time; when there are toll collection, ramp control, no-turn, or lane restrictions, corresponding costs or infeasibility constraints are added.

[0055] Based on generalized cost, the path proportion of each candidate path is determined according to a preset selection model (such as using the Logit random selection model): Formula (1), where, This represents the proportion of candidate paths (od, od) selected given an OD pair and a population category (g). Here, g represents the population category index (commuting, local activity, non-commuting), corresponding to the aforementioned proportion; r represents the number of the candidate path whose selection probability is being calculated; and h is the summation index, used for iteration. All candidate paths in the list; od represents the origin-end point pair number; Represents the set of candidate paths for OOD pairs; For the cost sensitivity parameter of the g-th group; Let the generalized cost of the r-th candidate path be, for example, This includes travel time, queue delays, signal waiting, and penalties for activated controls; Let represent the generalized cost of the h-th candidate path. This formula is a standard multinomial Logit model, where the numerator is the exponential utility of path r, and the denominator is the sum of the exponential utilities of all candidate paths for that OD, ensuring that the sum of the proportions of all paths within the same OD and the same population category equals 1. Different populations g use different... This allows for differentiated path allocation based on different population groups.

[0056] Based on the path ratio and the corresponding OD travel volume, the demand of each OD is multiplied by the corresponding candidate path ratio to determine the flow of each path. The flow of a certain OD undertaken by a certain road segment is obtained by adding the flow of all candidate paths passing through that road segment, thereby forming a path-road segment association relationship that can be updated with the control cycle. This is used in step 14 to determine whether a certain OD or path passes through the queuing source section and its contribution to the supply-demand gap.

[0057] Traffic flow from each path is loaded into the meso-cellular road network, and vehicle groups, queues, and lane occupancy are advanced according to preset simulation steps. In each simulation step, the meso-cellular road network engine calculates the number of vehicles that upstream cells can send and the number of vehicles that downstream cells can receive. The actual transfer volume is limited by both, along with lane connectivity, turning ratio, signal release, and ramp priority rules. When the downstream receiving capacity is less than the upstream sending demand, untransferred vehicles remain upstream and form queues. Once the queue reaches the capacity of the upstream cells, it overflows to higher upstream cells. Traffic flow corresponds to the number of vehicles passing through a cross-section per unit time, density corresponds to the number of vehicles per unit road length, speed is determined by the consistent traffic flow-density state, and queue length and lane occupancy are directly calculated from the same loading.

[0058] The outer control cycle is responsible for updating the OD prior matrix, candidate paths, and path ratios, consistent with the detector aggregation cycle and control command cycle, and can be set to 5 minutes or as needed for management. The inner meso-level simulation step is responsible for advancing vehicle groups, signal phases, queues, and lane occupancy, and can be set to 1 to 5 seconds. At the end of a control cycle, the simulation step-level states are summarized into flow rate, velocity, density, and queuing indicators consistent with the detection section, determining the meso-level simulation results.

[0059] After each load, a conservation and feasibility check is performed: the sum of the flows from each OD to all candidate paths equals the demand of that OD; changes in entry segments, exit segments, and inventory are conserved within the allowable error range; the number of vehicles, path ratios, speeds, and available capacity are non-negative; and no illegal transfers occur during red lights, no-turn zones, and closed lanes. If the check fails, the current state does not proceed to step 13, but returns to check path connectivity, simulation steps, or control boundaries.

[0060] The mesoscopic simulation results output in step 12 of this application include, but are not limited to: path proportion, path-segment correlation, and segment traffic flow, segment speed, segment density, queue length, and lane occupancy status statistically analyzed by segment or detection section. Signal timing, lane function, speed limit, and ramp release rate are simulation condition inputs and should not be mixed with the above status outputs.

[0061] It should be noted that free-flow time and restricted traffic constraints are used when candidate paths are initially generated. After closed-loop operation, the actual travel time, queuing delay, and signal waiting time obtained from the mesoscopic simulation of the previous control cycle are used to update the path cost. To avoid frequent changes in the candidate set under small fluctuations, paths that were still feasible in the previous cycle can be retained, and added or removed only when the path becomes unreachable, the detour rate exceeds the limit, the cost remains high, or a new effective alternative path appears. Emergency reallocation triggered by road interruptions or major events only applies to the affected ODs and does not recalculate irrelevant paths across the entire network.

[0062] In an optional embodiment, step 13 described above includes: The observation density is calculated based on the ratio of observed flow rate to observed velocity in the observed data; Based on the critical density and hysteresis interval of the macroscopic traffic flow basic graph, the branch of the basic graph in which the observed density is located is determined; The observed flow rate, observed speed, observed density, and road segment flow rate, road segment speed, and road segment density from the mesoscopic simulation results are projected onto the same basic graph branch to determine the corrected observation state and simulation state. Based on the deviation between the corrected observed state and the simulated state, flow residuals and velocity residuals are constructed. Using the flow residual, the speed residual, and the OD prior matrix as prior constraints, the OD is iteratively corrected to obtain the corrected OD; after each correction, the road segment status is updated based on the corrected OD to obtain the corrected road segment data, and the flow residual and the speed residual are verified. The iteration stops when the relative change of the corrected OD, the change of the flow residual, and the change of the velocity residual are all less than the corresponding preset thresholds, or when the preset iteration limit is reached. Output the corrected OD and the corrected road segment data.

[0063] It should be noted that before calculating residuals (such as flow rate residuals and speed residuals), the detectors, floating cars, and the aforementioned mesoscopic simulation results are aligned to the same cross-section, lane range, and time window. Flow rates are uniformly converted to the same statistical duration, and speeds are standardized in units, distinguishing between spatial average speed and time average speed. Availability tags are set for records that are missing, duplicated, out of range, constant for a long time, or significantly contradictory to adjacent cross-sections. Missing values ​​can be used for demonstrative imputation but are not included in OD updates as high-confidence residuals. For multi-lane cross-sections, flow rate and occupancy are aggregated according to detector coverage, and then the cross-section speed is formed by the number of vehicles or the number of effective lanes. If the on-site cross-section only covers part of the lanes, the aforementioned mesoscopic simulation results also only aggregate the same set of lanes. Data preprocessing simultaneously saves the reliability, coverage, and anomaly type of each cross-section for use in the flow rate weight, speed weight, and subsequent verification in the inversion target.

[0064] In this embodiment, the observation density is calculated for the observed flow rate and observed velocity of each observation data point (i.e., the effective detection section). To avoid division by zero due to low velocity or missing values, a lower limit protection is applied to the velocity. The formula for applying the lower limit protection to the observation density is as follows: , formula (2); This represents the observation density at the detection section d at time t; This represents the observed flow rate at the detection section d at time t; This represents the observation velocity at the detection section d at time t; This represents the lower speed limit protection value, which is a preset constant to prevent division by zero when the speed is low or missing.

[0065] Furthermore, based on the critical density and hysteresis interval of the macroscopic traffic flow basic map, the branch of the basic map where the observed density is located is determined. When the observed density is significantly lower than the critical density, it is judged as a free flow branch; when it is significantly higher than the critical density, it is judged as a congested branch. When it is in the critical interval, the branch, speed change and queuing status of the previous time period are referenced to avoid frequent branch jumps caused by observation noise.

[0066] Furthermore, by projecting the observed and simulated states onto the same branch, the corrected flow rate, velocity, and density are obtained: Formula (3); where, The projection operator is used to represent the feasible set of projections onto the base graph by weighted distance; the inputs are the observation triples and the simulation triples, respectively, to obtain the observation correction state and the simulation correction state on the same branch of the base graph. The triplet representing the flow rate-velocity-density at the detection section d at time t can be either the observed triplet or the simulated triplet. This represents the corrected triplet after projection, i.e., the corrected observation state or the corrected simulation state. This represents the feasible set of the macroscopic traffic flow basic map corresponding to the detection section d. The basic map is divided into free-flow branches or congested branches by the critical density. d represents the detection section index; t represents the current time or simulation layout index.

[0067] Projection constrains the original observation or simulation state back to feasible points on the same basic graph branch, so that subsequent residual calculations are performed within the same physical feasible region, avoiding misjudging observation noise or numerical errors as OD bias.

[0068] It should be noted that the critical density of the macroscopic traffic flow basic map can be derived from the model of the macroscopic traffic flow basic map, rather than a preset constant. The macroscopic traffic flow basic map is used to describe the relationship between flow rate, speed, and density. Different road segments have different basic map curves due to differences in road grade, number of lanes, design speed, and infrastructure conditions, and therefore correspond to different critical densities.

[0069] Critical density can be determined or updated in the following way: using historical flow-speed data (detector, floating car, or survey data) of a road segment, a basic graph curve for that road segment is fitted, and the density corresponding to the peak flow in the curve is taken as the critical density of that road segment. During the iteration in step 13, if the flow-speed relationship deviation of a certain road segment continues to exceed the preset range, the recalibration of the macroscopic traffic flow basic graph parameters can be triggered, and the corresponding critical density is updated. When data is insufficient, the critical density of roads of the same level or adjacent comparable road segments is used and assigned a lower confidence level. Each road segment (or the road segment corresponding to the detection section) saves its own critical density as part of the macroscopic traffic flow basic graph parameters, and is maintained together with parameters such as free flow velocity and congestion wave velocity.

[0070] For example, a basic macroscopic traffic flow map can be like this: Figure 3 As shown, the horizontal axis represents density, and the vertical axis represents flow rate. The curve is divided into two branches: the free-flow branch, where flow rate increases linearly with density from 0 to the critical density, and the velocity approximates the free-flow velocity; and the congested branch, where flow rate decreases with increasing density after the density exceeds the critical density, and the velocity drops to the level of the congestion wave velocity. The critical density is the density corresponding to the peak flow rate and is the dividing point between the free-flow and congested branches. Figure 3 Observation point A is located on a free-flow branch, and its observation density is lower than the critical density, so it is determined to be a free-flow branch. The hysteresis interval is a transition zone around the critical density: when the observation density is significantly lower than the lower limit of the interval, it is determined to be a free-flow branch; when it is significantly higher than the upper limit of the interval, it is determined to be a congested branch. When falling into the interval, the branch is maintained or switched based on the branch, velocity change trend and queuing status of the previous time period, so as to avoid single noise triggering the jump.

[0071] It should be noted that physical uniformity applies constraint correction to the flow, speed, and density triplet output by the meso-level cellular road network, without generating a separate set of parameters that do not correspond to the meso-level simulation; queue length and lane occupancy are directly calculated from the same meso-level loading.

[0072] Furthermore, based on the deviation between the corrected observed state and the simulated state, flow residuals and velocity residuals are constructed for flow rate and velocity, respectively: , formula (4-1); This represents the flow residual at the detection section d at time t; This represents the corrected observed flow rate after being projected onto the same basic graph branch using formula (3); This represents the corrected simulated flow rate after being projected onto the same basic graph branch using formula (3); This represents the lower limit protection value of the flow rate, preventing the residual from being excessively amplified when the denominator is zero or close to zero; max(·) is to take the larger of the corrected observed flow rate and the lower limit protection value as the normalized denominator.

[0073] , formula (4-2); This represents the velocity residual of the detection section d at time t; This indicates the corrected observation rate after being projected onto the same basic graph branch using formula (3); This indicates the corrected simulation speed after being projected onto the same basic graph branch using formula (3); This represents the free-flow velocity at the detection section d at time t; This represents the lower speed limit protection value, which is consistent with the value in formula (2).

[0074] Flow residuals and velocity residuals reflect different demand deviations: when the observed flow rate is consistently higher than the simulated flow rate, it indicates that one or more origin-destination (OD) points, departure times, or path proportions passing through that section are underestimated; when the observed velocity is lower than the simulated velocity on congested branches, it indicates that arrival demand is too low, paths are excessively dispersed, or the effective capacity setting is too high. Both types of residuals are used together within the same objective to prevent only fitting the flow rate without reproducing congestion velocity, or only fitting the velocity while violating the flow conservation of the section.

[0075] In the iterative correction of OD, the OD inversion objective is constructed using the flow residual, velocity residual, and OD prior matrix as prior constraints: Formula (5); where, This represents the OD matrix (i.e., the corrected OD) to be solved in the current control period t. This represents the objective function for OD inversion. The flow residual vector is represented by the values ​​of each section in formula (4-1). Arranged according to cross-section index, is Implicit functions (OD changes, path allocation, meso-level loading, residual changes); The velocity residual vector is derived from the values ​​of each section in formula (4-2). Arranged according to the cross-section index; This is the flow residual weight matrix; Represents the velocity residual weight matrix; This represents the OD prior matrix output in step 11; The formula represents the corrected OD of the previous control period t-1; α represents the prior constraint weight coefficient; and β represents the time smoothing weight coefficient. Both the prior constraint weight coefficient α and the time smoothing weight coefficient β are non-negative. α and β can be determined based on historical playback data or calibration data. The value of α ranges from 0.1 to 10, and the value of β ranges from 0.01 to 5. Initial reference values ​​can be α=1 and β=0.1, respectively. When the detection section coverage is insufficient or the OD identification is low, α can be appropriately increased to enhance the constraint of the OD prior matrix on the inversion results; when the traffic state continuity is strong, β can be appropriately increased to suppress abnormal fluctuations in OD between adjacent control periods. The above values ​​are only examples and can be determined through historical data calibration based on different road types, detection coverage, and control periods.

[0076] The two types of residuals shown in formulas (4-1) and (4-2) above are both incorporated into the OD inversion objective shown in formula (5), while retaining the OD prior bias and time smoothing term. The solution is obtained using projection gradient, sequential quadratic programming, finite difference search, or other numerical methods that satisfy non-negativity constraints.

[0077] Driven by the OD inversion target shown in formula (5), after each OD update, the path allocation and meso-level cellular road network loading in step 12 are re-executed, and the basic calculation is then performed. Figure 1 The corrected flow residual and velocity residual are updated by the same objective function shown in formula (5) until one of the following termination conditions is met: when the relative change of the corrected OD, the change of the flow residual, and the change of the velocity residual are all less than the corresponding preset threshold, the residual is determined to converge and the iteration stops; or when the preset iteration upper limit is reached, the iteration stops. The threshold corresponding to the relative change of OD can be 0.5% to 5%, the threshold corresponding to the change of flow residual can be 0.5% to 5%, and the threshold corresponding to the change of velocity residual can be 0.5% to 5%. When all three are lower than the corresponding threshold, it is considered that the improvement of the correction result by continuing the iteration is lower than the preset accuracy requirement, the residual is determined to converge and the iteration stops.

[0078] For ODs with sufficient detection coverage, the flow residual and velocity residual jointly determine their update. For ODs that pass through only a few detection sections or are highly collinear with other ODs, the prior constraint weight α is increased, single-cycle changes are restricted, or unidentifiable components are frozen. If a detector consistently provides contradictory directions for the two types of residuals, the weight of that section is first reduced, and the sensor, time alignment, basic graph branches, and path connections are checked to prevent the contradiction from being forcibly attributed to OD changes. If the detection coverage is insufficient, the prior constraint weight is increased to avoid large drifts in unobservable ODs. If the inversion does not converge but reaches the real-time computation limit, the current optimal feasible state with a low confidence label is output, and the automatic action range of step 15 is limited.

[0079] After the iteration terminates, the corrected OD is scaled proportionally according to the three component ratios output by the operator's base station positioning record processing, with each component remaining non-negative, and the sum of the three components equal to the corrected OD. Step 13 outputs the corrected OD and the corrected road segment data, wherein the corrected road segment data includes path ratio, path-road segment correlation, corrected road segment traffic flow, road segment speed, road segment density, queue length, and lane occupancy status.

[0080] It should be noted that the sources of each output are distinguished as follows: the corrected OD comes from the physical inversion in this step; the OD components of the three population groups come from the OD prior matrix of step 11, scaled proportionally to the corrected OD; the path ratio and path-segment correlation come from the final step 12; and the segment flow, speed, and density come from the final meso-level state after basic analysis. Figure 1 Consolidation; queue length and lane occupancy are directly statistically derived from the same mesoscopic loading. Optionally, at the end of this step, save the iteration number, initial target, final target, residuals of each section, updated and frozen OD lists, and termination reason.

[0081] In an optional embodiment, in step 14, determining the cause of congestion based on the corrected road segment data includes: Based on the corrected road segment data, a directed propagation chain for queuing is established along the road topology direction; Based on the directed propagation chain, when the queuing duration reaches the minimum confirmation duration and the passage speed of the passage direction section recovers to above the preset threshold, the queuing source section is determined. Based on the arrival demand and current effective traffic capacity of the queuing source section, when the arrival demand continuously exceeds the current effective traffic capacity, the signal timing, lane function, speed limit, ramp release rate and event status corresponding to the queuing source section are read. Based on the difference between the arrival demand and the current effective traffic capacity, as well as the read signal timing, lane function, speed limit, ramp release rate and event status, the cause of congestion is determined; the cause of congestion includes at least one of the following: demand overload, insufficient signal supply, lane configuration mismatch, ramp merging conflict, speed limit incoordination, sudden drop in event capacity, and decrease in traffic capacity.

[0082] In this embodiment, based on the corrected road segment data, the upstream expansion of the queue tail and low-speed zone is tracked according to the road connection direction within multiple consecutive simulation steps or detection cycles. The directed propagation chain is established according to the send-receive relationship of the meso-level cellular road network. For example, when a downstream cell is close to saturation, its receiving capacity limits upstream transmission, and untransmitted vehicles continue to accumulate upstream, a backflow relationship is recorded from the downstream restricted position to the upstream affected position. The relationships in consecutive time steps are superimposed according to road direction and occurrence time to obtain a directed propagation graph extending upstream from the active position and the set of affected road segments. A momentary transmission obstruction does not constitute a propagation edge and must satisfy the minimum duration window.

[0083] Based on the directed propagation chain, when the queuing duration reaches the minimum confirmation time and the passage speed of the cross-section in the release direction recovers to above a preset threshold, the queuing source cross-section is determined. The queuing source cross-section is not the road segment with the lowest current speed, but rather the root position in the directed propagation graph where a persistent supply-demand mismatch first appears, downstream areas show relative recovery after release, and can explain the upstream queue expansion.

[0084] Optionally, candidate queuing source sections should simultaneously meet the following conditions: their downstream areas have relatively unobstructed flow or recovery characteristics after being released; their upstream areas exhibit continuous queuing, sudden drops in speed, or increases in occupancy; and the duration of the state reaches the minimum confirmation window. If multiple candidate queuing source sections appear in an affected area, they are sorted by activation time, upstream queue length, downstream recovery degree, propagation coverage, and observation completeness, and the primary queuing source section and secondary queuing source sections are output; each queuing source section is then entered into the cause and contribution analysis.

[0085] For queuing source sections with continuous flow and velocity observations, cross-verification is performed using shock wave direction. The shock wave velocity between adjacent traffic states 1 and 2 is: Formula (6), where, and To congest traffic on both sides of the boundary, and This corresponds to the density.

[0086] Based on the flow rate, density, and boundary trajectory on both sides of the congestion boundary, the direction and speed of the congestion boundary moving upstream or downstream are estimated. If the spillover propagation root and the shock wave backtracking location are located in the same road segment or adjacent interpretable locations, the confidence level of the queuing source section is increased; if the deviation exceeds the spatial or temporal tolerance, the detectors, base map, topology, OD loading, and event logs are checked, and the result is marked as pending verification. Shock wave backtracking is only used to increase or decrease the confidence level of the queuing source section and does not independently determine the location of the queuing source section. Before verification is completed, only monitoring, alerts, or low-risk actions are allowed; strategies that would significantly alter traffic organization are not automatically issued.

[0087] It should be noted that congestion causes are retrieved on demand. All congestion causes require current demand, effective capacity, and the status of the queuing source section. Only when the public state (arrival demand at the queuing source section, current effective capacity, queuing duration, upstream and downstream speeds, and propagation range) indicates a corresponding probability will the corresponding signal timing, lane function, speed limit, ramp release rate, or event status data be retrieved. Missing dedicated data reduces the identifiability of the cause, and its absence will not be automatically interpreted as the cause not existing.

[0088] After identifying the queuing source section, the cause of congestion is determined based on the difference between the arrival demand and the current effective capacity, as well as the read signal timing, lane function, speed limit, ramp release rate, and event status. Specifically, cause identification follows the order of first using common states, and then retrieving specific data for candidate causes. The common states include the arrival demand at the queuing source section, the current effective capacity, the queuing duration, upstream and downstream speeds, and propagation range; the corresponding signal timing, lane function, speed limit, ramp release rate, or event status data is only read when the common states indicate a corresponding possibility; missing specific data reduces the identifiability of the cause, and its absence is not automatically interpreted as the cause not existing.

[0089] Optionally, Table 1 shows the criteria for determining various congestion causes and the corresponding candidate actions after triggering them. Table 1:

[0090] Optionally, the evidence for each type of cause mentioned above is normalized to the range of 0 to 1 and compared independently with the corresponding cause threshold. It is not mandatory to classify all causes as a single primary cause; multiple causes are allowed to be output simultaneously. If no cause reaches the corresponding threshold, or if the causes conflict with each other and the confidence difference is insufficient, then "unknown" or "pending verification" is output, and high-risk control strategies are not implemented accordingly.

[0091] Furthermore, in step 14 above, the supply-demand gap and the entities that contribute the most to congestion are identified, including: Based on the corrected OD, the arrival demand for entering the queuing source section is determined; Based on the supply parameters and confirmed events of the aforementioned meso-cellular road network, the effective traffic capacity is determined; The supply-demand gap is determined based on the difference between the arrival demand and the effective capacity; and; Based on the corrected OD, for the preset quantity combinations that rank high, keeping other conditions unchanged, the corresponding requirements are reduced and a rehearsal is performed, and the combination whose queue length change direction is consistent with the share ranking direction is taken as the contribution object.

[0092] It should be noted that, based on the corrected OD and the road segments traversed by each candidate path, within the activation time window of the queuing source section, the traffic from each origin and destination point to the queuing source section on each candidate path is broken down item by item to obtain multiple origin-destination point-path combinations and the arrival traffic of each combination; wherein, the activation time window is the time interval corresponding to the queuing duration of the queuing source section satisfying the minimum confirmation time; according to the proportion of the arrival traffic of each combination to the supply-demand gap, the multiple origin-destination point-path combinations are sorted from largest to smallest to obtain the sorting result. Based on the sorting result, a preset number of combinations with the highest sorting can be obtained.

[0093] In this embodiment, based on the corrected OD output in step 13, after final path allocation and meso-level cellular road network loading in step 12, the arrival flow of the corrected OD on each candidate path at the queuing source section is taken as the arrival demand entering the queuing source section. The arrival demand reflects the corrected actual arrival volume, rather than the road design capacity or prior estimate.

[0094] Based on the supply parameters and confirmed events of the meso-level cellular road network, the effective capacity of the queuing source section is determined. The effective capacity is jointly constituted by the basic supply parameters of the meso-level cellular road network, signal timing, lane function, ramp release rate, and confirmed events; the road design capacity is not used to replace the current effective capacity. Confirmed events include states such as accidents, construction, lane closures, or temporary lane occupancy input in step 13 or from an external event platform.

[0095] The supply-demand gap is determined based on the difference between arriving demand and effective capacity. When arriving demand exceeds effective capacity, a positive supply-demand gap is formed; when arriving demand does not exceed effective capacity, the supply-demand gap is not positive, no control strategy is generated, and only continued monitoring, data supplementation, manual review, or low-intensity protection actions are output.

[0096] When the supply-demand gap is positive, the congestion causes identified in step 14 determine the direction of handling the supply-demand gap: demand overload mainly translates into OD volume that needs to be delayed, restricted, or staggered; excessive path concentration mainly translates into traffic that needs to be transferred from high-contribution paths to available alternative paths; supply causes such as signal timing, lane function, or ramp release rate mainly translate into effective traffic capacity that needs to be released within a safe range; sudden drop in event capacity first creates handling and detour requirements, and demand-side measures only serve as temporary protection before the event is resolved.

[0097] For compound causes, the supply-demand gap is allocated into multiple coordinated control measures based on the degree to which each cause exceeds the threshold, the contribution of the contributing object, and the available resources. Capacity compensation does not exceed the physical upper limit of the facility, path transfer does not exceed the remaining capacity of alternative paths, and OD reduction does not exceed the scope of feasible entry control or demand management. If one resource is insufficient, the remaining gap is transferred to other matching actions without amplifying irrelevant actions.

[0098] Furthermore, based on the path-segment correlation in the corrected OD and the simulation results in step 12, within the preset activation time window of the queuing source section determined in step 14 (i.e., the time interval corresponding to the minimum confirmation time of the queuing source section), according to the path ratio determined above, each OD, candidate path, and the three types of crowd components output in step 13 are decomposed item by item in the arrival flow of the queuing source section to form an arrival flow account with OD-path-crowd granularity. Here, the OD-path combination is the flow component of a certain OD in the path allocation in step 12 that arrives at the queuing source section via a certain specific path. Combinations that do not pass through the queuing source section, have zero arrival flow, or only pass outside the preset activation time window are directly excluded.

[0099] For the remaining combinations, calculate the share of each combination's arrival flow relative to the excess arrival demand at the queuing source section. The excess arrival demand is the positive difference between the arrival demand and the effective capacity, i.e., the absolute value of the supply-demand gap determined in step 13. Sort by share from largest to smallest to form a rapid contribution ranking, locating potential major contributors with less computation.

[0100] For the pre-defined combinations of rapid contribution rankings, perform local counterfactual verification. While keeping other OD, events, control boundaries, and supply parameters unchanged, reduce the demand for the corresponding combinations by a pre-defined proportion (the pre-defined proportion is a sufficiently small fixed proportion or adaptive step size to generate observable marginal changes without altering the overall state magnitude). Perform a short-term rehearsal covering the activation and queue propagation of the queuing source section in the meso-cellular road network, calculating the marginal changes in the queue length or oversaturation index of the queuing source section. The queue length is the vehicle queue length statistically recorded at the queuing source section during the final meso-cellular road network loading; the oversaturation index is the excess amount where the ratio of the arriving flow to the effective capacity at the queuing source section exceeds 1.

[0101] For the m-th OD-path combination that has been quickly filtered, after reducing the preset ratio requirement while keeping other conditions unchanged, we obtain the counterfactual evaluation value after removing the m-th item.

[0102] After reducing the preset ratio requirement while keeping other conditions unchanged, we obtain the counterfactual evaluation quantity after removing the m-th term. E represents the baseline counterfactual evaluation quantity, i.e., the counterfactual evaluation quantity obtained from the same rehearsal without reducing any combined demand. The counterfactual evaluation quantity is the queue length or oversaturation index, both of which are the same physical quantity used in step 14 to determine the supply-demand gap. Its normalized marginal contribution is... Formula (7); where, This represents the normalized marginal contribution of the m-th OD-path combination, with a value between 0 and 1. (The remaining text appears to be incomplete and possibly contains errors.) The sum is close to 1, reflecting the relative contribution share of each combination to the congestion at the queuing source section; E represents the baseline counterfactual evaluation quantity, that is, the queuing length or oversaturation index of the queuing source section obtained by performing a short-time simulation in the meso-cellular road network without reducing the demand of any combination. This represents the counterfactual evaluation quantity after removing the m-th item. That is, while keeping the remaining OD, events, control boundaries and supply parameters unchanged, the queue length or oversaturation index of the queue source section obtained by re-executing the short-term simulation is reduced by a preset ratio only after the m-th combined demand is reduced by a preset ratio. This represents the counterfactual evaluation value after removing the j-th term, with the same meaning as... ; This indicates the positive part operation, i.e. =max(x,0), negative values ​​are zero, used to handle the situation where the queue length increases due to alternative path backflow after reducing the demand for a certain combination. In this case, the marginal contribution of the combination is recorded as zero instead of negative. To prevent tiny positive numbers with a denominator of zero.

[0103] If the queue length or oversaturation index at the queue source section shows a marginal decrease in the same direction as the share ranking, the corresponding combination is confirmed as the contributing object. If the change direction is opposite or offset by alternative path return flow, the contribution confidence of the corresponding combination is reduced, and the path reallocation causing the difference is recorded. The alternative path return flow refers to the situation where, after reducing the demand for a certain combination, some vehicles switch to other paths in step 12 path allocation, generating arrival flow again at the queue source section or other sections, thereby offsetting part of the marginal improvement. The contribution confidence is a confidence value in the range of 0 to 1 given by the consistency of the fast share and the counterfactual marginal improvement.

[0104] Optionally, the contribution objects can be output at three levels: OD, OD-path, or OD-path-population. OD-level results aggregate the contributions of all paths under the same OD, suitable for entry flow control or demand staggering in step 15; path-level results (OD-path) retain the distinction of specific paths, suitable for induced diversion in step 15; population-level results (OD-path-population) further distinguish the three population components output in step 13, suitable for differentiated setting of induction intensity and compliance rate in step 15. Each contribution record simultaneously saves the queuing source section it passed through, the preset activation time window, the fast share, the counterfactual marginal improvement, the contribution confidence, and the available alternative paths, for step 15 to select control objects and generate control strategies.

[0105] Optionally, if the supply-demand gap is positive and the cause of congestion has been confirmed, the supply-demand gap is converted into coordinated control quantities corresponding to adjustable resources using a cause-action mapping. The cause-action mapping includes at least: demand overload corresponding to inlet flow restriction, diversion guidance, and demand staggering; insufficient signal supply corresponding to green light ratio, cycle, and phase difference; lane configuration mismatch corresponding to turning lanes, tidal flow lanes, or lane functions; ramp merging conflicts corresponding to ramp release and upstream inlet coordination; speed limit incoordination corresponding to variable speed limits; and sudden drop in event capacity corresponding to event handling, temporary closure, and detour guidance. Each control quantity records its location, controlled object, direction, initial intensity, allowable range, and supporting causal evidence, forming a control anchor point for direct reading and generation of control strategies in step 15.

[0106] If the supply-demand gap is not positive, the confidence level of the queuing source section is insufficient, the cause of congestion is still unknown or needs to be verified, or there is insufficient adjustable resources, then no control strategy will be generated; only continued monitoring, data supplementation, manual verification, or low-intensity protective actions will be output. This avoids automatically expanding control when traffic conditions have recovered or the diagnosis is uncertain.

[0107] Here, the unified output of step 14 includes: the queuing source section and its confidence level, the propagation impact range determined by the directed propagation chain, one or more of the congestion causes and their normalized evidence, the contributing objects and their ranking, the supply-demand gap and the list of adjustable resources. The above output serves as the search anchor point for step 15, and step 15 generates and executes the control strategy based on this anchor point.

[0108] Furthermore, step 15 above includes: The action type is determined based on the cause of the congestion, the target object is determined based on the contributing object, and the search range is determined based on the supply and demand gap. Based on the action type, the target object, and the search range, candidate control strategies are generated within preset safety constraints; The candidate control strategy is input into the meso-level cellular road network to determine the path cost after the action is implemented. The path ratio is recalculated based on the path cost after the action is implemented to determine the meso-level simulation result after iteration. Using the corrected OD and the iterative mesoscale simulation results as initial conditions, the candidate control strategy is pre-simulated in the mesoscale cellular road network to determine the pre-simulation results; A security check is performed on the pre-simulation results. Candidate control strategies that fail the check are eliminated. Candidate control strategies that pass the check are sorted according to management objectives, and the optimal solution is selected. If all candidates are eliminated, the current status is retained. The selected optimal solution is compiled into an execution terminal instruction and issued. After execution, the pre-simulation results are compared with the actual test results. The corresponding parameter group is updated according to the source of the deviation. The corresponding parameter group is frozen when the data is abnormal.

[0109] In this embodiment, based on the congestion causes output in step 14, the action type is determined as follows: overloaded demand corresponds to entrance flow restriction, diversion guidance, and demand staggering; insufficient signal supply corresponds to green light ratio, cycle, and phase difference; lane configuration mismatch corresponds to turning lanes, tidal flow lanes, or lane functions; ramp merging conflict corresponds to ramp release and upstream entrance coordination; speed limit incoordination corresponds to variable speed limits; sudden drop in event capacity corresponds to event handling, temporary closure, and detour guidance. Action types in step 14 that do not reach the cause threshold, are determined to be unknown, or are pending review are closed by default and do not enter the candidate generation.

[0110] Based on the contributing objects output in step 14, the target objects are determined. The contributing objects were already output in step 14 at three levels: OD, OD-path, or OD-path-crowd. This step directly reads the corresponding level as the target object for control. For demand overload, priority is given to induction or inlet flow restriction for ODs with high contribution ranking and alternative paths; for insufficient signal supply, only the phase of the inlet related to the queuing source section is adjusted; for ramp merging conflicts, the search is only conducted within the coordinated control range of the ramp and its upstream sections; for speed limit incoordination, variable speed limiting is only activated if the counterfactual simulation in step 14 has proven to have an improvement effect.

[0111] The search scope is determined based on the supply-demand gap and adjustable resources output in step 14. Step 14 has mapped the supply-demand gap into multiple coordinated control quantities according to cause-action, and recorded the facility physical upper limit, remaining capacity of alternative paths, and entry control or demand management scope corresponding to each control quantity. The search scope is centered on the control quantity in step 14, and the search width is determined by the contribution confidence and resource availability: the search scope is wider when the contribution confidence is high, the target is concentrated, and resource availability is clear; the scope is narrowed when the confidence is low or the target is dispersed.

[0112] Based on the action type, the target object, and the search range, candidate control strategies are generated within preset safety constraints. Each candidate control strategy consists of the following components: OD entry restriction ratio, path-guided transfer ratio, signal green light increment, cycle and phase difference, variable speed limit, ramp release rate, lane organization scheme, and event handling status. The system activates the corresponding subset based on the composite cause set from step 14, without simultaneously optimizing all components each time.

[0113] The preset safety constraints include at least the following: (1) the signal cycle, green ratio, minimum green light, clearing time and phase sequence meet the intersection safety rules; (2) the upper and lower limits of variable speed limits, the difference between adjacent road segments and the rate of change per unit time meet the equipment and road safety rules; (3) the ramp release rate and the entrance flow limit are not lower than the safety lower limit and do not exceed the equipment upper limit; (4) the guided transfer flow does not exceed the remaining capacity of the alternative path, and the total OD demand is non-negative and remains conserved; (5) lane functions and tidal lanes are discrete actions, which are enumerated through predefined safety schemes and are not changed directly by continuous gradients; (6) when the event status is not cleared or the data quality is lower than the threshold, the automatic action amplitude is limited and manual review is performed.

[0114] Continuous components are processed within the real-time computational budget using methods such as bounded grid search, Bayesian optimization, response surface search, or finite difference; discrete components are enumerated according to pre-validated safety schemes. The search primarily focuses on small-scale combinations near the anchor point in step 14. If all combinations are infeasible, the search is gradually expanded within the safety boundary, avoiding direct unconstrained global search. Each candidate control strategy retains its source: corresponding causal evidence, contributing entities, search center, and search boundary, for implementation personnel to trace.

[0115] After the candidate control strategies are written into the meso-cellular road network, signal, flow restriction, ramp, lane, or speed limit actions change the travel time, queuing delay, signal waiting time, and control penalty of each candidate path, forming the path cost after the action is implemented. Based on the candidate path set output in step 12 and the Logit allocation mechanism, the path ratio is recalculated using the path cost after the action is implemented; if necessary, new alternative paths not exceeding the detour limit are added for the affected ODs. The path cost is the value recalculated after the control action is modified by the generalized cost function used in the Logit allocation in step 12, including free travel time, queuing delay, signal waiting time, and control penalty components.

[0116] Different population groups read the cost sensitivity and induced compliance probability output in step 11. Stable commuters only partially migrate under the constraints of fixed arrival time and route habits; non-commuters respond quickly to changes in induced and detour costs; local residents only migrate when short-distance alternative routes are clearly advantageous. The migration volume of each population group is constrained by feasible paths, the remaining capacity of alternative paths, and OD conservation. Traffic that does not comply with the induced flow continues to be allocated at the original cost.

[0117] Based on the recalculated path proportions, an alternating manager-traveler iterative process is performed in the mesoscopic cellular road network: the manager maintains the candidate control strategy unchanged, while the traveler performs rerouting based on the cost sensitivity and compliance probability of each group, repeating this process until the path proportions and road segment states are basically stable, yielding the iterative mesoscopic simulation results. The iterative mesoscopic simulation results include the flow rate, speed, density, queue length, and lane occupancy status of each road segment. If the real-time calculation budget is exhausted, the best feasible solution among the currently validated candidates and the non-converged flag are returned; if traveler rerouting causes overload on alternative paths, the diversion intensity is reduced or a new candidate path is switched before recalculation.

[0118] The management evaluation primarily targets oversaturation at the queuing source section and queuing reduction, while also penalizing secondary congestion in adjacent areas, implementation costs, decreased road network efficiency, and cause-action mismatch. Evaluation weights are configured based on management objectives, road grade, and safety regulations, and remain consistent across all candidate solutions.

[0119] Using the corrected OD and the iterative mesoscopic simulation results as initial conditions, a formal pre-run of the candidate control strategy is executed in the mesoscopic cellular road network. The initial conditions are a snapshot of the mesoscopic state reached after iteration of the corrected OD output in step 13, including the current queue distribution, signal phase time, and supply and demand status of each road segment. The pre-run time domain covers multiple control cycles, encompassing strategy activation, queue response, and possible secondary congestion formation processes. The pre-run maintains consistency between event states, boundary inflows, and equipment availability with the control decision-making process, and writes the candidate control strategy into the mesoscopic cellular road network engine shared with OD inversion, running under the same OD, event, and boundary conditions. The pre-run results include at least: queue length or oversaturation level at the queue source section, total regional delay, speed recovery of critical road segments, alternative path load, and safety agent indicators.

[0120] The pre-simulation results are first subjected to a hard safety check, which checks each item: signal phases do not conflict and the minimum green light and clearing time are not disrupted; lane organization switching meets transition and equipment conditions; the alternative path after diversion does not exceed the remaining capacity; queuing at entrances or ramps does not overflow to critical nodes; adjacent speed limits and rate of change per unit time do not exceed the safety limit; affected lanes are not open when the event is not fully handled. If any item is not met, the candidate control strategy is eliminated, and soft scoring is not used to offset the safety violation. Eliminated candidates return to the step of "generating candidate control strategies within preset safety constraints based on the action type, the target, and the search range" for regeneration.

[0121] Candidate control strategies that pass hard safety verification are then ranked according to management objectives. In addition to queuing at the queuing source section, oversaturation, total regional delay, and speed recovery, the load on alternative routes, peripheral queuing growth, the impact of public transportation or emergency lanes, the magnitude of control actions, and execution costs are also compared. If multiple solutions show similar improvement, the solution with fewer actions, smaller magnitude, clearer causal basis, and lower impact on secondary areas is prioritized. When all candidates are eliminated, the current status is preserved, and specific reasons for failure are output, such as insufficient alternative route capacity, unmet signal constraints, or unresolved event conditions. It is suggested that continued monitoring, supplementary manual guidance, or expansion of the candidate facility scope be provided, but unenforceable strategies are not forcibly assigned.

[0122] The selected optimal solution is compiled into execution terminal instructions and issued. Signal actions are translated into target controllers, phases, cycles, green lights, and effective times; variable speed limits are translated into road segments, directions, speed values, and gradual change processes; traffic guidance and diversion are translated into routes and validity periods via variable message signs, navigation interfaces, or travel information platforms; entrance and ramp control are translated into release rates, flow restriction ratios, or control queue limits; lane and event actions are translated into predefined scheme numbers and manual confirmation requirements. Multiple types of actions are issued collaboratively according to the same effective time and scope; event handling and road safety actions have the highest priority, and signal, ramp, entrance, and guidance actions take effect in stages according to the queue propagation order. If some equipment is offline or instruction confirmation fails, the remaining available actions are re-rehearsed or downgraded, and the remaining actions in the original combination are not directly executed as usual.

[0123] After execution, the measured improvement is calculated from the same cross-sections, road segments, and areas as the pre-simulation using the preset evaluation window, and the actual effective value of the equipment is verified. Feedback between the pre-simulation and actual measurements uses component vectors instead of a single total score. Formula (8); where, This indicates a deviation in the congestion assessment. The improvement value of the congestion evaluation at the queuing source section mentioned in the formal rehearsal is the difference between the queuing length or oversaturation index at the queuing source section before and after the rehearsal. This refers to the measured improvement in congestion assessment calculated using the same cross-section and caliber after implementation. The difference between the two reflects the deviation of the simulation in predicting the congestion reduction effect. (To improve the deviation in speed). This represents the speed improvement value of key road sections during the rehearsal, which is the difference in average speed of key road sections before and after the rehearsal. This represents the measured speed improvement value for the corresponding road segment after implementation. This is due to deviations in the security agent metrics. The safety agent indicators calculated during the simulation include comprehensive measures such as alternative path load, peripheral queue growth, impact of public transportation or emergency lanes, control action range, and execution costs. These are the corresponding security proxy metrics measured after execution. This represents the deviation in the path transfer achievement rate. The predicted path transfer achievement rate in the simulation is the ratio of the actual traffic transferred from the original path to the alternative path to the total traffic of the contributing object in the simulation. The actual path transition success rate after execution is obtained through statistical comparison of traffic flow from navigation interfaces, travel information platforms, or detection sections. This component bias is mainly attributed to traveler response modeling, updates to directional trigger path cost sensitivity, crowd response parameters, and induced compliance probability.

[0124] Here, each component triggers a targeted update independently, without simultaneously modifying all parameters: the first component deviation updates the OD prior or expansion coefficients; the second component deviation updates the basic graph capacity and critical density; the third component deviation updates the safety constraint parameters; and the fourth component deviation updates the path cost sensitivity and induced compliance probability. The cause threshold is only updated when the congestion cause is confirmed as a misjudgment by manual verification or multi-source evidence. Any parameter group is frozen and not updated if data quality is insufficient, the event is not resolved, the device does not execute the command, or there are two consecutive feedback directions that are opposite, to avoid a single abnormal execution damaging the long-term model. Each update saves the old value, new value, evidence, scope of application, and verification results, and can be rolled back to the previous stable version if necessary.

[0125] It should be noted that the method in this application can be applied to control systems. Before deployment, the system completes traffic zone division, road network topology and lane connection modeling, detection section alignment, basic map and initial capacity calibration, candidate path generation rule configuration, cause threshold calibration, and execution device interface checks. After going online, the online process is triggered by the control cycle, with second-level meso-level advancement for sustainable operation, and can also be restored from the previous cycle's state snapshot. Each online cycle sequentially completes data quality checks, OD prior generation, crowd decomposition, path loading, OD inversion, queuing source section and congestion cause diagnosis, contribution object and supply-demand gap calculation, candidate strategy solution, and pre-rehearsal; it only enters the deployment when on-site control is required and safety gates are passed. When there is no congestion, the supply-demand gap is not positive, or the diagnosis is uncertain, the current state is saved and rolled to the next cycle, without mandatory control execution.

[0126] The control cycle can be 5 minutes or configured according to the control center, with path proportions and strategies updated at cycle boundaries; the meso-simulation step can be 1 to 5 seconds, used for vehicle grouping, queuing, and signal status advancement; the number of candidate paths is usually no more than 5, constrained by detour rate and accessibility; the minimum confirmation window can be 2 to 3 signal cycles or an equivalent continuous detection window, used to eliminate false bottlenecks caused by instantaneous fluctuations; the OD convergence threshold is set separately according to relative change and residual change, and iteration stops when all three conditions are met simultaneously; the cause threshold is calibrated by historical annotations or replay samples, and unknown causes are allowed to be output when the threshold is not reached; the number of counterfactual objects is only taken from the top few items in the fast contribution ranking, balancing interpretation accuracy and real-time performance; the upper and lower limits of actions are configured by roads, equipment, and safety regulations, and are not fixed in the algorithm. The above values ​​are examples for ease of implementation and do not limit other settings that can achieve the same technical effect.

[0127] Initial parameter values ​​are primarily derived from road design data, controller configuration, historical test calibrations, and playback experiments. OD, path, supply, and cause diagnosis parameters are set up with separate calibration and validation datasets. After calibration, the applicable road type, time period, and event scope are recorded. Incremental updates after deployment do not exceed safety regulations and equipment physical boundaries; if they exceed the existing applicable scope, manual verification or re-offline calibration is required.

[0128] In one specific embodiment, the morning peak control of a certain city's expressway-ground road composite corridor is taken as an example. The road network is divided into traffic zones, road segments, detection sections, and signal inlets. The control cycle is set to 5 minutes, the simulation step of the meso-cellular road network is set to 2 seconds, and each OD retains no more than 5 candidate paths that meet the restrictions on driving and detours.

[0129] During the 07:30 control cycle, the system reads historical OD data from the most recent cycles, weekday morning peak time stamps, detector traffic speeds, and queuing status from the previous cycle to generate OD priors for several future control cycles. Mobile phone signaling, after base station coverage-traffic cell overlay, stop point deduplication, and cross-day anchor point identification, only provides OD composition ratio and total calibration information, without directly replacing vehicle ODs. The OD prior for a certain corridor increases from 350 vehicles / cycle to 490 vehicles / cycle, with commuter traffic making up the majority. The system allocates routes according to the candidate path generation rules and Logit allocation mechanism in step 12, using a generalized cost including driving, queuing, and signal waiting, with approximately 62% of demand via the main path. After loading the mesoscopic cellular road network, deviations in traffic flow and speed occur upstream of section Y5, prompting step 13 for inversion.

[0130] Preprocessing confirmed that the upstream flow and velocity of Y5 were valid and consistent with the coverage of the simulation section. The initial mesoscopic loading results showed that the simulated flow was lower than the observed flow and the simulated velocity was higher than the observed flow under congested conditions, indicating that the current prior knowledge of the demand passing through Y5 was still underestimated. The system first projects the observed and simulated states to the same congested branch and inverts the relevant OD according to formulas (4-1), (4-2) and (5). After each update, path allocation and mesoscopic cellular network loading are re-executed, and the proportion of the main path and the queue are recalculated until the relative change of OD, the change of flow residual, and the change of velocity residual are all lower than the threshold.

[0131] Around 07:47, the propagation chain showed that the downstream section of Y5 merging exhibited speed recovery characteristics after release, while the upstream queue continued to grow. The shock wave boundary propagated upstream from this location, and was located near the root node of the propagation chain at the same cross section. Y5 was identified as the high-confidence queuing source cross section. Based on separately calibrated and non-mandatorily unified evidence scoring, the demand overload score was 0.38, the signal supply insufficiency score was 0.27, the ramp merging conflict score was 0.22, and the speed limit incoordination score was 0.09. The event evidence was below the threshold. The first three categories of causes exceeded the threshold of this embodiment, forming a composite cause of demand overload, signal supply insufficiency, and ramp merging conflict; the speed limit evidence was insufficient, and variable speed limiting was not activated for this cycle.

[0132] The rapid contribution screening indicates that the arrival flow of OD from Y area to CBD is the largest within the preset activation time window of Y5, with its main path contribution ratio being approximately 0.62. Counterfactual simulation of a small-scale load reduction for this combination simultaneously reduces oversaturation on Y5 and upstream queuing, confirming it as a contributing factor. Supply-demand gap calculations indicate that approximately 10% of effective capacity needs to be released in relevant inlet and merging areas, while approximately 35% of traffic on the critical path needs to be diverted, delayed, or restricted. This value serves as the action search anchor point in step 15, constrained by the remaining capacity of alternative paths and the upper limit of control equipment.

[0133] The strategy generation module only allows three types of actions corresponding to compound causes: reducing the OD main path guidance ratio, increasing the effective green light for Y5 related inlets, and adjusting upstream inlets and ramp releases within the equipment's allowable range. After screening by preset safety constraints, candidate solutions include guiding the main path ratio from 62% to approximately 40%, adding a 10-second effective green light to relevant phases, and setting the inlet flow restriction coefficient around 0.42; the actual values ​​are determined jointly by the remaining capacity of alternative paths and queue overflow constraints. Lane organization and variable speed limits are not included in this round of search due to the lack of corresponding causal evidence.

[0134] For each candidate route, the system first recalculates the route proportions for the three groups of people based on the adjusted travel time, queuing, signal waiting, and control penalties. Commuters only partially accept detours, while non-commuters have a relatively higher transfer rate. After the transfer, traffic is reloaded onto the main and alternative routes, and the system iterates alternately between administrators and travelers until the route proportions and key road segment status are basically stable. If the expected transfer volume exceeds the remaining capacity of the alternative routes, the system automatically reduces the guidance ratio and iterates again.

[0135] All candidate schemes were run within a 30-minute rehearsal window, using the corrected OD, queue, signal phase, and event status at 07:45 as common initial conditions. One aggressive diversion scheme, while reducing Y5 queueing, caused signal overflow at the downstream inlet of the alternative path and was eliminated by hard safety checks; another scheme violated the minimum green light constraint for the relevant phase and also did not enter soft sorting. The final scheme, without triggering secondary queuing or equipment safety limitations, achieved better performance in terms of target queue source section oversaturation, total regional delay, and speed recovery than the maintain-status scheme, and was executed after confirmation by the control platform.

[0136] After the strategy is issued, the signal controller, entrance and ramp equipment, and guidance platform receive the instructions at the unified effective time. The system verifies the actual green light increment, entrance release ratio, and guidance release time of the controller; if any key terminal fails to confirm, the combined strategy will not continue to be executed according to the original parameters, and will return to the pre-rehearsal stage to recalculate based on available equipment.

[0137] After execution, the component vectors were compared between the pre- and actual measurements according to formula (8). In this execution, the direction of queuing and speed improvement at the target queuing source section was consistent with the pre-execution, but the actual path transfer rate of commuters was lower than that in the pre-execution; there was no systematic deviation in the flow-speed relationship and on-site capacity. The system only reduced the path cost sensitivity and induced compliance probability of commuters, while the basic map capacity, critical density, and other population parameters remained unchanged. Updates took effect in small steps in the next control cycle, and the parameters before this execution were retained as the rollback version.

[0138] For multiple concurrent queuing source sections, the system first splits or merges bottleneck groups according to propagation relationships, and then forms cause-contributing object-supply-demand gaps respectively; when the scope of multiple strategies overlaps, constraints are jointly checked in the unified meso-level cellular road network pre-simulation. For cases with no valid mobile signaling, no alternative path, zero OD, or insufficient cause confidence, historical classification ratios, supply-side control only, skipping contribution calculation, or conservative handling of unknown causes are adopted respectively to ensure that the process has a definite degradation path.

[0139] In summary, compared with schemes that separate prediction, diagnosis, and control, the present invention has at least the following beneficial effects: 1) The role and output dimensions of historical OD, temporal characteristics, and road network status are clearly defined; mobile signaling, after spatial mapping, stable anchor point classification, and total calibration, only participates in OD composition estimation, avoiding the direct equation of base station or grid trajectory with vehicle OD. 2) Candidate paths, generalized costs, update frequencies, and mesoscopic simulation steps are all clearly defined, ensuring that corrected OD can be stably mapped to path and segment states, without omitting queuing delays and signal waiting. 3) Flow residuals and velocity residuals simultaneously enter the same OD inversion target, and the feasible region of the basic graph is used for correcting state branches, avoiding overlapping roles of multiple inversion mechanisms and ensuring that the current corrected OD, path ratio, and segment state have traceable sources. 4) Queuing source sections, congestion causes, contributing objects, and supply-demand gaps jointly limit the action type, role location, and amplitude range of candidate strategies; unknown causes are allowed to be output when the cause threshold is not reached, reducing the risk of miscontrol. 5) The same meso-level cellular road network engine is used for both OD state loading and policy pre-simulation. If the safety gating fails, it returns to the policy generation stage. The component deviation after execution is updated according to the source, and the closed loop is controllable and maintainable.

[0140] The various methods of the embodiments of this application have been described above. Apparatus for implementing the above methods will now be provided.

[0141] Please refer to Figure 4 This application also provides a congestion tracing and strategy generation device based on path backpropagation, comprising: The first processing module 41 is used to generate a priori matrix of origin and destination points (OD) corresponding to traffic zones and the prediction time domain, based on historical travel data, time characteristics and the road network status of the previous control cycle in the current control cycle. The second processing module 42 is used to generate candidate paths based on the non-zero ODs in the OD prior matrix, determine the path ratio of the candidate paths according to a preset selection model, determine the traffic of each path based on the path ratio, load the path traffic into the mesoscale cellular road network, and determine the mesoscale simulation results. The third processing module 43 is used to project the observation data of the detection section and the mesoscopic simulation results onto the same macroscopic traffic flow basic map, and iteratively correct the OD with the deviation between observation and simulation and the OD prior matrix as constraints. After each correction, the road segment status is updated and the residual is checked until the residual converges or reaches the preset iteration limit, and the corrected OD and the corrected road segment data are output. The fourth processing module 44 is used to output the causes of congestion, the supply-demand gap, and the object that contributes the most to the congestion based on the corrected OD and the corrected road segment data. The fifth processing module 45 is used to generate candidate control strategies based on the congestion causes, the supply-demand gap, and the contributing objects, perform a pre-run and safety verification of the candidate control strategies in the meso-level cellular road network, and send them to the execution terminal after the verification is passed, and update the corresponding parameter group according to the deviation between the pre-run and the actual measurement.

[0142] Optionally, the apparatus of this application further includes: The storage module is used to save the corrected OD, the path ratio, the corrected road segment data, and the strategy execution record; The sixth processing module is used to replace the corrected OD and the path ratio of the current control cycle with the road network state of the previous control cycle in the next control cycle to generate the OD prior matrix of the next control cycle. The seventh processing module is used to recalibrate using the observation data of the next control cycle if the observation data of the next control cycle conflicts with the corrected OD.

[0143] Optionally, the apparatus of this application further includes: The acquisition module is used to acquire operator base station location records, map the base station coverage area to the traffic cell, perform jitter removal on short-term jumps of the same terminal between adjacent base stations, and retain records exceeding a preset displacement threshold as valid traffic cell transfers. The first determining module is used to identify the stop points for the effective traffic community transfer, and to determine the residential anchor point and the employment anchor point respectively based on a preset time period; The eighth processing module is used to classify the effective traffic zone transfer into commuting, local activities and non-commuting based on the spatial relationship and round-trip periodicity between the residential anchor point and the employment anchor point. The second determining module is used to expand the effective traffic cell transfers by deduplication of devices and travel conversion coefficients to determine the composition ratio of commuting, local activities and non-commuting for each traffic cell pair in the OD prior matrix; the composition ratio is used to determine the path ratio of the candidate path.

[0144] Optionally, the first processing module 41 described above is specifically used for: Establish spatial indexes based on traffic zones, road sections, and detection sections; align historical OD sequences, time characteristics, and road network status of the previous control cycle with a unified timestamp to determine the aligned input data. Based on the historical OD sequence in the aligned input data, the spatial coding result is determined by aggregating adjacent and upstream / downstream traffic zone information according to adjacency and historical travel connections using the traffic zone map. Based on the spatial coding results, the temporal features and the road network status are taken as exogenous variables in chronological order and fused with the spatial coding results to determine the temporal coding results of each traffic zone. Traffic cells with missing or insufficient historical origins in the time-encoding results are assigned low-confidence markers, and the missing values ​​are replaced with preset smoothing values; Based on the time coding results of each traffic cell, a priori matrix of origin and destination points (OD) corresponding to the traffic cell and the prediction time domain is generated.

[0145] Optionally, the second processing module 42 described above is specifically used for: Based on the non-zero ODs in the OD prior matrix, prohibited road sections and paths with unreachable turns are eliminated to generate candidate paths; Based on the corrected road segment data from the previous control cycle, the generalized cost of the candidate path is calculated. Based on the generalized cost, the path ratio of each candidate path is determined according to a preset selection model; Based on the path ratio and the corresponding OD travel volume, the traffic flow for each path is determined; The traffic flow of each path is loaded into the mesoscopic cellular road network to determine the mesoscopic simulation results; the mesoscopic simulation results include at least one of the following: path ratio, path-segment correlation, segment traffic flow, segment speed, segment density, queue length, and lane occupancy status.

[0146] Optionally, the third processing module 43 mentioned above is specifically used for: The observation density is calculated based on the ratio of observed flow rate to observed velocity in the observed data; Based on the critical density and hysteresis interval of the macroscopic traffic flow basic graph, the branch of the basic graph in which the observed density is located is determined; The observed flow rate, observed speed, observed density, and road segment flow rate, road segment speed, and road segment density from the mesoscopic simulation results are projected onto the same basic graph branch to determine the corrected observation state and simulation state. Based on the deviation between the corrected observed state and the simulated state, flow residuals and velocity residuals are constructed. Using the flow residual, the speed residual, and the OD prior matrix as prior constraints, the OD is iteratively corrected to obtain the corrected OD; after each correction, the road segment status is updated based on the corrected OD to obtain the corrected road segment data, and the flow residual and the speed residual are verified. The iteration stops when the relative change of the corrected OD, the change of the flow residual, and the change of the velocity residual are all less than the corresponding preset thresholds, or when the preset iteration limit is reached. Output the corrected OD and the corrected road segment data.

[0147] Optionally, the fourth processing module 44 described above is specifically used for: Based on the corrected road segment data, a directed propagation chain for queuing is established along the road topology direction; Based on the directed propagation chain, when the queuing duration reaches the minimum confirmation duration and the passage speed of the passage direction section recovers to above the preset threshold, the queuing source section is determined. Based on the arrival demand and current effective traffic capacity of the queuing source section, when the arrival demand continuously exceeds the current effective traffic capacity, the signal timing, lane function, speed limit, ramp release rate and event status corresponding to the queuing source section are read. Based on the difference between the arrival demand and the current effective traffic capacity, as well as the read signal timing, lane function, speed limit, ramp release rate and event status, the cause of congestion is determined; the cause of congestion includes at least one of the following: demand overload, insufficient signal supply, lane configuration mismatch, ramp merging conflict, speed limit incoordination, sudden drop in event capacity, and decrease in traffic capacity.

[0148] Optionally, the fourth processing module 44 described above is also specifically used for: Based on the corrected OD, the arrival demand for entering the queuing source section is determined; Based on the supply parameters and confirmed events of the aforementioned meso-cellular road network, the effective traffic capacity is determined; The supply-demand gap is determined based on the difference between the arrival demand and the effective capacity; and; Based on the corrected OD, for the preset quantity combinations that rank high, keeping other conditions unchanged, the corresponding requirements are reduced and a rehearsal is performed, and the combination whose queue length change direction is consistent with the share ranking direction is taken as the contribution object.

[0149] Optionally, the fifth processing module 45 mentioned above is specifically used for: The action type is determined based on the cause of the congestion, the target object is determined based on the contributing object, and the search range is determined based on the supply and demand gap. Based on the action type, the target object, and the search range, candidate control strategies are generated within preset safety constraints; The candidate control strategy is input into the meso-level cellular road network to determine the path cost after the action is implemented. The path ratio is recalculated based on the path cost after the action is implemented to determine the meso-level simulation result after iteration. Using the corrected OD and the iterative mesoscale simulation results as initial conditions, the candidate control strategy is pre-simulated in the mesoscale cellular road network to determine the pre-simulation results; A security check is performed on the pre-simulation results. Candidate control strategies that fail the check are eliminated. Candidate control strategies that pass the check are sorted according to management objectives, and the optimal solution is selected. If all candidates are eliminated, the current status is retained. The selected optimal solution is compiled into an execution terminal instruction and issued. After execution, the pre-simulation results are compared with the actual test results. The corresponding parameter group is updated according to the source of the deviation. The corresponding parameter group is frozen when the data is abnormal.

[0150] It should be noted that the device in this embodiment corresponds to the device applied to the method described above. The implementation methods in each of the above embodiments are also applicable to the embodiments of this device and can achieve the same technical effect. The device provided in this application embodiment can implement all the method steps implemented in the above method embodiments and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiments and the beneficial effects will not be described in detail.

[0151] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described path-backpropagation-based congestion tracing and strategy generation method embodiments, achieving the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0152] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described congestion tracing and strategy generation method embodiments based on path backpropagation, and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0153] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0154] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0155] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for congestion source tracing and strategy generation based on path backpropagation, characterized in that, include: In the current control cycle, based on historical travel data, time characteristics, and the road network status of the previous control cycle, a priori matrix of origin and destination points (OD) corresponding to traffic zones and the prediction time domain is generated. Based on the non-zero OD in the OD prior matrix, candidate paths are generated, the path ratio of the candidate paths is determined according to a preset selection model, the traffic of each path is determined based on the path ratio, and the path traffic is loaded into the mesoscale cellular road network to determine the mesoscale simulation results. The observation data of the detection section and the mesoscopic simulation results are projected onto the same macroscopic traffic flow basic map. The OD is iteratively corrected with the deviation between observation and simulation and the OD prior matrix as constraints. After each correction, the road segment status is updated and the residual is checked until the residual converges or reaches the preset iteration limit. The corrected OD and the corrected road segment data are then output. Based on the corrected OD and the corrected road segment data, determine the causes of congestion, the supply-demand gap, and the objects that contribute the most to congestion. Based on the causes of congestion, the supply-demand gap, and the contributing objects, candidate control strategies are generated. The candidate control strategies are then pre-rendered and safety-verified in the meso-level cellular road network. After the verification is passed, the strategies are sent to the execution terminal, and the corresponding parameter groups are updated according to the deviation between the pre-rendering and the actual measurement.

2. The method according to claim 1, characterized in that, After the current control cycle ends, the method further includes: Save the corrected OD, the path ratio, the corrected road segment data, and the strategy execution record; In the next control cycle, the corrected OD and the path ratio of the current control cycle are used to replace the road network state of the previous control cycle to generate the OD prior matrix of the next control cycle. If the observation data of the next control cycle conflicts with the corrected OD, the observation data of the next control cycle shall be used for recalibration.

3. The method according to claim 1, characterized in that, In the current control cycle, after generating the origin-destination (OD) prior matrix corresponding to traffic zones and the prediction time domain based on historical travel data, time characteristics, and the road network status of the previous control cycle, the method further includes: Obtain operator base station location records, map the base station coverage area to the traffic cell, perform jitter removal on short-term jumps of the same terminal between adjacent base stations, and retain records exceeding a preset displacement threshold as valid traffic cell transfers; For the effective traffic community transfer and stop points, based on a preset time period, residential anchor points and employment anchor points are determined respectively; Based on the spatial relationship and round-trip periodicity between the residential anchor point and the employment anchor point, the effective traffic zone transfer is divided into commuting, local activities and non-commuting. The effective traffic cell transfers are expanded by deduplicating devices and using travel conversion coefficients to determine the composition ratio of commuting, local activities, and non-commuting for each traffic cell pair in the OD prior matrix; the composition ratio is used to determine the path ratio of the candidate paths.

4. The method according to claim 1, characterized in that, In the current control cycle, based on historical travel data, time characteristics, and the road network status of the previous control cycle, a priori origin-destination (OD) matrix corresponding to traffic zones and the prediction time domain is generated, including: Establish spatial indexes based on traffic zones, road sections, and detection sections; align historical OD sequences, time characteristics, and road network status of the previous control cycle with a unified timestamp to determine the aligned input data. Based on the historical OD sequence in the aligned input data, the spatial coding result is determined by aggregating adjacent and upstream / downstream traffic zone information according to adjacency and historical travel connections using the traffic zone map. Based on the spatial coding results, the temporal features and the road network status are taken as exogenous variables in chronological order and fused with the spatial coding results to determine the temporal coding results of each traffic zone. Traffic cells with missing or insufficient historical origins in the time-encoding results are assigned low-confidence markers, and the missing values ​​are replaced with preset smoothing values; Based on the time coding results of each traffic cell, a priori matrix of origin and destination points (OD) corresponding to the traffic cell and the prediction time domain is generated.

5. The method according to claim 3 or 4, characterized in that, Based on the non-zero ODs in the OD prior matrix, candidate paths are generated. The path proportions of these candidate paths are determined according to a preset selection model. The traffic flow for each path is determined based on these path proportions, and the path traffic is loaded into the mesoscopic cellular road network. The mesoscopic simulation results are then determined, including: Based on the non-zero ODs in the OD prior matrix, prohibited road sections and paths with unreachable turns are eliminated to generate candidate paths; Based on the corrected road segment data from the previous control cycle, the generalized cost of the candidate path is calculated. Based on the generalized cost, the path ratio of each candidate path is determined according to a preset selection model; Based on the path ratio and the corresponding OD travel volume, the traffic flow for each path is determined; The traffic flow of each path is loaded into the mesoscopic cellular road network to determine the mesoscopic simulation results; the mesoscopic simulation results include at least one of the following: path ratio, path-segment correlation, segment traffic flow, segment speed, segment density, queue length, and lane occupancy status.

6. The method according to claim 1, characterized in that, The observed data from the detection section and the mesoscopic simulation results are projected onto the same macroscopic traffic flow basic map. The OD is iteratively corrected using the deviation between observation and simulation and the OD prior matrix as constraints. After each correction, the road segment status is updated and the residuals are checked until the residuals converge or reach the preset iteration upper limit. The corrected OD and the corrected road segment data are output, including: The observation density is calculated based on the ratio of observed flow rate to observed velocity in the observed data; Based on the critical density and hysteresis interval of the macroscopic traffic flow basic graph, the branch of the basic graph in which the observed density is located is determined; The observed flow rate, observed speed, observed density, and road segment flow rate, road segment speed, and road segment density from the mesoscopic simulation results are projected onto the same basic graph branch to determine the corrected observation state and simulation state. Based on the deviation between the corrected observed state and the simulated state, flow residuals and velocity residuals are constructed. Using the flow residual, the speed residual, and the OD prior matrix as prior constraints, the OD is iteratively corrected to obtain the corrected OD; after each correction, the road segment status is updated based on the corrected OD to obtain the corrected road segment data, and the flow residual and the speed residual are verified. The iteration stops when the relative change of the corrected OD, the change of the flow residual, and the change of the velocity residual are all less than the corresponding preset thresholds, or when the preset iteration limit is reached. Output the corrected OD and the corrected road segment data.

7. The method according to claim 6, characterized in that, Based on the corrected road segment data, the causes of congestion are determined, including: Based on the corrected road segment data, a directed propagation chain for queuing is established along the road topology direction; Based on the directed propagation chain, when the queuing duration reaches the minimum confirmation duration and the passage speed of the passage direction section recovers to above the preset threshold, the queuing source section is determined. Based on the arrival demand and current effective traffic capacity of the queuing source section, when the arrival demand continuously exceeds the current effective traffic capacity, the signal timing, lane function, speed limit, ramp release rate and event status corresponding to the queuing source section are read. Based on the difference between the arrival demand and the current effective traffic capacity, as well as the read signal timing, lane function, speed limit, ramp release rate and event status, the cause of congestion is determined; the cause of congestion includes at least one of the following: demand overload, insufficient signal supply, lane configuration mismatch, ramp merging conflict, speed limit incoordination, sudden drop in event capacity, and decrease in traffic capacity.

8. The method according to claim 7, characterized in that, Identify the supply-demand gap and the entities that contribute most to congestion, including: Based on the corrected OD, the arrival demand for entering the queuing source section is determined; Based on the supply parameters and confirmed events of the aforementioned meso-cellular road network, the effective traffic capacity is determined; The supply-demand gap is determined based on the difference between the arrival demand and the effective capacity; and, Based on the corrected OD, for the preset quantity combinations that rank high, keeping other conditions unchanged, the corresponding requirements are reduced and a rehearsal is performed, and the combination whose queue length change direction is consistent with the share ranking direction is taken as the contribution object.

9. The method according to claim 1, characterized in that, Based on the congestion causes, the supply-demand gap, and the contributing factors, candidate control strategies are generated. These strategies are then rehearsed and validated within the meso-level cellular road network. Upon successful validation, the strategies are sent to the execution terminal, and the corresponding parameter groups are updated based on the deviation between the rehearsal and the actual measurements. This includes: The action type is determined based on the cause of the congestion, the target object is determined based on the contributing object, and the search range is determined based on the supply and demand gap. Based on the action type, the target object, and the search range, candidate control strategies are generated within preset safety constraints; The candidate control strategy is input into the meso-level cellular road network to determine the path cost after the action is implemented. The path ratio is recalculated based on the path cost after the action is implemented to determine the meso-level simulation result after iteration. Using the corrected OD and the iterative mesoscale simulation results as initial conditions, the candidate control strategy is pre-simulated in the mesoscale cellular road network to determine the pre-simulation results; A security check is performed on the pre-simulation results. Candidate control strategies that fail the check are eliminated. Candidate control strategies that pass the check are sorted according to management objectives, and the optimal solution is selected. If all candidates are eliminated, the current status is retained. The selected optimal solution is compiled into an execution terminal instruction and issued. After execution, the pre-simulation results are compared with the actual test results. The corresponding parameter group is updated according to the source of the deviation. The corresponding parameter group is frozen when the data is abnormal.

10. A congestion tracing and strategy generation device based on path backpropagation, characterized in that, include: The first processing module is used to generate a priori origin-destination (OD) matrix corresponding to traffic zones and the prediction time domain, based on historical travel data, time characteristics, and the road network status of the previous control cycle in the current control cycle. The second processing module is used to generate candidate paths based on the non-zero ODs in the OD prior matrix, determine the path ratio of the candidate paths according to a preset selection model, determine the traffic of each path based on the path ratio, load the path traffic into the mesoscale cellular road network, and determine the mesoscale simulation results. The third processing module is used to project the observation data of the detection section and the mesoscopic simulation results onto the same macroscopic traffic flow basic map, and iteratively correct the OD with the deviation between observation and simulation and the OD prior matrix as constraints. After each correction, the road segment status is updated and the residual is checked until the residual converges or reaches the preset iteration limit, and the corrected OD and the corrected road segment data are output. The fourth processing module is used to output the causes of congestion, the supply-demand gap, and the objects that contribute the most to congestion based on the corrected OD and the corrected road segment data. The fifth processing module is used to generate candidate control strategies based on the congestion causes, the supply-demand gap, and the contributing objects, perform a pre-run and safety verification of the candidate control strategies in the meso-level cellular road network, and send them to the execution terminal after the verification is passed, and update the corresponding parameter group according to the deviation between the pre-run and the actual measurement.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 9.

12. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 9.

Images