A multi-scale traffic operation state prediction method and system

Abstract

This invention relates to the fields of intelligent transportation and road construction management technology, and particularly to a multi-scale traffic operation state prediction method and system. The invention achieves high-precision traffic state prediction by constructing a three-layer spatial topology architecture, designing and building a dual-flow independent LSTM architecture, and establishing a recursive rolling and interactive input mechanism. Under complex conditions such as lane closures, temporary traffic diversions, and perception blind spots, this invention, based on a hierarchical spatial perception system and a dual-flow deep temporal model, can perform high-precision short-term predictions of micro, meso, and macro traffic states, providing decision support for lane-level refined management and control in construction areas.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and road construction management technology, and in particular to a multi-scale traffic operation status prediction method and system. Background Technology

[0002] Currently, existing research on highway traffic condition prediction mainly focuses on macro-level cross-section prediction or deep learning modeling in general road network scenarios. While these studies have achieved some success under normal conditions, they still have the following significant shortcomings when dealing with the special scenario of highway reconstruction and expansion construction: (1) Insufficient sensing coverage, resulting in monitoring blind spots: Current approach: The mainstream method relies on data from ETC gantry or loop detectors at fixed locations.

[0003] Disadvantages: During the reconstruction and expansion construction, due to lane closures, temporary diversions, equipment relocation, and other reasons, a large number of bottleneck road sections with a length of 100 meters are in monitoring blind spots. The existing model has not built a spatial mapping mechanism for non-detection points, making it difficult to achieve dynamic status tracking of the entire construction area, especially micro-road sections, and thus failing to meet the "100-meter level" control requirements.

[0004] (2) Gradient interference exists in multivariate joint training: Existing approaches: Some studies attempt to predict both velocity and flow simultaneously, but these typically employ single-flow shared weights or multi-task learning frameworks.

[0005] Disadvantages: It does not fully consider the significant differences in dimensions and dynamic range among different traffic parameters. During backpropagation, the loss function is easily dominated by high-dimensional variables (such as vehicle speed), resulting in insufficient gradient updates for low-dimensional variables (such as the congestion index), and a significant decrease in the model's sensitivity to identifying the initial stage of congestion.

[0006] (3) Long-term recursive prediction trajectory divergence: Current approach: Existing recursive models (such as standard LSTM and its variants) typically use the predicted value of the previous step as the input of the next step when predicting future states in multiple steps.

[0007] Disadvantages: It lacks explicit modeling of the physical evolution of traffic, such as "low speed inducing congestion and congestion inhibiting vehicle speed". This leads to the continuous accumulation of errors in the recursive process, resulting in significant trajectory drift and instability in long-term predictions (such as the next 60 minutes), and even producing physically impossible prediction results.

[0008] (4) Lack of a hierarchical spatial perception system: Current approaches: Although some studies have used GPS floating car data or ETC data to improve models, most of them are single-scale analyses.

[0009] Disadvantages: A hierarchical spatial perception system of "micro-meta-macro" for reconstruction and expansion construction scenarios has not yet been established, making it difficult to support full-scale decision-making needs from road network scheduling (macro) to lane-level intervention (micro).

[0010] In summary, existing technologies are insufficient to meet the urgent need for accurate traffic condition prediction during the construction period of highway reconstruction and expansion in all four aspects mentioned above.

[0011] Therefore, it is necessary to provide a multi-scale traffic operation state prediction method and system to make high-precision short-term predictions of micro, meso and macro traffic states under complex conditions such as lane closures, temporary diversions, and perception blind spots, based on a hierarchical spatial perception system and a dual-flow deep time series model, so as to provide decision support for lane-level refined management and control in construction areas. Summary of the Invention

[0012] The purpose of this invention is to provide a multi-scale traffic operation state prediction method and system, which, based on a hierarchical spatial perception system and a dual-flow depth time series model, can make high-precision short-term predictions of micro, meso and macro traffic states under complex conditions such as lane closures, temporary diversions, and perception blind spots, providing decision support for lane-level refined management and control in construction areas.

[0013] To address the problems existing in the prior art, this invention provides a multi-scale traffic operation state prediction method, comprising the following steps: S1: Construct a three-layer spatial topology architecture consisting of micro, meso, and macro levels; S2: Process the data and construct a high-dimensional feature tensor. Constructing the high-dimensional feature tensor includes constructing input vectors for traffic state features and time context features. S3: Construct a prediction model based on dual-stream recursive LSTM; S31: The LSTM feature extraction unit uses forget gate, input gate and output gate structure to control the retention and forgetting of information; S32: Construct a dual-stream independent LSTM architecture, including constructing two parallel network streams with the same topology but completely independent weight parameters, namely the congestion prediction stream and the velocity prediction stream; S33: Establish a recursive scrolling and interactive input mechanism; S331: A recursive strategy is used to serialize the single-step prediction model to generate the trajectory for the next 60 minutes; S332: Dual-stream interactive input, the method is as follows: the low vehicle speed predicted at time t is used as a feature input into the congestion model, causing the congestion index to rise at time t+1; the predicted high congestion index is used as a feature input into the speed model, suppressing the recovery of vehicle speed at time t+1.

[0014] Optionally, in the multi-scale traffic operation state prediction method, the three-layer spatial topology architecture of micro, meso, and macro levels is constructed as follows: The continuous road space is discretized into three levels: micro-unit, meso-unit, and macro-unit. The micro-unit is: using the gantry or monitoring section in the construction area as the reference anchor point, the continuous road is discretized into a sequence of road segments with a fixed length of 100m based on the vehicle's driving trajectory; The meso-level unit is a set formed by merging several micro-level road segments with the same functional attributes based on the road's functional attributes. The macro-unit is a collection that aggregates the traffic conditions of all micro-level road segments based on the same highway name and direction of travel.

[0015] Optionally, in the multi-scale traffic operation state prediction method, processing the data and constructing a high-dimensional feature tensor includes the following steps: S21: Data collection: Acquire ETC gantry transaction data and road checkpoint monitoring data near the highway reconstruction and expansion section at a preset frequency; S22: Trajectory Reconstruction and Mapping: Based on the time and speed of the vehicle passing through the gantry, the spatiotemporal trajectory of the vehicle on each 100m micro-road segment is reconstructed using a linear interpolation algorithm, and the average cross-sectional speed and congestion index of the current road segment are calculated. S23: Outlier handling: Remove outlier records with a speed of 0 or exceeding physical limits; use spatiotemporal neighborhood interpolation to fill in missing time slices; S24: Feature Construction: Constructing input vectors for traffic state features and temporal context features; S25: Sample partitioning: Partition the dataset according to the time axis.

[0016] Optionally, in the multi-scale traffic operation state prediction method, LSTM is an abbreviation for Long Short-Term Memory. Building a dual-stream independent LSTM architecture also includes: Shared input: Both prediction streams receive a complete feature vector containing the other's historical state, ensuring that the physical interactions between variables are captured; Independent optimization: Minimize the mean square error of congestion index and vehicle speed separately to avoid training skew caused by differences in dimensions; Network configuration: Each prediction stream contains 2 stacked LSTM layers, 64 hidden units, and a Dropout ratio of 0.2.

[0017] Optionally, in the multi-scale traffic operation state prediction method, a prediction model based on a dual-flow recursive LSTM predicts the future. The iterative steps are as follows, where H is the maximum number of prediction steps: State deduction: Utilizing the current input window, compute the first state in parallel. Congestion index and speed predictions for the bus route: ;in, This is a predicted value for the congestion index. For the sub-model of predicting the congestion index, This is the current input window; ;in, For speed prediction, For the sub-model that predicts average vehicle speed, This is the current input window; Feature vector reconstruction: Construction New feature vector at time step : Among them, traffic state characteristics are represented by predicted values. and ; Embed vectors for temporal context features, based on the actual timestamps. It is generated after vectorization encoding; Sliding window update: Updates the input sequence for the next prediction step; Remove the historical data at the very front of the window and use the newly constructed feature vector. Concat means concatenating the historical feature sequence with the new feature vector at the current moment along the time dimension to achieve dynamic updating of the sliding window; Repeat the state deduction, feature vector reconstruction, and sliding window update steps until... The final output sequence .

[0018] This invention also provides a multi-scale traffic operation status prediction system, comprising: A three-layer spatial topology architecture, comprising micro-units, meso-units, and macro-units; The data module is configured to process data and construct a high-dimensional feature tensor. Constructing the high-dimensional feature tensor includes constructing input vectors for traffic state features and time context features. The prediction model based on dual-stream recursive LSTM includes an LSTM feature extraction unit, a dual-stream independent LSTM architecture, and a recursive scrolling and interactive input mechanism. The LSTM feature extraction unit is configured to control the retention and forgetting of information using forget gate, input gate, and output gate structures; The dual-stream independent LSTM architecture involves constructing two parallel network streams with the same topology but completely independent weight parameters: a congestion prediction stream and a velocity prediction stream. The recursive rolling and interactive input mechanism is configured to use a recursive strategy to serialize the single-step prediction model to generate the trajectory for the next 60 minutes; it is also configured to use dual-stream interactive input, in the following way: the low vehicle speed predicted at time t is used as a feature input into the congestion model to cause the congestion index to rise at time t+1; the predicted high congestion index is used as a feature input into the speed model to suppress the recovery of vehicle speed at time t+1.

[0019] Optionally, in the multi-scale traffic operation status prediction system, the three-layer spatial topology architecture discretizes the continuous road space into three levels: micro-unit, meso-unit, and macro-unit. The micro-unit is: using the gantry or monitoring section in the construction area as the reference anchor point, the continuous road is discretized into a sequence of road segments with a fixed length of 100m based on the vehicle's driving trajectory; The meso-level unit is a set formed by merging several micro-level road segments with the same functional attributes based on the road's functional attributes. The macro-unit is a collection that aggregates the traffic conditions of all micro-level road segments based on the same highway name and direction of travel.

[0020] Optionally, in the multi-scale traffic operation status prediction system, the data module includes the following units: Data acquisition unit: configured to acquire ETC gantry transaction data and road checkpoint monitoring data near the highway reconstruction and expansion section at a preset frequency; Trajectory Reconstruction and Mapping Unit: Configured to reconstruct the spatiotemporal trajectory of the vehicle on each 100m micro-road segment based on the time and speed of the vehicle passing through the gantry, and calculate the average cross-sectional vehicle speed and congestion index of the current road segment. Outlier handling unit: configured to remove outlier records with a speed of 0 or exceeding physical limits; and to complete missing time slices using spatiotemporal neighborhood interpolation. Feature construction unit: configured as an input vector to construct traffic state features and temporal context features; Sample partitioning unit: Configured to partition the dataset in chronological order.

[0021] Optionally, in the multi-scale traffic operation state prediction system, the dual-stream independent LSTM architecture further includes the following units: Shared input unit: configured so that both prediction streams receive a complete feature vector containing the other's historical state, ensuring that physical interactions between variables are captured; Independent optimization unit: configured to minimize the mean square error of congestion index and vehicle speed respectively, avoiding training skew caused by differences in dimensions; Network configuration unit: Configured to contain 2 stacked LSTM layers per prediction stream, 64 hidden units, and a Dropout ratio of 0.2.

[0022] Optionally, in the multi-scale traffic operation state prediction system, the prediction model based on dual-flow recursive LSTM is configured to predict the future. The iteration steps, where H is the maximum number of prediction steps, are as follows: State deduction: Utilizing the current input window, compute the first state in parallel. Congestion index and speed predictions for the bus route: ;in, This is a predicted value for the congestion index. For the sub-model of predicting the congestion index, This is the current input window; ;in, For speed prediction, For the sub-model that predicts average vehicle speed, This is the current input window; Feature vector reconstruction: Construction New feature vector at time step : Among them, traffic state characteristics are represented by predicted values. and ; Embed vectors for temporal context features, based on the actual timestamps. It is generated after vectorization encoding; Sliding window update: Updates the input sequence for the next prediction step; Remove the historical data at the very front of the window and use the newly constructed feature vector. Concat means concatenating the historical feature sequence with the new feature vector at the current moment along the time dimension to achieve dynamic updating of the sliding window; Repeat the state deduction, feature vector reconstruction, and sliding window update steps until... The final output sequence .

[0023] Compared with the prior art, the present invention has the following advantages: (1) Significantly improved prediction accuracy: At the most challenging micro-granularity, the mean absolute error (MAE) of congestion index prediction is reduced by about 35% compared to the single-flow model (Std-LSTM) (from 1.25 to 0.82), significantly enhancing the sensitivity to the initial stage of congestion. In the long-term prediction of the next 60 minutes, the root mean square error (RMSE) of vehicle speed prediction is reduced by 27% (from 9.75 to 7.12).

[0024] (2) Strong multi-scale adaptability: The model outperforms the historical average method (HA), support vector regression (SVR) and standard LSTM model at the micro, meso and macro spatial granularities. The mean absolute percentage error (MAPE) of vehicle speed prediction at the macro granularity is as low as 3.25%, accurately depicting the overall trend of the road network.

[0025] (3) High robustness and physical consistency: It effectively overcomes the trajectory divergence problem of traditional recursive prediction and avoids logically contradictory prediction results such as "high congestion but high vehicle speed".

[0026] (4) Supporting refined management and control: It has achieved full coverage perception of micro road sections at the 100m level, filled the monitoring blind spot in the construction area, and provided reliable data support for lane-level active management and control. Attached Figure Description

[0027] Figure 1 A flowchart for multi-scale traffic operation status prediction provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a dual-stream independent prediction architecture provided in an embodiment of the present invention. Detailed Implementation

[0028] The specific embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. The advantages and features of the present invention will become clearer from the following description. It should be noted that the drawings are all in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0029] In the following, if the methods described herein include a series of steps, the order of these steps presented herein is not necessarily the only order in which these steps can be performed, and some of the steps described may be omitted and / or some other steps not described herein may be added to the method.

[0030] In the existing technology, research on highway traffic condition prediction mainly focuses on macro-level cross-section prediction or deep learning modeling in general road network scenarios. Although it has achieved certain results under normal conditions, it still has many shortcomings when dealing with the special scenario of highway reconstruction and expansion construction period.

[0031] To address the problems existing in the prior art, this invention provides a multi-scale traffic operation state prediction method, such as... Figure 1 As shown, it includes the following steps: S1: Construct a three-layer spatial topology architecture consisting of micro, meso, and macro levels; The following is a method for constructing a three-tiered spatial topology architecture consisting of micro, meso, and macro levels: The continuous road space is discretized into three levels: micro-unit, meso-unit, and macro-unit. (1) The micro-unit is: taking the gantry or monitoring section in the construction area as the reference anchor point, the continuous road is discretized into a sequence of road segments with a fixed length of 100m according to the vehicle driving trajectory; this is the basic physical unit for the model to extract features and predict the state, aiming to accurately capture the position of the queue tail and its dissipation process.

[0032] (2) The meso-level unit is a set of micro-level road segments with the same functional attributes, which are merged according to the road functional attributes (such as the main line basic section, ramp merging area, interchange area, tunnel, etc.); it is used to analyze the local operation status of a specific road type.

[0033] (3) The macro unit is a set of traffic conditions of all micro road segments based on the same highway name and driving direction, which is used to reflect the overall service level of the entire highway in that direction.

[0034] S2: Process the data and construct a high-dimensional feature tensor, including the following steps: S21: Data collection: Acquire ETC gantry transaction data and road checkpoint monitoring data near the highway reconstruction and expansion section at a preset frequency, with the sampling frequency uniformly standardized to 5 minutes; S22: Trajectory Reconstruction and Mapping: Based on the time and speed of the vehicle passing through the gantry, the spatiotemporal trajectory of the vehicle on each 100m micro-road segment is reconstructed using a linear interpolation algorithm, and the average cross-sectional speed and congestion index of the current road segment are calculated. S23: Outlier handling: Remove outlier records with a speed of 0 or exceeding physical limits (e.g., >140 km / h); use spatiotemporal neighborhood interpolation to fill in missing time slices; S24: Feature Construction: Constructing input vectors for traffic state features and temporal context features; ① Traffic state features: Congestion Index (CI), average speed (Speed). ② Temporal context features: Time period coding (morning peak / off-peak / evening peak / night), weekend identification, holiday type.

[0035] S25: Sample Splitting: The dataset is strictly split according to the timeline (70% training set, 15% validation set, and 15% test set) to prevent future information leakage. The historical observation window is the past 50 minutes (10 time steps), and the prediction target is the next 60 minutes (12 time steps).

[0036] S3: Construct a prediction model based on dual-stream recursive LSTM; S31: The LSTM feature extraction unit uses the forget gate, input gate, and output gate structure to control the retention and forgetting of information, effectively capturing the long-range dependencies of traffic flow; S32: As Figure 2 As shown, a dual-stream independent LSTM architecture (i.e., Dual-Stream Recursive LSTM, DSR-LSTM) is constructed, which includes building two parallel network streams with the same topology but completely independent weight parameters, namely the congestion prediction stream and the velocity prediction stream. LSTM is an abbreviation for Long Short-Term Memory. Building a dual-stream independent LSTM architecture also includes: Shared input: Both prediction streams receive a complete feature vector containing the other's historical state, ensuring that the physical interactions between variables are captured; Independent optimization: Minimize the mean square error (MSE) of congestion index and vehicle speed separately to avoid training skew caused by differences in dimensions; Network configuration: Each prediction stream contains 2 stacked LSTM layers, 64 hidden units, and a Dropout ratio of 0.2.

[0037] S33: Establish a recursive scrolling and interactive input mechanism; S331: A recursive strategy is used to serialize the single-step prediction model to generate the trajectory for the next 60 minutes; S332: Dual-stream interactive input method: The low vehicle speed predicted at time t is used as a feature input into the congestion model, causing the congestion index to increase at time t+1; the predicted high congestion index is used as a feature input into the speed model, inhibiting the recovery of vehicle speed at time t+1. Implicit physical constraints are imposed to ensure the logical consistency of long-term predictions.

[0038] In the dual-flow model, let the sub-model for predicting the congestion index be... The sub-model for predicting average vehicle speed is For the start time The historical input sequence is , Let T represent the set of real numbers, where T is the length of the sliding window and D is the feature dimension at each time step. A prediction model based on a two-stream recursive LSTM predicts the future. The iterative steps are as follows, where H is the maximum number of prediction steps: State deduction: Utilizing the current input window, compute the first state in parallel. Congestion index and speed predictions for the bus route: ;in, This is a predicted value for the congestion index. For the sub-model of predicting the congestion index, This is the current input window; ;in, For speed prediction, For the sub-model that predicts average vehicle speed, This is the current input window; Feature vector reconstruction: Construction New feature vector at time step : Among them, traffic state characteristics are represented by predicted values. and ; Embed vectors for temporal context features, based on the actual timestamps. It is generated after vectorization encoding; Sliding window update: Updates the input sequence for the next prediction step; Remove the historical data at the very front of the window and use the newly constructed feature vector. Concat means concatenating the historical feature sequence with the new feature vector at the current moment along the time dimension to achieve dynamic updating of the sliding window; Repeat the state deduction, feature vector reconstruction, and sliding window update steps until... The final output sequence .

[0039] The main objective of this invention is to address the three key technical bottlenecks in traffic condition prediction during highway reconstruction and expansion, providing reliable decision support for lane-level refined management during the construction period: (1) Solving the problem of perception blind spots: Traditional cross-section detection equipment cannot cover bottleneck road sections of hundreds of meters in construction scenarios such as lane closure and temporary diversion, making it difficult to support lane-level fine management and control.

[0040] (2) Solving the model instability caused by multivariate coupling: The dimensions of the congestion index (0-10) and the average vehicle speed (0-120 km / h) are significantly different. The use of a single-stream shared weight structure is prone to gradient interference in the back propagation process, reducing the sensitivity to capturing weak signals such as the congestion start-up stage.

[0041] (3) Solving the problem of lack of physical constraints in long-term prediction: Existing recursive prediction methods cause trajectory divergence due to error accumulation in rolling inference, which cannot guarantee the physical consistency of traffic state evolution in the next 60 minutes.

[0042] In one experiment, (1) Experimental environment and data: ①Data source: A section of a highway reconstruction and expansion project, covering ETC gantry and checkpoint data.

[0043] ② Data scale: Covering a 31-day continuous time series, a total of 5,006 independent 100m micro-road segment units were divided (2,513 for uphill and 2,493 for downhill).

[0044] ③ Dataset splitting: training set (first 21 days, 70%), validation set (middle 5 days, 15%), test set (last 5 days, 15%).

[0045] (2) Model parameter configuration: ① Input / Output: 10-step history window (50 minutes), 12-step prediction step (60 minutes).

[0046] ②Network structure: 2 stacked LSTM layers, 64 hidden units, Dropout 0.2.

[0047] ③ Training strategy: Adam optimizer, initial learning rate 0.001, batch size 1024, maximum epochs 100, early stopping patience value 10.

[0048] (3) Comparison of experimental results: Three statistical indicators widely used in time series forecasting were selected: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE).

[0049] The historical average (HA), support vector regression (SVR), and standard single-stream long short-term memory (Std-LSTM) networks were selected as benchmarks. The test results at the micro-granularity level are shown in Table 1. Table 1. Comparison of model performance at different spatial granularities

[0050] Experimental data fully demonstrate that the method proposed in this invention is significantly superior to existing technologies in terms of accuracy, stability, and multi-scale adaptability, and can meet the high-precision prediction requirements during the construction period of highway reconstruction and expansion.

[0051] Preferably, the present invention also provides a multi-scale traffic operation status prediction system, comprising: 1. Three-layer spatial topology architecture: The three-layer spatial topology architecture discretizes the continuous road space into three levels, namely micro-unit, meso-unit and macro-unit; (1) The micro-unit is: taking the gantry or monitoring section in the construction area as the reference anchor point, the continuous road is discretized into a sequence of road segments with a fixed length of 100m according to the vehicle driving trajectory; this is the basic physical unit for the model to extract features and predict the state, aiming to accurately capture the position of the queue tail and its dissipation process.

[0052] (2) The meso-level unit is a set of micro-level road segments with the same functional attributes, which are merged according to the road functional attributes (such as the main line basic section, ramp merging area, interchange area, tunnel, etc.); it is used to analyze the local operation status of a specific road type.

[0053] (3) The macro unit is a set of traffic conditions of all micro road segments based on the same highway name and driving direction, which is used to reflect the overall service level of the entire highway in that direction.

[0054] 2. Data module, configured to process data and construct high-dimensional feature tensors. Constructing high-dimensional feature tensors includes constructing input vectors for traffic state features and time context features. Specifically, the data module includes the following units: Data acquisition unit: configured to acquire ETC gantry transaction data and road checkpoint monitoring data near the highway reconstruction and expansion section at a preset frequency, with the sampling frequency uniformly standardized to 5 minutes; Trajectory Reconstruction and Mapping Unit: Configured to reconstruct the spatiotemporal trajectory of the vehicle on each 100m micro-road segment based on the time and speed of the vehicle passing through the gantry, and calculate the average cross-sectional vehicle speed and congestion index of the current road segment. Outlier handling unit: configured to remove outlier records with a speed of 0 or exceeding physical limits (e.g., >140 km / h); and to complete missing time slices using spatiotemporal neighborhood interpolation. Feature Construction Unit: Configured as the input vector for constructing traffic state features and temporal context features; ① Traffic State Features: Congestion Index (CI), Average Speed. ② Temporal Context Features: Time Period Code (morning peak / off-peak / evening peak / night), Weekend Identifier, Holiday Type.

[0055] Sample partitioning: The dataset is configured to be partitioned strictly according to the timeline (70% training set, 15% validation set, and 15% test set) to prevent future information leakage. The historical observation window is the past 50 minutes (10 time steps), and the prediction target is the next 60 minutes (12 time steps).

[0056] 3. A prediction model based on dual-stream recursive LSTM, including LSTM feature extraction units, dual-stream independent LSTM architecture, and recursive rolling and interactive input mechanisms; The LSTM feature extraction unit is configured to control the retention and forgetting of information using forget gate, input gate and output gate structure, effectively capturing the long-range dependencies of traffic flow; like Figure 2 As shown, the Dual-Stream Recursive LSTM (DSR-LSTM) architecture involves constructing two parallel network streams with the same topology but completely independent weight parameters, namely the congestion prediction stream and the velocity prediction stream. LSTM is an abbreviation for Long Short-Term Memory. The dual-stream independent LSTM architecture also includes the following units: Shared input unit: configured so that both prediction streams receive a complete feature vector containing the other's historical state, ensuring that physical interactions between variables are captured; Independent optimization unit: configured to minimize the mean square error of congestion index and vehicle speed respectively, avoiding training skew caused by differences in dimensions; Network configuration unit: Configured to contain 2 stacked LSTM layers per prediction stream, 64 hidden units, and a Dropout ratio of 0.2.

[0057] The recursive rolling and interactive input mechanism is configured to use a recursive strategy to serialize the single-step prediction model to generate the trajectory for the next 60 minutes. It is also configured to use dual-stream interactive input, in the following way: the low vehicle speed predicted at time t is used as a feature input into the congestion model to cause the congestion index to rise at time t+1; the predicted high congestion index is used as a feature input into the speed model to suppress the recovery of vehicle speed at time t+1, implicitly imposing physical constraints to ensure the logical consistency of long-term prediction.

[0058] In the dual-flow model, let the sub-model for predicting the congestion index be... The sub-model for predicting average vehicle speed is For the start time The historical input sequence is , Let T represent the set of real numbers, where T is the length of the sliding window and D is the feature dimension at each time step. A prediction model based on a two-stream recursive LSTM is configured to predict the future. The iteration steps, where H is the maximum number of prediction steps, are as follows: State deduction: Utilizing the current input window, compute the first state in parallel. Congestion index and speed predictions for the bus route: ;in, This is a predicted value for the congestion index. For the sub-model of predicting the congestion index, This is the current input window; ;in, For speed prediction, For the sub-model that predicts average vehicle speed, This is the current input window; Feature vector reconstruction: Construction New feature vector at time step : Among them, traffic state characteristics are represented by predicted values. and ; Embed vectors for temporal context features, based on the actual timestamps. It is generated after vectorization encoding; Sliding window update: Updates the input sequence for the next prediction step; Remove the historical data at the very front of the window and use the newly constructed feature vector. Concat means concatenating the historical feature sequence with the new feature vector at the current moment along the time dimension to achieve dynamic updating of the sliding window; Repeat the state deduction, feature vector reconstruction, and sliding window update steps until... The final output sequence .

[0059] Compared with the prior art, the present invention has the following advantages: (1) Significantly improved prediction accuracy: At the most challenging micro-granularity, the mean absolute error (MAE) of congestion index prediction is reduced by about 35% compared to the single-flow model (Std-LSTM) (from 1.25 to 0.82), significantly enhancing the sensitivity to the initial stage of congestion. In the long-term prediction of the next 60 minutes, the root mean square error (RMSE) of vehicle speed prediction is reduced by 27% (from 9.75 to 7.12).

[0060] (2) Strong multi-scale adaptability: The model outperforms the historical average method (HA), support vector regression (SVR) and standard LSTM model at the micro, meso and macro spatial granularities. The mean absolute percentage error (MAPE) of vehicle speed prediction at the macro granularity is as low as 3.25%, accurately depicting the overall trend of the road network.

[0061] (3) High robustness and physical consistency: It effectively overcomes the trajectory divergence problem of traditional recursive prediction and avoids logically contradictory prediction results such as "high congestion but high vehicle speed".

[0062] (4) Supporting refined management and control: It has achieved full coverage perception of micro road sections at the 100m level, filled the monitoring blind spot in the construction area, and provided reliable data support for lane-level active management and control.

[0063] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A multi-scale traffic operation state prediction method, characterized in that, Includes the following steps: S1: Construct a three-layer spatial topology architecture consisting of micro, meso, and macro levels; S2: Process the data and construct a high-dimensional feature tensor. Constructing the high-dimensional feature tensor includes constructing input vectors for traffic state features and time context features. S3: Construct a prediction model based on dual-stream recursive LSTM; S31: The LSTM feature extraction unit uses forget gate, input gate and output gate structure to control the retention and forgetting of information; S32: Construct a dual-stream independent LSTM architecture, including constructing two parallel network streams with the same topology but completely independent weight parameters, namely the congestion prediction stream and the velocity prediction stream; S33: Establish a recursive scrolling and interactive input mechanism; S331: A recursive strategy is used to serialize the single-step prediction model to generate the trajectory for the next 60 minutes; S332: Dual-stream interactive input, the method is as follows: the low vehicle speed predicted at time t is used as a feature input into the congestion model, causing the congestion index to rise at time t+1; the predicted high congestion index is used as a feature input into the speed model, suppressing the recovery of vehicle speed at time t+1.

2. The multi-scale traffic operation state prediction method as described in claim 1, characterized in that, The following is a method for constructing a three-tiered spatial topology architecture consisting of micro, meso, and macro levels: The continuous road space is discretized into three levels: micro-unit, meso-unit, and macro-unit. The micro-unit is: using the gantry or monitoring section in the construction area as the reference anchor point, the continuous road is discretized into a sequence of road segments with a fixed length of 100m based on the vehicle's driving trajectory; The meso-level unit is a set formed by merging several micro-level road segments with the same functional attributes based on the road's functional attributes. The macro-unit is a collection that aggregates the traffic conditions of all micro-level road segments based on the same highway name and direction of travel.

3. The multi-scale traffic operation status prediction method as described in claim 2, characterized in that, Processing data and constructing a high-dimensional feature tensor involves the following steps: S21: Data collection: Acquire ETC gantry transaction data and road checkpoint monitoring data near the highway reconstruction and expansion section at a preset frequency; S22: Trajectory Reconstruction and Mapping: Based on the time and speed of the vehicle passing through the gantry, the spatiotemporal trajectory of the vehicle on each 100m micro-road segment is reconstructed using a linear interpolation algorithm, and the average cross-sectional speed and congestion index of the current road segment are calculated. S23: Outlier handling: Remove outlier records with a speed of 0 or exceeding physical limits; use spatiotemporal neighborhood interpolation to fill in missing time slices; S24: Feature Construction: Constructing input vectors for traffic state features and temporal context features; S25: Sample partitioning: Partition the dataset according to the time axis.

4. The multi-scale traffic operation status prediction method as described in claim 2, characterized in that, LSTM is an abbreviation for Long Short-Term Memory. Building a dual-stream independent LSTM architecture also includes: Shared input: Both prediction streams receive a complete feature vector containing the other's historical state, ensuring that the physical interactions between variables are captured; Independent optimization: Minimize the mean square error of congestion index and vehicle speed separately to avoid training skew caused by differences in dimensions; Network configuration: Each prediction stream contains 2 stacked LSTM layers, 64 hidden units, and a Dropout ratio of 0.

2.

5. The multi-scale traffic operation state prediction method as described in claim 4, characterized in that, Predicting the future based on a two-stream recursive LSTM model The iterative steps are as follows, where H is the maximum number of prediction steps: State deduction: Utilizing the current input window, compute the first state in parallel. Congestion index and speed predictions for the bus route: ;in, This is a predicted value for the congestion index. A sub-model for predicting the congestion index, The current input window; ;in, For speed prediction, For the sub-model that predicts average vehicle speed, The current input window; Feature vector reconstruction: Construction New feature vector at time step : Among them, traffic state characteristics are represented by predicted values. and ; Embed vectors for temporal context features, based on the actual timestamps. It is generated after vectorization encoding; Sliding window update: Updates the input sequence for the next prediction step; Remove the historical data at the very front of the window and use the newly constructed feature vector. Concat means concatenating the historical feature sequence with the new feature vector at the current moment along the time dimension to achieve dynamic updating of the sliding window; Repeat the state deduction, feature vector reconstruction, and sliding window update steps until... The final output sequence .

6. A multi-scale traffic operation status prediction system, characterized in that, include: A three-layer spatial topology architecture, comprising micro-units, meso-units, and macro-units; The data module is configured to process data and construct a high-dimensional feature tensor. Constructing the high-dimensional feature tensor includes constructing input vectors for traffic state features and time context features. The prediction model based on dual-stream recursive LSTM includes an LSTM feature extraction unit, a dual-stream independent LSTM architecture, and a recursive scrolling and interactive input mechanism. The LSTM feature extraction unit is configured to control the retention and forgetting of information using forget gate, input gate, and output gate structures; The dual-stream independent LSTM architecture involves constructing two parallel network streams with the same topology but completely independent weight parameters: a congestion prediction stream and a velocity prediction stream. The recursive rolling and interactive input mechanism is configured to use a recursive strategy to serialize the single-step prediction model to generate the trajectory for the next 60 minutes; it is also configured to use dual-stream interactive input, in the following way: the low vehicle speed predicted at time t is used as a feature input into the congestion model to cause the congestion index to rise at time t+1; the predicted high congestion index is used as a feature input into the speed model to suppress the recovery of vehicle speed at time t+1.

7. The multi-scale traffic operation status prediction system as described in claim 6, characterized in that, The three-layer spatial topology architecture discretizes the continuous road space into three levels: micro-unit, meso-unit, and macro-unit. The micro-unit is: using the gantry or monitoring section in the construction area as the reference anchor point, the continuous road is discretized into a sequence of road segments with a fixed length of 100m based on the vehicle's driving trajectory; The meso-level unit is a set formed by merging several micro-level road segments with the same functional attributes based on the road's functional attributes. The macro-unit is a collection that aggregates the traffic conditions of all micro-level road segments based on the same highway name and direction of travel.

8. The multi-scale traffic operation status prediction system as described in claim 7, characterized in that, The data module includes the following units: Data acquisition unit: configured to acquire ETC gantry transaction data and road checkpoint monitoring data near the highway reconstruction and expansion section at a preset frequency; Trajectory Reconstruction and Mapping Unit: Configured to reconstruct the spatiotemporal trajectory of the vehicle on each 100m micro-road segment based on the time and speed of the vehicle passing through the gantry, and calculate the average cross-sectional vehicle speed and congestion index of the current road segment. Outlier handling unit: configured to remove outlier records with a speed of 0 or exceeding physical limits; and to fill in missing time slices using spatiotemporal neighborhood interpolation. Feature construction unit: configured as an input vector to construct traffic state features and temporal context features; Sample partitioning unit: Configured to partition the dataset in chronological order.

9. The multi-scale traffic operation status prediction system as described in claim 7, characterized in that, The dual-stream independent LSTM architecture also includes the following units: Shared input unit: configured so that both prediction streams receive a complete feature vector containing the other's historical state, ensuring that physical interactions between variables are captured; Independent optimization unit: configured to minimize the mean square error of congestion index and vehicle speed respectively, avoiding training skew caused by differences in dimensions; Network configuration unit: Configured to contain 2 stacked LSTM layers per prediction stream, 64 hidden units, and a Dropout ratio of 0.

2.

10. The multi-scale traffic operation status prediction system as described in claim 9, characterized in that, A prediction model based on a two-stream recursive LSTM is configured to predict the future. The iteration steps, where H is the maximum number of prediction steps, are as follows: State deduction: Utilizing the current input window, compute the first state in parallel. Congestion index and speed predictions for the bus route: ;in, This is a predicted value for the congestion index. A sub-model for predicting the congestion index, The current input window; ;in, For speed prediction, For the sub-model that predicts average vehicle speed, The current input window; Feature vector reconstruction: Construction New feature vector at time step : Among them, traffic state characteristics are represented by predicted values. and ; Embed vectors for temporal context features, based on the actual timestamps. It is generated after vectorization encoding; Sliding window update: Updates the input sequence for the next prediction step; Remove the historical data at the very front of the window and use the newly constructed feature vector. Concat means concatenating the historical feature sequence with the new feature vector at the current moment along the time dimension to achieve dynamic updating of the sliding window; Repeat the state deduction, feature vector reconstruction, and sliding window update steps until... The final output sequence .

Images