A trunk green wave band optimization system based on traffic situation prediction

Abstract

This invention provides a trunk green wave optimization system based on traffic situation prediction, including a dynamic data fusion module that collects real-time vehicle trajectory data through an internet platform and performs data preprocessing to map multi-cycle trajectories to a unified time axis; a traffic situation spatiotemporal identification module that statistically analyzes the collected trajectory data and identifies oversaturation states; a traffic flow spatiotemporal prediction module that outputs predicted values ​​of future traffic flow characteristics based on collected historical traffic flow characteristics, parameters of the real-time signal timing scheme, and static spatiotemporal characteristics; and a trunk green wave optimization module that optimizes the current signal timing scheme based on the predicted values ​​of future traffic flow characteristics, with minimizing the number of oversaturated phases as the primary objective and minimizing total vehicle delay as the secondary objective, and sends the optimization results to the traffic signal controller for execution. Multi-cycle trajectory aggregation increases the amount of effective data; and prioritizing the number of oversaturated phases reduces vehicle delays.

Description

Technical Field

[0001] This invention belongs to the field of traffic signal control, and in particular relates to a trunk green wave optimization system based on traffic situation prediction. Background Technology

[0002] Green wave traffic signal control is an intelligent traffic management strategy that aims to reduce waiting time and improve traffic efficiency by coordinating the timing of traffic lights at consecutive intersections on main roads, ensuring vehicles encounter green lights continuously at a set speed. The implementation of green wave traffic signal control relies on a traffic signal controller, which adjusts the green light sequence at each intersection based on road distance and vehicle speed. By setting appropriate vehicle speeds, the signal control system automatically adjusts the green light sequence at each intersection, ensuring that traffic enjoys continuous green light service during its journey.

[0003] Existing patent CN116665470A discloses a method and device for coordinated control of adjacent intersections based on low-permeability trajectory data. The method includes the following steps: Step S1, acquiring low-permeability trajectory data of two adjacent intersections and calculating vehicle flow rate information; Step S2, constructing a phase difference optimization model, and based on the flow rate information, constructing optimization functions to optimize the phase difference with the objectives of maximizing and minimizing the average vehicle delay of the coordinated phase, respectively, and obtaining the maximum and minimum average vehicle delay by changing the phase difference; Step S3, based on the maximum and minimum average vehicle delay, determining whether coordinated control is required for adjacent intersections, and if so, implementing coordinated control for the two adjacent intersections. This invention determines whether adjacent intersections are suitable for coordinated control from the perspective of the coordinated control results. Summary of the Invention

[0004] Existing traffic signal control systems lack end-to-end intelligent capabilities from data fusion to control execution, exhibiting significant shortcomings, particularly in low-quality data processing, multi-objective collaborative optimization, and dynamic scenario adaptation, resulting in limited improvements in arterial road traffic efficiency. Traditional traffic situation analysis relies primarily on infrastructure such as loop detectors, which suffers from high maintenance costs and limited coverage. While emerging methods based on probe vehicle trajectory data are cost-effective, they struggle to adapt to the low penetration rates of 3%-12% and sampling intervals of 10-30 seconds in real-world scenarios. When probe penetration reaches 10%, the delay prediction error of traditional methods can exceed 40%, rendering signal optimization schemes ineffective. Furthermore, traditional green wave control models are based on fixed periods and historical traffic flow data, failing to integrate real-time traffic conditions and neglecting the impact of fluctuations in green wave design speeds. Additionally, existing optimization algorithms are prone to getting trapped in local optima when processing low-resolution trajectory data, exhibiting slow convergence speeds and failing to meet real-time control requirements.

[0005] Meanwhile, existing technologies have shortcomings in signal control under oversaturated traffic conditions. Current oversaturated phase recognition methods rely on full trajectory data, achieving an accuracy rate of less than 60% at low penetration rates. Furthermore, they do not consider extreme congestion scenarios, with accuracy further dropping below 50% under severe congestion. Regarding traffic flow prediction, traditional time series models cannot capture the spatiotemporal correlation of traffic flow. In arterial traffic scenarios, the prediction error for traffic flow 15 minutes in the future exceeds 25%, making it difficult to support real-time signal optimization.

[0006] To address the aforementioned technical problems, the present invention provides the following technical solution: a trunk green wave optimization system based on traffic situation prediction, comprising a dynamic data fusion module that collects real-time vehicle trajectory data through an internet platform and performs data preprocessing to map multi-cycle trajectories to a unified time axis; a traffic situation spatiotemporal identification module that statistically analyzes the collected trajectory data and identifies oversaturation states; a traffic flow spatiotemporal prediction module that outputs predicted values ​​of future traffic flow characteristics based on collected historical traffic flow characteristics, parameters of the real-time signal timing scheme, and static spatiotemporal characteristics; and a trunk green wave optimization module that optimizes the current signal timing scheme based on the predicted values ​​of future traffic flow characteristics, with the primary objective of minimizing the number of oversaturated phases and the secondary objective of minimizing total vehicle delay, and sends the optimization results to the traffic signal controller for execution.

[0007] Specifically, the dynamic data fusion module obtains trajectory data from ride-hailing and navigation apps through internet platform APIs, including timestamps, latitude and longitude, speed, and heading angle, while simultaneously collecting traffic data from geomagnetic detectors as auxiliary verification.

[0008] Specifically, the dynamic data fusion module performs coordinate transformation and velocity filtering on the collected trajectory data, and extracts key parameters after grouping by phase. The key parameters include the initial departure time ratio, initial delay, and initial period of the trajectory under the current initial signal timing scheme. The initial period is the length of the initial signal period under the current initial signal timing scheme, the initial departure time ratio is the proportion of the trajectory's departure time to the initial period, and the initial delay is the trajectory's delay in the initial signal timing scheme.

[0009] Specifically, the dynamic data fusion module aggregates trajectory data across cycles, maps trajectory data of the same phase in multiple consecutive cycles to a base cycle, forming an aggregated trajectory set, and maps multi-cycle trajectories to a unified time axis according to the principle that when the signal cycle is adjusted from the initial cycle to the new cycle and the green signal ratio is adjusted from the initial value to the new green signal ratio, the difference between the proportion of the trajectory's departure time to the cycle and the proportion of the delay to the cycle remains unchanged.

[0010] Specifically, when the dynamic data fusion module aggregates trajectory data across periods, it performs spatiotemporal alignment of multi-period data based on the trajectory time offset. The time offset of the trajectory in a certain period is equal to the difference between the index of that period and 1, multiplied by the initial period, plus the proportion of the initial departure time of the trajectory multiplied by the initial period.

[0011] Specifically, the traffic situation spatiotemporal recognition module divides the green light duration of a phase into an initial green segment and a final green segment, each segment lasting 10% of the total green light duration. It counts the trajectory of vehicles passing the stop line in the initial and final green segments. If the trajectory count in the final green segment is greater than or equal to the trajectory count in the initial green segment, the phase is determined to be oversaturated; otherwise, it is determined to be non-oversaturated. If the traffic situation spatiotemporal recognition module detects that the number of trajectories in the final green segment is more than 5% greater than that in the initial green segment, it forcibly sets the green signal ratio of that phase to the maximum value.

[0012] Specifically, the traffic flow spatiotemporal prediction module adopts a three-layer bidirectional LSTM spatiotemporal prediction model, which integrates historical traffic flow, real-time signal timing, and spatiotemporal features to predict future traffic flow feature values. The module takes the collected historical traffic flow features, real-time signal timing parameters, and static spatiotemporal features as inputs. The historical traffic flow features are the traffic flow features of the past 1 hour, with 5-minute time steps, including the flow rate, average speed, and queue length of each phase. The real-time signal timing parameters include the current cycle and the green light ratio of each phase. The static spatiotemporal features include the distance to adjacent intersections and the direction weight.

[0013] Specifically, the traffic flow spatiotemporal prediction module outputs predicted values ​​every 5 minutes for the next 30 minutes, including flow rate, delay, and queue length for each phase. The three-layer bidirectional LSTM spatiotemporal prediction model uses a hybrid loss function, which consists of three weighted parts: the first part is the mean squared error, calculated as the average of the squares of the predicted values ​​minus the actual values ​​of all samples; the second part is the mean absolute error, calculated as the average of the absolute values ​​of the predicted values ​​minus the actual values ​​of all samples; and the third part is the mean absolute percentage error, calculated as the average of the absolute values ​​of all samples divided by the average of the actual values. The traffic flow spatiotemporal prediction module fine-tunes the model parameters every 15 minutes using newly collected traffic data.

[0014] Specifically, the trunk green wave optimization module establishes a hierarchical optimization system with minimizing the number of oversaturated phases as the primary objective and minimizing total vehicle delay as the secondary objective. It integrates genetic algorithms and artificial fish swarm algorithms for iterative optimization, and improves convergence speed and global optimality through dynamic parameter adjustment, behavioral coordination, and optimal solution caching mechanisms. At the same time, when the trunk green wave optimization module finally adjusts the signal timing scheme, it calculates the phase difference between intersections based on the distance between adjacent intersections and the green wave design speed to ensure that vehicles pass through green lights continuously when traveling at the design speed.

[0015] Specifically, when the trunk green wave optimization module integrates genetic algorithm and artificial fish swarm algorithm for optimization, it dynamically adjusts the crossover and mutation probabilities of the genetic algorithm based on the population fitness to avoid the algorithm getting trapped in local optima; and sets up a behavior cooperative optimization mechanism for the artificial fish swarm algorithm, including foraging behavior optimization, swarm behavior cooperation, and tail-chasing behavior guidance; the trunk green wave optimization module maintains a global optimal solution cache in real time, recording the current optimal artificial fish state and fitness; after each iteration, if the artificial fish state is better than the optimal solution in the cache, the optimal artificial fish state and fitness are updated, and the continuous non-update counter is reset; when the continuous non-update counter counts 20 consecutive times without updating, a global neighborhood search is triggered: a neighborhood solution space is randomly generated for 20% of the artificial fish to expand the search range.

[0016] The beneficial effects of this invention are as follows: It proposes a multi-cycle trajectory aggregation algorithm based on the principle of proportionality, overcoming the limitation of low probe penetration rate and increasing the effective data volume, significantly improving data utilization and matching accuracy compared to traditional methods; it innovatively designs an oversaturated phase recognition model that does not require full trajectory data, judging the oversaturation state by the difference in trajectory density between the initial and final green segments, significantly improving recognition accuracy in low-penetration scenarios; it constructs a three-layer bidirectional LSTM network integrating 12-dimensional spatiotemporal features, reducing errors in future traffic flow prediction, and periodically adaptively and dynamically updating model parameters to adapt to the time-varying characteristics of traffic flow; it employs a hybrid optimization algorithm combining adaptive genetics and artificial fish swarms, integrating adaptive crossover mutation, fish swarm behavior collaboration, and optimal solution caching mechanisms, improving the convergence speed of trunk line optimization, increasing the probability of obtaining the global optimal solution, and shortening the single optimization time; simultaneously, it for the first time takes the number of oversaturated phases as the primary optimization objective, forming a "capacity-efficiency" dual-optimal control mode, reducing the number of oversaturated phases and lowering vehicle delays; and it integrates real-time prediction data and speed fluctuation feedback for dynamic green wave bandwidth adaptation, improving the green wave bandwidth retention rate. Attached Figure Description

[0017] Figure 1 This is a system structure diagram of the present invention.

[0018] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Example 1: A trunk green wave optimization system based on traffic situation prediction includes a dynamic data fusion module, which collects real-time vehicle trajectory data through an internet platform and performs data preprocessing to map multi-cycle trajectories to a unified time axis; a traffic situation spatiotemporal identification module statistically analyzes the collected trajectory data to identify oversaturation states; a traffic flow spatiotemporal prediction module outputs predicted values ​​of future traffic flow characteristics based on collected historical traffic flow characteristics, parameters of the real-time signal timing scheme, and static spatiotemporal characteristics; and a trunk green wave optimization module optimizes the current signal timing scheme based on the predicted values ​​of future traffic flow characteristics, with the primary objective of minimizing the number of oversaturated phases and the secondary objective of minimizing total vehicle delay, and sends the optimization results to the traffic signal controller for execution.

[0021] The dynamic data fusion module is primarily responsible for data collection and preprocessing. It acquires trajectory data (including timestamps, latitude and longitude, speed, and heading angle) from ride-hailing and navigation apps via internet platform APIs, with sampling intervals of 1-15 seconds. Simultaneously, it collects traffic flow data from geomagnetic detectors as auxiliary verification, achieving data collection coverage of over 80% of urban main roads. The module also performs coordinate transformation (WGS84 to GCJ-02), speed filtering (removing outliers exceeding 150% of the road speed limit), and extracts key parameters after grouping by phase. Key parameters include: Initial departure time ratio: The proportion of the departure time of the trajectory in the initial scheme (i.e., the current signal timing scheme) to the initial cycle (range 0-1). Initial delay: The delay of the trajectory in the initial plan (unit: seconds); Initial period: The length of the initial signal period (unit: seconds).

[0022] Meanwhile, the dynamic data fusion module utilizes the distributed computing framework (Spark) to perform real-time cleaning, denoising, and cross-cycle aggregation of trajectory data. Through the Same Proportion Principle (SRPs), multi-cycle trajectories are mapped to a unified time axis, ultimately achieving a data density increase of over 40%. Cross-cycle aggregation assumes that vehicle arrivals follow an approximately uniform distribution, mapping trajectory data from the same phase across multiple consecutive cycles to a base cycle, forming an aggregated trajectory set. The trajectory time offset is calculated as follows: the time offset of a trajectory in a given cycle equals (the index of that cycle minus 1) multiplied by the initial cycle, plus the proportion of the trajectory's initial departure time multiplied by the initial cycle. This calculation achieves spatiotemporal alignment of multi-cycle data. When the signal cycle is adjusted from the initial cycle to a new cycle, and the green signal ratio is adjusted from the initial value to the new green signal ratio, the difference between the proportion of the trajectory's departure time to the cycle and the proportion of delay to the cycle remains unchanged. Trajectories from different periods (initial period and new period) and with different green light ratios can be aligned to a base period based on the principle of proportionality. Sparse trajectories that were originally scattered across multiple periods are aggregated into a base period through proportional mapping, filling the gaps in the time dimension. The principle of proportionality ensures that the trajectory delay ratio is consistent across different periods through the principle of geometric similarity, effectively compensating for the sparsity of data under low penetration rates. By mapping and aggregating trajectory data using this method, the utilization rate of trajectory data can be increased by more than 40%.

[0023] For low penetration scenarios of 3%-12%, the dynamic data fusion module controls trajectory matching error through a collaborative scheme of "coordinate transformation + velocity filtering + hidden Markov model (HMM) map matching + cross-cycle aggregation".

[0024] The traffic situation spatiotemporal identification module is responsible for judging oversaturation states, mainly through a trajectory counting mechanism and an oversaturation judgment model. The module divides the green light duration (green light ratio multiplied by the cycle) of a certain phase into a green start segment and a green end segment, each segment being 10% of the total green light duration. Oversaturation states are identified by the difference in trajectory distribution density between the two segments. If a trajectory crosses the stop line in the green start segment, it is marked as a green start segment trajectory; otherwise, it is not marked. If a trajectory crosses the stop line in the green end segment, it is marked as a green end segment trajectory; otherwise, it is not marked. Oversaturation states are judged by the difference in trajectory counts between the green end segment and the green start segment. If the number of trajectories in the green end segment is greater than or equal to that in the green start segment, the phase is judged as oversaturated. When the number of trajectories in the green end segment is more than 5% greater than that in the green start segment, the green light ratio of that phase is forcibly set to its maximum value (the ratio of the maximum green light time to the cycle) to ensure maximum traffic capacity under oversaturation conditions. The maximum green light time is determined by road capacity and pedestrian safety requirements.

[0025] The traffic flow spatiotemporal prediction module constructs a three-layer bidirectional LSTM spatiotemporal prediction model, integrating historical traffic flow, real-time timing, and spatiotemporal features to improve prediction accuracy. The three-layer bidirectional LSTM spatiotemporal prediction model uses a three-layer bidirectional LSTM network (128 neurons per layer) and inputs 12-dimensional features, including: Historical traffic flow characteristics over 1 hour: 12 time steps in total, every 5 minutes, including flow rate, average speed, and queue length for each phase; Real-time signal timing parameters: current period, green light ratio of each phase; Static spatiotemporal characteristics: distance between adjacent intersections and directional weights (0.8 for straight-through direction on main roads and 0.2 for branch roads); The output consists of 6-dimensional features, including predicted values ​​for each time step every 5 minutes within the next 30 minutes, specifically including flow rate, delay, and queue length for each phase.

[0026] The model training strategy for the traffic flow spatiotemporal prediction module employs a hybrid loss function, consisting of three weighted components with weights of 0.5, 0.3, and 0.2. The first component is the mean squared error, calculated as the average of the squares of the predicted value minus the actual value for all samples. The second component is the mean absolute error, calculated as the average of the absolute values ​​of the predicted value minus the actual value for all samples. The third component is the mean absolute percentage error, calculated as the average of the absolute value of the predicted value minus the actual value divided by the actual value for all samples.

[0027] Meanwhile, the traffic flow spatiotemporal prediction module incorporates a dynamic update mechanism, fine-tuning the model parameters every 15 minutes using newly collected traffic data to adapt to the time-varying characteristics of traffic flow. Real-world testing shows that the model's traffic flow prediction error for the next 15 minutes is controlled within 12%, which is more than 25% better than the traditional ARIMA model; the error for the next 30 minutes is ≤15%, meeting real-time optimization requirements.

[0028] The trunk line green wave optimization module constructs a hierarchical multi-objective optimization framework, establishing a hierarchical optimization system with "minimizing the number of oversaturated phases" as the primary objective and "minimizing total vehicle delay" as the secondary objective, ensuring priority is given to trunk line traffic capacity, and then optimizing operational efficiency.

[0029] The primary objective of oversaturation control is to minimize the number of oversaturated phases, which is achieved by summing the oversaturation binary variables of all phases marked by the traffic situation spatiotemporal identification module and taking the minimum value. The total number of phases is determined based on the signal structure (e.g., the total number of phases is 8 in a NEMA dual-ring eight-phase system). The secondary objective of delay optimization is to minimize the total delay of all trajectories, which is achieved by summing the delays of all trajectories under all phases and taking the minimum value. The calculation rule for trajectory delay is as follows: if the trajectory is a non-stagnant trajectory (the stagnant binary variable is 0), the delay is 0; if the trajectory is a stagnant trajectory (the stagnant binary variable is 1), the delay is equal to "(new departure time ratio minus (new departure cycle index minus 1) multiplied by (1 - new green signal ratio)) multiplied by the new cycle".

[0030] The trunk green wave optimization module derives the optimal solution using an adaptive genetic-artificial fish swarm hybrid optimization algorithm, thereby optimizing the signal timing scheme. This algorithm combines the global search capability of the genetic algorithm with the local optimization capability of the artificial fish swarm algorithm, improving convergence speed and global optimality through dynamic parameter adjustment, behavioral coordination, and an optimal solution caching mechanism.

[0031] The genetic algorithm employs a dynamic adaptive crossover and mutation strategy, dynamically adjusting the crossover and mutation probabilities based on the population fitness (lower fitness values ​​are better) to avoid getting trapped in local optima. This mechanism, by dynamically adjusting the crossover and mutation probabilities, improves the algorithm's convergence speed by 30% and enhances its global search capability by 25%.

[0032] The artificial fish swarm algorithm employs a behavioral collaborative optimization mechanism, specifically including foraging behavior optimization, swarming behavior coordination, and tail-chasing behavior guidance. Under the foraging behavior optimization mechanism, artificial fish randomly select states within their visual range; if a direction has higher fitness, they move in that direction. This behavior simulates the characteristic of fish moving towards food-rich areas, improving local search capabilities. Under the swarming behavior coordination mechanism, the algorithm calculates the center position of artificial fish within the neighborhood; if the center position has better fitness and is less crowded, the artificial fish moves towards the center. This behavior promotes population aggregation and accelerates the discovery of high-quality solutions. Under the tail-chasing behavior guidance mechanism, the algorithm searches for the optimal artificial fish in the neighborhood; if its fitness is better and it is less crowded, the artificial fish moves towards it. This behavior simulates the characteristic of fish following the optimal individual, improving the algorithm's convergence speed.

[0033] The trunk line green wave optimization module also features an optimal solution cache update mechanism. Under this mechanism, a global optimal solution cache is maintained in real-time during optimization, recording the current state and fitness of the optimal artificial fish. After each iteration, if the artificial fish's state is better than the optimal solution in the cache, the cache is updated, and the consecutive non-updated counter is reset. When there are 20 consecutive non-updated iterations, a global neighborhood search is triggered: a neighborhood solution space is randomly generated for 20% of the artificial fish, expanding the search range. This mechanism increases the probability of the algorithm escaping local optima by 40%.

[0034] Meanwhile, for the NEMA dual-loop eight-phase phase design method commonly used in traffic signal control, this embodiment introduces dual-loop structural constraint processing to ensure the feasibility of the optimized scheme. The main constraints include: Total cycle duration constraint: The sum of the green light ratios of all phases, plus the ratio of the total cycle loss time (valued at 15 seconds) to the cycle, equals 1; Minimum / Maximum Green Light Ratio Constraints: The green light ratio of each phase must be greater than or equal to the ratio of the minimum green light time to the cycle of that phase, and less than or equal to the ratio of the maximum green light time to the cycle of that phase. The minimum green light time must meet the pedestrian safety crossing requirements (usually ≥15 seconds).

[0035] In this embodiment, when the trunk green wave optimization module finally adjusts the signal timing scheme, it also constructs a green wave phase difference optimization model to optimize the phase difference. The green wave phase difference optimization model calculates the phase difference between intersections based on the distance between adjacent intersections and the green wave design speed, ensuring that vehicles continuously pass through green lights while traveling at the design speed. The phase difference between adjacent intersections is equal to the distance between adjacent intersections (in kilometers) multiplied by 3600 (unit conversion factor), then divided by (green wave design speed (in kilometers per hour) multiplied by the signal period (in seconds)). Relying on the green wave phase difference optimization model, combined with a dynamic speed adjustment strategy, bandwidth stability under fluctuating scenarios can be further improved. In actual implementation, when traffic flow fluctuates by ±20%, the green wave bandwidth maintenance rate reaches over 85%.

[0036] Example 2: A trunk green wave optimization system based on traffic situation prediction includes a dynamic data fusion module, which collects real-time vehicle trajectory data through an internet platform and performs data preprocessing to map multi-cycle trajectories to a unified time axis; a traffic situation spatiotemporal identification module statistically analyzes the collected trajectory data to identify oversaturation states; a traffic flow spatiotemporal prediction module outputs predicted values ​​of future traffic flow characteristics based on collected historical traffic flow characteristics, parameters of the real-time signal timing scheme, and static spatiotemporal characteristics; and a trunk green wave optimization module optimizes the current signal timing scheme based on the predicted values ​​of future traffic flow characteristics, with the primary objective of minimizing the number of oversaturated phases and the secondary objective of minimizing total vehicle delay, and sends the optimization results to the traffic signal controller for execution.

[0037] The dynamic data fusion module is primarily responsible for data collection and preprocessing. It acquires trajectory data (including timestamps, latitude and longitude, speed, and heading angle) from ride-hailing and navigation apps via internet platform APIs, with sampling intervals of 1-15 seconds. Simultaneously, it collects traffic flow data from geomagnetic detectors as auxiliary verification, achieving data collection coverage of over 80% of urban main roads. The module also performs coordinate transformation (WGS84 to GCJ-02), speed filtering (removing outliers exceeding 150% of the road speed limit), and extracts key parameters after grouping by phase. Key parameters include: Initial departure time ratio: The proportion of the departure time of the trajectory in the initial scheme (i.e., the current signal timing scheme) to the initial cycle (range 0-1). Initial delay: The delay of the trajectory in the initial plan (unit: seconds); Initial period: The length of the initial signal period (unit: seconds).

[0038] Meanwhile, the dynamic data fusion module utilizes the distributed computing framework (Spark) to perform real-time cleaning, denoising, and cross-cycle aggregation of trajectory data. Through the Same Proportion Principle (SRPs), multi-cycle trajectories are mapped to a unified time axis, ultimately achieving a data density increase of over 40%. Cross-cycle aggregation assumes that vehicle arrivals follow an approximately uniform distribution, mapping trajectory data from the same phase across multiple consecutive cycles to a base cycle, forming an aggregated trajectory set. The trajectory time offset is calculated as follows: the time offset of a trajectory in a given cycle equals (the index of that cycle minus 1) multiplied by the initial cycle, plus the proportion of the trajectory's initial departure time multiplied by the initial cycle. This calculation achieves spatiotemporal alignment of multi-cycle data. When the signal cycle is adjusted from the initial cycle to a new cycle, and the green signal ratio is adjusted from the initial value to the new green signal ratio, the difference between the proportion of the trajectory's departure time to the cycle and the proportion of delay to the cycle remains unchanged. Trajectories from different periods (initial period and new period) and with different green light ratios can be aligned to a base period based on the principle of proportionality. Sparse trajectories that were originally scattered across multiple periods are aggregated into a base period through proportional mapping, filling the gaps in the time dimension. The principle of proportionality ensures that the trajectory delay ratio is consistent across different periods through the principle of geometric similarity, effectively compensating for the sparsity of data under low penetration rates. By mapping and aggregating trajectory data using this method, the utilization rate of trajectory data can be increased by more than 40%.

[0039] For low penetration scenarios of 3%-12%, the dynamic data fusion module controls trajectory matching error through a collaborative scheme of "coordinate transformation + velocity filtering + hidden Markov model (HMM) map matching + cross-cycle aggregation".

[0040] In this embodiment, the dynamic data fusion module synchronizes the trajectory timestamps by calibrating them with an NTP server during trajectory data preprocessing, with an error of < 50ms; it uses a Hidden Markov Model (HMM) to match trajectory points to the road topology to complete map matching, with a matching accuracy of > 95%; during anomaly filtering, it removes abnormal trajectories with speeds > 1.5 times the speed limit and dwell times > 10 minutes, with a filtering rate of approximately 5%.

[0041] In this embodiment, when the dynamic data fusion module performs cross-cycle aggregation, it first divides the cycle into units of a base cycle (default 90 seconds) and maps the trajectory data of 10 consecutive cycles to a unified time axis. Then, it calculates the time offset. For a trajectory in a certain cycle, it calculates the time point in the base cycle by multiplying the index of the cycle by 1 by the base cycle and adding the proportion of the initial departure time of the trajectory by the base cycle. Finally, it aggregates the data, groups it by phase, and counts the number of trajectories, average speed and delay for each time point to generate an aggregated trajectory matrix (dimension: number of phases × time point × trajectory features).

[0042] In this embodiment, when selecting the baseline period, the data aggregation effect was tested under three baseline periods: 60 seconds, 90 seconds, and 120 seconds. The test results showed that the aggregation effect of the 90-second baseline period was the best, with the highest data utilization, matching accuracy, and recognition accuracy. Therefore, 90 seconds was adopted as the default baseline period.

[0043] The traffic situation spatiotemporal identification module is responsible for judging oversaturation states, mainly through a trajectory counting mechanism and an oversaturation judgment model. The module divides the green light duration (green light ratio multiplied by the cycle) of a certain phase into a green start segment and a green end segment, each segment being 10% of the total green light duration. Oversaturation states are identified by the difference in trajectory distribution density between the two segments. If a trajectory crosses the stop line in the green start segment, it is marked as a green start segment trajectory; otherwise, it is not marked. If a trajectory crosses the stop line in the green end segment, it is marked as a green end segment trajectory; otherwise, it is not marked. Oversaturation states are judged by the difference in trajectory counts between the green end segment and the green start segment. If the number of trajectories in the green end segment is greater than or equal to that in the green start segment, the phase is judged as oversaturated. When the number of trajectories in the green end segment is more than 5% greater than that in the green start segment, the green light ratio of that phase is forcibly set to its maximum value (the ratio of the maximum green light time to the cycle) to ensure maximum traffic capacity under oversaturation conditions. The maximum green light time is determined by road capacity and pedestrian safety requirements.

[0044] The traffic flow spatiotemporal prediction module constructs a three-layer bidirectional LSTM spatiotemporal prediction model, integrating historical traffic flow, real-time timing, and spatiotemporal features to improve prediction accuracy. The three-layer bidirectional LSTM spatiotemporal prediction model uses a three-layer bidirectional LSTM network (128 neurons per layer) and inputs 12-dimensional features, including: Historical traffic flow characteristics over 1 hour: 12 time steps in total, every 5 minutes, including flow rate, average speed, and queue length for each phase; Real-time signal timing parameters: current period, green light ratio of each phase; Static spatiotemporal characteristics: distance between adjacent intersections and directional weights (0.8 for straight-through direction on main roads and 0.2 for branch roads); The output consists of 6-dimensional features, including predicted values ​​for each time step every 5 minutes within the next 30 minutes, specifically including flow rate, delay, and queue length for each phase.

[0045] The model training strategy for the traffic flow spatiotemporal prediction module employs a hybrid loss function, consisting of three weighted components with weights of 0.5, 0.3, and 0.2. The first component is the mean squared error, calculated as the average of the squares of the predicted value minus the actual value for all samples. The second component is the mean absolute error, calculated as the average of the absolute values ​​of the predicted value minus the actual value for all samples. The third component is the mean absolute percentage error, calculated as the average of the absolute value of the predicted value minus the actual value divided by the actual value for all samples.

[0046] Meanwhile, the traffic flow spatiotemporal prediction module incorporates a dynamic update mechanism, fine-tuning the model parameters every 15 minutes using newly collected traffic data to adapt to the time-varying characteristics of traffic flow. Real-world testing shows that the model's traffic flow prediction error for the next 15 minutes is controlled within 12%, which is more than 25% better than the traditional ARIMA model; the error for the next 30 minutes is ≤15%, meeting real-time optimization requirements.

[0047] In this embodiment, when the LSTM prediction model performs prediction, it first acquires data by obtaining the input features (12 dimensions) from the data processing layer for the past hour every 5 minutes. Then, it performs feature standardization using Z-Score standardization, calculated as "(original feature value minus feature mean) divided by feature standard deviation". Finally, it performs prediction by inputting the standardized features into the pre-trained LSTM model and outputting the predicted values ​​(6 dimensions) for the next 30 minutes. After outputting the results, it performs inverse standardization by using the feature mean and standard deviation used during standardization to restore the predicted values ​​to the actual physical quantities. Kalman filtering is used to calibrate the error, and the error after calibration is <10%.

[0048] The trunk line green wave optimization module constructs a hierarchical multi-objective optimization framework, establishing a hierarchical optimization system with "minimizing the number of oversaturated phases" as the primary objective and "minimizing total vehicle delay" as the secondary objective, ensuring priority is given to trunk line traffic capacity, and then optimizing operational efficiency.

[0049] The primary objective of oversaturation control is to minimize the number of oversaturated phases, which is achieved by summing the oversaturation binary variables of all phases marked by the traffic situation spatiotemporal identification module and taking the minimum value. The total number of phases is determined based on the signal structure (e.g., the total number of phases is 8 in a NEMA dual-ring eight-phase system). The secondary objective of delay optimization is to minimize the total delay of all trajectories, which is achieved by summing the delays of all trajectories under all phases and taking the minimum value. The calculation rule for trajectory delay is as follows: if the trajectory is a non-stagnant trajectory (the stagnant binary variable is 0), the delay is 0; if the trajectory is a stagnant trajectory (the stagnant binary variable is 1), the delay is equal to "(new departure time ratio minus (new departure cycle index minus 1) multiplied by (1 - new green signal ratio)) multiplied by the new cycle".

[0050] The trunk green wave optimization module derives the optimal solution using an adaptive genetic-artificial fish swarm hybrid optimization algorithm, thereby optimizing the signal timing scheme. This algorithm combines the global search capability of the genetic algorithm with the local optimization capability of the artificial fish swarm algorithm, improving convergence speed and global optimality through dynamic parameter adjustment, behavioral coordination, and an optimal solution caching mechanism.

[0051] The genetic algorithm employs a dynamic adaptive crossover and mutation strategy, dynamically adjusting the crossover and mutation probabilities based on the population fitness (lower fitness values ​​are better) to avoid getting trapped in local optima. This mechanism, by dynamically adjusting the crossover and mutation probabilities, improves the algorithm's convergence speed by 30% and enhances its global search capability by 25%.

[0052] The artificial fish swarm algorithm employs a behavioral collaborative optimization mechanism, specifically including foraging behavior optimization, swarming behavior coordination, and tail-chasing behavior guidance. Under the foraging behavior optimization mechanism, artificial fish randomly select states within their visual range; if a direction has higher fitness, they move in that direction. This behavior simulates the characteristic of fish moving towards food-rich areas, improving local search capabilities. Under the swarming behavior coordination mechanism, the algorithm calculates the center position of artificial fish within the neighborhood; if the center position has better fitness and is less crowded, the artificial fish moves towards the center. This behavior promotes population aggregation and accelerates the discovery of high-quality solutions. Under the tail-chasing behavior guidance mechanism, the algorithm searches for the optimal artificial fish in the neighborhood; if its fitness is better and it is less crowded, the artificial fish moves towards it. This behavior simulates the characteristic of fish following the optimal individual, improving the algorithm's convergence speed.

[0053] The trunk line green wave optimization module also features an optimal solution cache update mechanism. Under this mechanism, a global optimal solution cache is maintained in real-time during optimization, recording the current state and fitness of the optimal artificial fish. After each iteration, if the artificial fish's state is better than the optimal solution in the cache, the cache is updated, and the consecutive non-updated counter is reset. When there are 20 consecutive non-updated iterations, a global neighborhood search is triggered: a neighborhood solution space is randomly generated for 20% of the artificial fish, expanding the search range. This mechanism increases the probability of the algorithm escaping local optima by 40%.

[0054] In this embodiment, when the trunk green wave band optimization module optimizes using the adaptive genetic-artificial fish swarm hybrid algorithm, it first initializes the population by setting the population size to 50 and encoding the artificial fish status as "[period, green ratio of the first phase, green ratio of the second phase, ..., green ratio of the eighth phase]", where the period ranges from 40 to 150 seconds and the green ratio of each phase ranges from 0.1 to 0.7. Next, iterative optimization is performed, with each artificial fish sequentially performing foraging, swarming, and tail-chasing behaviors. Dynamic adaptive crossover mutation is triggered every 50 iterations. The final termination condition is reaching the maximum number of iterations (200) or failing to update the optimal solution for 20 consecutive iterations. After termination, a scheme is generated, selecting the scheme with the fewest oversaturated phases and the lowest total delay, ensuring a period error ≤ 5 seconds and a green ratio error ≤ 0.05.

[0055] Meanwhile, for the NEMA dual-loop eight-phase phase design method commonly used in traffic signal control, this embodiment introduces dual-loop structural constraint processing to ensure the feasibility of the optimized scheme. The main constraints include: Total cycle duration constraint: The sum of the green light ratios of all phases, plus the ratio of the total cycle loss time (valued at 15 seconds) to the cycle, equals 1; Minimum / maximum green light ratio constraint: The green light ratio of each phase must be greater than or equal to the ratio of the minimum green light time to the cycle of that phase, and less than or equal to the ratio of the maximum green light time to the cycle of that phase. The minimum green light time must meet the pedestrian safety crossing requirements (usually ≥15 seconds).

[0056] In this embodiment, when the trunk green wave optimization module finally adjusts the signal timing scheme, it also constructs a green wave phase difference optimization model to optimize the phase difference. The green wave phase difference optimization model calculates the phase difference between intersections based on the distance between adjacent intersections and the green wave design speed, ensuring that vehicles continuously pass through green lights while traveling at the design speed. The phase difference between adjacent intersections is equal to the distance between adjacent intersections (in kilometers) multiplied by 3600 (unit conversion factor), then divided by (green wave design speed (in kilometers per hour) multiplied by the signal period (in seconds)). Relying on the green wave phase difference optimization model, combined with a dynamic speed adjustment strategy, bandwidth stability under fluctuating scenarios can be further improved. In actual implementation, when traffic flow fluctuates by ±20%, the green wave bandwidth maintenance rate reaches over 85%.

[0057] In this embodiment, when the trunk green wave optimization module sends the new signal timing scheme to the signal controller for signal timing scheme conversion, it first performs cycle conversion, converting the optimized cycle into a format acceptable to the signal controller (accuracy of 1 second); then it performs green ratio conversion, calculating the green light time for each phase by multiplying the green ratio by the cycle (accuracy of 1 second), ensuring that the minimum green light time constraint is met (e.g., pedestrian phase ≥ 15 seconds); finally, it performs phase difference calculation, obtaining the phase difference between adjacent intersections according to the calculation rules of the green wave phase difference optimization model, and converting it into controller offset (accuracy of 1 second). When the specific scheme is issued and feedback is performed, a protocol conversion is first performed, and the timing scheme is issued to the signal controller via the NTCIP protocol, with a response time < 100ms; finally, feedback is executed, with the controller transmitting the actual timing parameters in real time for comparison with the optimized scheme. If the error exceeds 5%, re-optimization is triggered. In specific applications, dynamic adjustment of the green wave design speed was also tested, and the bandwidth retention rate under different green wave speed fluctuations was tested. Test results show that the dynamic speed adjustment threshold for green wave design is ±10km / h. When the actual speed fluctuation exceeds this threshold, the phase difference is recalculated and the timing scheme is updated to ensure that the bandwidth retention rate is ≥80%.

[0058] This embodiment proposes a multi-cycle trajectory aggregation algorithm based on the principle of proportionality, which breaks through the limitation of low probe penetration rate (3%-12%), increases the effective data volume by more than 40%, and controls the trajectory matching error within 8% under low penetration rate. Compared with traditional methods, the data utilization rate and matching accuracy are significantly improved. In terms of data fusion and low-quality data processing innovation, this embodiment innovatively designs an oversaturated phase recognition model that does not require the full range of trajectories by using sampling trajectory density oversaturation recognition technology. It judges the oversaturation state by the difference in trajectory density between the initial and final green segments. The recognition accuracy reaches 92% at a penetration rate of 10% (over 88% in extreme congestion scenarios), which is 32 percentage points higher than the SCATS system.

[0059] In terms of prediction and optimization algorithm innovation, the LSTM spatiotemporal prediction model used in this embodiment constructs a three-layer bidirectional LSTM network that integrates 12-dimensional spatiotemporal features. The traffic flow prediction error in the next 15 minutes is ≤12%, which is 25% better than the traditional ARIMA model. The model parameters are dynamically updated every 15 minutes to adapt to the time-varying characteristics of traffic flow. This embodiment employs an adaptive genetic-artificial fish swarm hybrid optimization algorithm, which integrates adaptive crossover mutation, fish swarm behavior collaboration, and optimal solution caching mechanism. This improves the convergence speed of trunk line optimization by 30%, increases the probability of obtaining the global optimal solution by 25%, and reduces the single optimization time from 30 minutes to less than 5 minutes.

[0060] In terms of innovation in trunk line green wave control technology, this embodiment adopts a hierarchical multi-objective optimization framework, and for the first time takes the number of oversaturated phases as the primary optimization objective, forming a dual-optimal control mode of "capacity-efficiency". Actual measurements show that the number of oversaturated phases is reduced by 50% and vehicle delays are reduced by 45%. At the same time, dynamic green wave bandwidth adaptive technology is adopted, which integrates real-time prediction data and speed fluctuation feedback. When traffic flow fluctuates by ±20% and speed fluctuates by ±10km / h, the green wave bandwidth maintenance rate is ≥80%, which is 35 percentage points higher than the traditional Maxband method.

Claims

1. A trunk road green wave optimization system based on traffic situation prediction, characterized in that, It includes a dynamic data fusion module that collects real-time vehicle trajectory data through an internet platform and performs data preprocessing to map multi-cycle trajectories to a unified time axis; a traffic situation spatiotemporal identification module that statistically analyzes the collected trajectory data and identifies oversaturation states; and a traffic flow spatiotemporal prediction module that outputs predicted values ​​of future traffic flow characteristics based on collected historical traffic flow characteristics, parameters of real-time signal timing schemes, and static spatiotemporal characteristics. The trunk green wave optimization module optimizes the current signal timing scheme based on the predicted values ​​of future traffic flow characteristics, with the primary objective of minimizing the number of oversaturated phases and the secondary objective of minimizing total vehicle delays, and then sends the optimized signal timing scheme to the traffic signal controller for execution.

2. The trunk green wave optimization system based on traffic situation prediction according to claim 1, characterized in that, The dynamic data fusion module obtains trajectory data from ride-hailing and navigation apps via internet platform APIs, including timestamps, latitude and longitude, speed, and heading angle, while simultaneously collecting traffic data from geomagnetic detectors as auxiliary verification.

3. The trunk green wave optimization system based on traffic situation prediction according to claim 1 or 2, characterized in that, The dynamic data fusion module performs coordinate transformation and velocity filtering on the collected trajectory data, and extracts key parameters after grouping by phase. The key parameters include the initial departure time ratio, initial delay, and initial period of the trajectory under the current initial signal timing scheme. The initial period is the length of the initial signal period under the current initial signal timing scheme, the initial departure time ratio is the proportion of the trajectory's departure time to the initial period, and the initial delay is the trajectory's delay in the initial signal timing scheme.

4. The trunk green wave optimization system based on traffic situation prediction according to claim 3, characterized in that, The dynamic data fusion module aggregates trajectory data across cycles, maps trajectory data of the same phase in multiple consecutive cycles to a reference cycle, forming an aggregated trajectory set. It also maps multi-cycle trajectories to a unified time axis according to the principle that when the signal cycle is adjusted from the initial cycle to the new cycle and the green signal ratio is adjusted from the initial value to the new green signal ratio, the difference between the departure time of the trajectory and the delay of the trajectory remains unchanged.

5. The trunk green wave optimization system based on traffic situation prediction according to claim 4, characterized in that, When the dynamic data fusion module aggregates trajectory data across periods, it performs spatiotemporal alignment of multi-period data based on the trajectory time offset. The time offset of the trajectory in a certain period is equal to the difference between the index of that period and 1, multiplied by the initial period, plus the proportion of the initial departure time of the trajectory multiplied by the initial period.

6. The trunk green wave optimization system based on traffic situation prediction according to claim 1, characterized in that, The traffic situation spatiotemporal recognition module divides the green light duration of a phase into an initial green segment and a final green segment, each segment being 10% of the total green light duration. It counts the trajectory of vehicles passing the stop line in the initial and final green segments. If the trajectory count in the final green segment is greater than or equal to the trajectory count in the initial green segment, the phase is determined to be oversaturated; otherwise, it is determined to be non-oversaturated. If the traffic situation spatiotemporal recognition module detects that the number of trajectories in the final green segment is more than 5% greater than that in the initial green segment, it forcibly sets the green light ratio of that phase to the maximum value.

7. The trunk green wave optimization system based on traffic situation prediction according to claim 1, characterized in that, The traffic flow spatiotemporal prediction module adopts a three-layer bidirectional LSTM spatiotemporal prediction model, which integrates historical traffic flow, real-time signal timing and spatiotemporal features to predict the feature values ​​of future traffic flow. The collected historical traffic flow features, parameters of the real-time signal timing scheme and static spatiotemporal features are used as inputs. The historical traffic flow features are the traffic flow features of the past 1 hour, with 1 time step every 5 minutes, including the flow rate, average speed and queue length of each phase. Real-time signal timing parameters include the current cycle and the green light ratio of each phase; static spatiotemporal characteristics include the distance between adjacent intersections and the direction weight.

8. The trunk green wave optimization system based on traffic situation prediction according to claim 7, characterized in that, The traffic flow spatiotemporal prediction module outputs predicted values ​​every 5 minutes for the next 30 minutes, including traffic flow, delay, and queue length for each phase. The three-layer bidirectional LSTM spatiotemporal prediction model employs a hybrid loss function, consisting of three weighted components: the first component is the mean squared error, calculated as the average of the squares of the predicted values ​​minus the actual values ​​for all samples; the second component is the mean absolute error, calculated as the average of the absolute values ​​of the predicted values ​​minus the actual values ​​for all samples; and the third component is the mean absolute percentage error, calculated as the average of the absolute values ​​of all samples divided by the actual values. The traffic flow spatiotemporal prediction module fine-tunes the model parameters every 15 minutes using newly collected traffic data.

9. The trunk green wave optimization system based on traffic situation prediction according to claim 1, characterized in that, The trunk green wave optimization module establishes a hierarchical optimization system with minimizing the number of oversaturated phases as the primary objective and minimizing total vehicle delay as the secondary objective. It integrates genetic algorithms and artificial fish swarm algorithms for iterative optimization, and improves convergence speed and global optimality through dynamic parameter adjustment, behavioral coordination, and optimal solution caching mechanisms. At the same time, when the trunk green wave optimization module finally adjusts the signal timing scheme, it calculates the phase difference between intersections based on the distance between adjacent intersections and the green wave design speed to ensure that vehicles pass through the green light continuously when traveling at the design speed.

10. The trunk green wave optimization system based on traffic situation prediction according to claim 9, characterized in that, When the trunk green wave optimization module integrates genetic algorithm and artificial fish swarm algorithm for optimization, it dynamically adjusts the crossover probability and mutation probability of the genetic algorithm according to the population fitness to avoid the algorithm getting trapped in local optima; and sets up a behavior cooperative optimization mechanism for the artificial fish swarm algorithm, including foraging behavior optimization, swarm behavior cooperation, and tail-chasing behavior guidance; the trunk green wave optimization module maintains a global optimal solution cache in real time, recording the current optimal artificial fish state and fitness; after each iteration, if the artificial fish state is better than the optimal solution in the cache, the optimal artificial fish state and fitness are updated, and the continuous non-update counter is reset; when the continuous non-update counter counts 20 consecutive times without updating, a global neighborhood search is triggered: a neighborhood solution space is randomly generated for 20% of the artificial fish to expand the search range.

Images