Traffic signal lamp control method and system based on AI large model, and intelligent agent

Abstract

The invention relates to a traffic signal lamp control method and system based on an AI large model, and an intelligent agent. The traffic signal lamp control method comprises the following steps: caching traffic condition data from an intelligent camera, a radar and the Internet at different time windows; for each intersection in the pre-modeled road traffic network, preprocessing the traffic condition data to generate traffic condition features; and selecting an optimal scheme from a pre-configured traffic signal lamp timing selection scheme list by adopting deep reinforcement learning based on the traffic condition characteristics. According to the traffic signal lamp control method and system provided by the invention, real-time intelligent decision-making for a large number of intersections is realized through a deep reinforcement learning method based on multi-dimensional data.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent traffic signal control technology, and more specifically, to a traffic signal control method, system, and intelligent agent based on an AI big model. Background Technology

[0002] Traditional traffic lights typically use fixed-duration timing schemes, which cannot be dynamically adjusted according to real-time traffic flow, leading to severe congestion during peak hours. With the development of artificial intelligence, AI methods, including reinforcement learning, are gradually being applied to traffic light control. However, most of these current AI methods rely on single-source traffic condition data, such as data from geomagnetic coil detectors, to evaluate and calculate traffic light timing schemes. This is highly dependent on road network construction and cannot achieve real-time processing of large amounts of data.

[0003] In view of this, how to provide a traffic signal control method that can quickly and efficiently calculate traffic signal timing schemes from diverse traffic condition data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, this disclosure provides a method and system for evaluating and calculating traffic signal timing schemes based on multivariate traffic condition data, which achieves real-time intelligent decision-making for a large number of intersections through deep reinforcement learning.

[0005] To achieve the above objectives, in a first aspect of this disclosure, a traffic signal control method based on an AI large model is provided, comprising: caching traffic condition data from smart cameras, radar, and the Internet at different time windows; preprocessing the traffic condition data for each intersection in a pre-modeled road traffic network to generate traffic condition features; and selecting the optimal scheme from a pre-configured list of traffic signal timing candidate schemes based on the traffic condition features using deep reinforcement learning.

[0006] Optionally, traffic condition data from smart cameras includes flow data and queue data; traffic condition data from radar includes spillover event data; and traffic condition data from the internet includes online data from online maps, wherein the online maps include at least one of Baidu Maps, Gaode Maps, Google Maps, and Tencent Maps. Preprocessing the traffic condition data to generate traffic condition features further includes: preprocessing flow data to generate flow features; preprocessing queue data to generate queue features; preprocessing spillover event data to generate spillover event features; and preprocessing online data to generate online data features. Based on the traffic condition features, selecting the optimal solution from a pre-configured list of traffic signal timing candidates using deep reinforcement learning further includes: establishing a set of state vectors containing flow features, queue features, spillover event features, online data features, current time features, and features of each traffic signal timing candidate in the list; feeding the set of state vectors into a deep Q-network model for model inference to generate a Q-value for each traffic signal timing candidate; and selecting the traffic signal timing candidate with the highest Q-value as the optimal solution.

[0007] Optionally, the state vector set further includes traffic prediction features and queuing prediction features for the current 10-minute time window; the traffic prediction features and queuing prediction features are generated through the following steps: in response to a working day, historical traffic data and historical queuing data for the past 10 days are obtained, and in response to a non-working day, historical traffic data and historical queuing data for the past 3 days are obtained; the historical traffic data and historical queuing data are divided into 113 10-minute time windows from 05:00 to 23:50 each day at 10-minute time window intervals; and the historical traffic data and historical queuing data within each 10-minute time window are averaged and preprocessed to generate traffic prediction features and queuing prediction features for each 10-minute time window.

[0008] Optionally, traffic data from smart cameras can be cached in a time window of 240 to 600 seconds; traffic data from radar can be cached in a time window of 500 to 700 seconds; and traffic data from the Internet can be cached in a time window of 1600 to 2000 seconds.

[0009] Optionally, after selecting the optimal solution for each intersection in the pre-modeled road traffic network, the method further includes: identifying adjacent intersection pairs in the pre-modeled road traffic network; detecting whether there is a conflict between the optimal solutions of adjacent intersection pairs; if there is a conflict, adjusting the optimal solution of the intersection with higher traffic volume according to traffic condition data to avoid the intersection with lower traffic volume; otherwise, no adjustment is made.

[0010] Optionally, after selecting the optimal solution for each intersection in the pre-modeled road traffic network, the method further includes: identifying key intersections and their adjacent intersections in the pre-modeled road traffic network; obtaining traffic condition data and the optimal solution for the key intersections and their adjacent intersections; and adjusting the optimal solution for the key intersections and their adjacent intersections based on the traffic condition data.

[0011] Optionally, a list of candidate traffic signal timing schemes can be established based on traffic condition data from the past 30 days.

[0012] Optionally, for each intersection in the pre-modeled road traffic network, a decision is made and an optimal solution is generated every 50 seconds; the traffic signal control method also includes: storing the traffic condition data cached during the previous decision, the optimal solution generated during the previous decision, the traffic signal timing candidate scheme used during the previous actual execution, and the traffic condition data cached during the current decision in an experience playback pool.

[0013] In a second aspect, this disclosure also provides a traffic signal control system based on an AI large model, comprising: a data receiving module configured to cache traffic condition data from smart cameras, radar, and the Internet at different time windows; a data processing module configured to preprocess the traffic condition data to generate traffic condition features; an AI decision-making module configured to select the optimal solution from a pre-configured list of traffic signal timing candidate solutions based on the traffic condition features using deep reinforcement learning; and a result output module configured to output the optimal solution.

[0014] In a third aspect, this disclosure also provides an AI agent configured to perform the traffic signal control method according to the first aspect above.

[0015] Compared with existing technologies, the traffic signal control method and system based on AI large model provided in this disclosure have the following beneficial effects: By acquiring multidimensional data from various sources at different time windows, the traffic signal control method provided in this disclosure has improved its ability to perceive real road conditions; based on multidimensional data, the Q value of each candidate timing scheme in the candidate timing scheme list is evaluated through a deep Q-network, thereby calculating the timing scheme with the best application effect, which has a faster response speed and better decision-making effect compared with traditional adaptive control methods; it identifies conflicts between timing schemes of adjacent intersections in the road traffic network, thereby realizing "green wave" at multiple intersections and significantly reducing the total number of stops; it predicts traffic flow data and queuing data for a certain time period and according to time windows, enabling timing adjustments to be made in advance using time windows as buffer time to avoid congestion. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of this disclosure, the embodiments of this disclosure will be further explained and described with reference to the following drawings. These drawings are only used to more conveniently and specifically describe the embodiments of this disclosure and are not intended to limit this disclosure. In the drawings:

[0017] Figure 1 A flowchart of a traffic light control method based on an AI large model according to one embodiment of the present disclosure is shown;

[0018] Figure 2 As shown Figure 1 The flowchart shows the sub-steps of the traffic light control method based on an AI large model.

[0019] Figure 3 A schematic diagram of the intersection coordination process of a traffic signal control method based on an AI large model according to one embodiment of the present disclosure is shown; and

[0020] Figure 4 A schematic diagram of the structure of a traffic signal control system based on an AI large model according to one embodiment of the present disclosure is shown. Detailed Implementation

[0021] In the description of this disclosure, unless otherwise expressly specified and limited, the term "comprising" and similar terms shall be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" shall be understood as "at least partially based on". The term "one embodiment" or "the embodiment" shall be understood as "at least one embodiment". The term "some embodiments" shall be understood as "at least some embodiments". Other explicit and implicit definitions may also be included below.

[0022] This disclosure provides a traffic signal control method based on a large AI model. Figure 1 This is a flowchart of a traffic signal control method based on an AI large model according to one embodiment of the present disclosure. Figure 1As shown, the traffic signal control method 100 may include: step 120, acquiring traffic condition data from smart cameras (e.g., high-definition cameras or other camera devices specifically deployed at intersections to capture traffic condition images or videos), radar, and the Internet at different time windows; and step 140, preprocessing the traffic condition data for each intersection in a pre-modeled road traffic network to generate traffic condition features; and selecting the optimal solution from a pre-configured list of traffic signal timing candidate solutions based on the traffic condition features using deep reinforcement learning. By collecting multi-dimensional data from various monitoring devices, including smart cameras, radar, and the Internet, and combining it with artificial intelligence deep reinforcement learning methods, the AI-based large model-based traffic signal control method provided in this disclosure can select the optimal solution from a given list of traffic signal timing candidate solutions for execution or reference.

[0023] In a preferred embodiment of this disclosure, traffic data from smart cameras may include flow data and queue data; traffic data from radar may include spillover event data; and traffic data from the internet may include online data from web maps, wherein these web maps may include at least one of Baidu Maps, Gaode Maps, Google Maps, and Tencent Maps, as well as other existing web maps known to those skilled in the art and any future web maps that may emerge. References Figure 2 , Figure 2 Is it like this? Figure 1 The flowchart shown is a sub-step of step 140 in the traffic light control method based on a large AI model. Figure 2As shown, the preprocessing of traffic condition data to generate traffic condition features in step 140 can further include the following sub-steps: Step 141, preprocessing traffic flow data to generate traffic flow features; Step 142, preprocessing queuing data to generate queuing features; Step 143, preprocessing backflow event data to generate backflow event features; and Step 144, preprocessing online data to generate online data features; the selection of the optimal scheme from the pre-configured traffic signal timing candidate scheme list based on traffic condition features using deep reinforcement learning in step 140 can further include the following sub-steps: Step 145, establishing a set of state vectors containing traffic flow features, queuing features, backflow event features, online data features, current time features, and features of each traffic signal timing candidate scheme in the traffic signal timing candidate scheme list; Step 146, feeding the set of state vectors into a deep Q-network model for model inference to generate Q-values ​​for each traffic signal timing candidate scheme; and Step 147, selecting the traffic signal timing candidate scheme with the highest Q-value as the optimal scheme. Those skilled in the art will understand that data preprocessing steps 141 to 144 can be executed in parallel. Based on the deep Q-network model, using the traffic condition characteristics described above as the state, and each traffic signal timing candidate scheme in the traffic signal timing candidate scheme list as the action, the reward value (i.e., Q value) of all traffic signal timing candidate schemes is calculated for each intersection. The Q value reflects the effectiveness of the traffic signal timing candidate scheme (specifically, it can be comprehensively set by combining factors such as the total passage time of the intersection, the total waiting time of vehicles at the intersection, the number of times vehicles stop at the intersection, and the differences between these factors caused by adopting the candidate scheme compared with the original timing scheme). The higher the Q value, the more significant the effectiveness of the traffic signal timing candidate scheme. This traffic signal control method ultimately returns the traffic signal timing candidate scheme with the highest Q value for execution or reference.

[0024] In a preferred embodiment of this disclosure, the traffic signal control method is implemented based on a pre-established road traffic network model. This model includes all relevant locations, lanes, and intersections, as well as the mapping relationships between intelligent cameras, radar, and internet monitoring devices and the aforementioned locations, lanes, and intersections. (Reference) Figure 2 In a preferred embodiment of this disclosure, traffic condition data from a smart camera may include flow data, as illustrated in the following pseudocode. Figure 2 The process of traffic data preprocessing described in step 141 is as follows: This pseudocode takes a list of traffic flow data within a time window as input and outputs traffic flow statistics for each intersection and direction as traffic flow features. These features reflect the real-time traffic conditions at each intersection. The location number identifies the smart camera installed at each intersection in the road network, and the lane number identifies the lane captured by the smart camera. From the location number, the intersection ID and the current traffic direction at the intersection can be obtained (where L / R / U / D indicate westbound / eastbound / northbound / southbound traffic, respectively). Based on the traffic flow data captured by the smart camera, the corresponding traffic flow count for each direction at the intersection is accumulated, timestamped, and saved as the traffic flow features described in this paper.

[0025] Continue to refer to Figure 2 In a preferred embodiment of this disclosure, traffic condition data from a smart camera may include queuing data, as illustrated in the following pseudocode. Figure 2 The process of queuing data preprocessing operation described in step 142 is as follows: This pseudocode takes a list of queue data within an input time window and outputs the maximum queue length and total number of vehicles for each lane in each direction at each intersection as queue features. These queue features reflect the long-term queue trend at each intersection. As those skilled in the art will understand, traffic flow data and queue data can be specifically identified and summarized from images / videos captured by smart cameras based on conventional image recognition techniques. For simplicity, the process of identifying and summarizing traffic flow data and queue data from traffic condition data from smart cameras is not shown in steps 141 and 142 above.

[0026] Still referencing Figure 2 In a preferred embodiment of this disclosure, traffic condition data from radar may include radar event data, which generally includes queue length data, vehicle speed data, and traffic flow data. Based on this data, overflow event data with greater reference value for intersection traffic adjustment can be further generated. The following pseudocode exemplarily illustrates the implementation as follows: Figure 2 The process of the overflow event data preprocessing operation described in step 143 is as follows: This pseudocode takes radar event data as input and outputs overflow event features for each intersection and direction. Specifically, the overflow event features (overflowMap) include the state (0 / 1, indicating no overflow / overflow exists respectively), distance and time for each intersection and direction of traffic, which are incorporated into the state vector set as overflow event features.

[0027] Still referencing Figure 2In a preferred embodiment of this disclosure, the online data preprocessing process described in step 144 may include: for each intersection in the network map corresponding to the road traffic network model preset in this disclosure, segmenting the online data from the network map into online data features within a certain time window (e.g., 10 minutes, 150 seconds, or shorter, such as 5 seconds to 1 second), and saving them as online data features. Specifically, the online data features may be the congestion level of each intersection, for example, divided into four levels from smooth to congested and assigned values ​​from 0 to 3. In an alternative embodiment of this disclosure, the online data from the network map and / or radar event data from radar may be structured data, which can be directly used by the AI ​​large model provided in this disclosure without a preprocessing process.

[0028] In a preferred embodiment of this disclosure, the state vector set further includes traffic prediction features and queuing prediction features for the current 10-minute time window; the traffic prediction features and queuing prediction features are generated through the following steps: acquiring historical traffic data and historical queuing data for the past 10 days in response to the current date being a working day, and acquiring historical traffic data and historical queuing data for the past 3 days in response to the current date being a non-working day; dividing the historical traffic data and historical queuing data into 113 10-minute time windows at 10-minute intervals from 05:00 to 23:50 each day; and averaging and preprocessing the historical traffic data and historical queuing data within each 10-minute time window to generate traffic prediction features and queuing prediction features for each 10-minute time window. In another preferred embodiment of this disclosure, the state vector set further includes traffic prediction features and queuing prediction features for the current 150-second time window, wherein the historical traffic data and historical queuing data are divided into 452 150-second time windows at 150-second intervals from 05:00 to 23:50 each day, and traffic prediction features and queuing prediction features are generated for these time windows. The following pseudocode exemplifies the process of generating traffic prediction features with 10-minute and 150-second time windows, respectively. The code for generating queuing prediction features is similar:

[0029] Return to reference Figure 1In a preferred embodiment of this disclosure, step 120 may further include: caching traffic condition data from smart cameras in a 240-600 second time window, for example, caching traffic flow data in a 600-second time window for long-term trend analysis, and caching queue data in a 240-second time window for short-term condition assessment; caching traffic condition data from radar in a 500-700 second time window, for example, caching radar event data including queue length data, vehicle speed data, and traffic flow data in a 600-second time window for radar event detection; and caching traffic condition data from the internet in a 1600-2000 second time window, for example, caching online data from web maps in an 1800-second time window for road network condition perception and information gap filling. By caching data in different time windows based on the characteristics of data from different sources, the overall traffic condition at the intersection can be perceived more accurately.

[0030] In a preferred embodiment of this disclosure, the state vector set may further include one or more of the following: a short-window traffic flow characteristic, i.e., a traffic flow characteristic generated from the most recent 150-second time window traffic flow data extracted from the aforementioned 600-second time window traffic data, which reflects a more real-time traffic flow characteristic; current time characteristics, including hour, minute, and weekday; stage characteristics of traffic signal timing candidate schemes matching the current time, typically traffic signal timing schemes are configured based on time periods, for example, 08:00 to 10:00 can be divided into four stages in 30-minute intervals (the stage duration is dynamic, different for different dates and times, the larger the traffic flow, the narrower the stage division, and the shorter the stage duration) and numbered separately, and different traffic signal timing schemes are configured for each stage, the stage characteristics including the current stage number and duration, which can be used to obtain the list of currently applicable traffic signal timing candidate schemes; and characteristics of each traffic signal timing candidate scheme in the list of traffic signal timing candidate schemes, i.e., the timing duration of each traffic signal timing candidate scheme for each traffic phase. In a preferred embodiment of this disclosure, the state vector containing one or more features as described in any of the preceding sections can be further normalized to obtain a state tensor (e.g., different weights are assigned to different features), and model inference is performed from this state tensor based on a deep Q-network model to obtain Q-values. The Q-values ​​of each traffic signal timing candidate scheme in the list of traffic signal timing candidate schemes are stored in a Q-value dictionary, and the ID of the traffic signal timing candidate scheme with the largest Q-value and the largest Q-value are returned to the operator for execution or reference. The following pseudocode exemplarily illustrates the process from constructing the state vector to selecting the optimal scheme: This pseudocode takes multidimensional features from multidimensional data and a list of candidate traffic signal timing schemes as input, and outputs the ID and Q value of the optimal scheme.

[0031] refer to Figure 3 , Figure 3 This is a schematic diagram of the intersection coordination process 300 of the traffic signal control method based on an AI large model according to one embodiment of the present disclosure. In a preferred embodiment of the present disclosure, after selecting the optimal solution for each intersection in the pre-modeled road traffic network, the traffic signal control method based on the AI ​​large model provided by the present disclosure may further include an adjacent intersection coordination process 310, which includes the following steps: step 311, identifying adjacent intersection pairs in the pre-modeled road traffic network; step 312, detecting whether there is a conflict between the optimal solutions of the adjacent intersection pairs; and step 313, if a conflict exists, adjusting the optimal solution of the intersection with higher traffic volume according to traffic condition data to make it avoid the intersection with lower traffic volume, otherwise no adjustment is made. The conflicts mainly include the following conflict types: intersection, merging, separation, and conflicts between vehicles and pedestrians. The existence and specific conflict types can be predicted by the optimal solution of each intersection and the current traffic volume (specifically, since the online data coverage of the Internet is the most extensive, online data from the Internet can be used to determine the traffic volume of adjacent intersections). In a preferred embodiment of this disclosure, the optimal solution for adjacent intersection pairs is further adjusted by phase-time offset or duration fine-tuning based on topology / traffic conditions, so that conflicting phases are staggered, reducing congestion caused by simultaneous passage. The following pseudocode exemplarily illustrates the implementation of the adjacent intersection coordination process 310:

[0032] Continue to refer to Figure 3 In a preferred embodiment of this disclosure, after selecting the optimal solution for each intersection in the pre-modeled road traffic network, the traffic light control method based on the AI ​​large model provided by this disclosure may further include a focus intersection coordination process 320, which includes the following steps: step 321, identifying focus intersections and their adjacent intersections in the pre-modeled road traffic network; step 322, acquiring traffic condition data (including the distance between the focus intersections identified in the road traffic network and their adjacent intersections, and vehicle speed data obtained from radar event data) and the optimal solution for the focus intersections and their adjacent intersections; and step 323, adjusting the optimal solution for the focus intersections and their adjacent intersections based on the traffic condition data (for example, calculating the travel time using the aforementioned distance and vehicle speed data, thereby optimizing the phase offset of the optimal solution for the focus intersections and their adjacent intersections), thereby obtaining a "green wave" (i.e., vehicles continuously pass through the green light). In a preferred embodiment of this disclosure, as... Figure 3As shown, after selecting the optimal solution for each intersection in the pre-modeled road traffic network, the traffic signal control method based on the AI ​​large model provided in this disclosure may further include a backflow event intersection coordination process 330, which includes the following steps: step 331, identifying intersections judged to have backflow events from the backflow event data; step 332, identifying intersections located downstream of the intersections with backflow events; and step 323, adjusting the optimal solution for the intersections located downstream of the intersections with backflow events, so that the downstream intersections delay traffic to alleviate the backflow event. Those skilled in the art will understand that in the above process of selecting the optimal solution for each intersection in the pre-modeled road traffic network and in the process of coordinating specific intersections, the selected or adjusted traffic signal timing candidate solutions are all within a reasonable phase duration and their phase switching conforms to the rules.

[0033] In a preferred embodiment of this disclosure, the traffic signal timing candidate scheme list described herein can be configured by traffic police managing traffic signals based on experience. In another preferred embodiment of this disclosure, the traffic signal timing candidate scheme list described herein can be established based on collected traffic condition data (e.g., traffic condition data collected within the last 30 days). Preferably, for intersections with smooth traffic flow, the traffic signal timing candidate scheme list can be configured by traffic police based on experience; for intersections with high traffic volume and peak hours at intersections, the traffic signal timing candidate scheme list can be generated based on collected traffic condition data. In a preferred embodiment of this disclosure, the traffic signal timing candidate scheme list may include an inner table and an outer table, wherein the inner table is generated based on the methods described herein, and the outer table can be adaptively adjusted by traffic police. When the traffic signal timing candidate scheme list is updated using the methods provided herein, the updated list is stored in the inner table. In a preferred embodiment of this disclosure, if the outer table remains unchanged for 30 days, it is automatically updated to match the inner table. The following pseudocode exemplifies the process of creating a list of candidate traffic signal timing schemes for peak hours at an intersection:

[0034] In a preferred embodiment of this disclosure, for each intersection in the pre-modeled road traffic network, a decision is made and an optimal solution is generated every 50 seconds; the traffic signal control method further includes: storing the traffic condition data cached during the previous decision, the optimal solution generated during the previous decision, the traffic signal timing candidate scheme used during the previous actual execution, and the traffic condition data cached during the current decision in an experience playback pool for learning. In a preferred embodiment of this disclosure, the traffic signal control method based on an AI large model defines an AI agent for each intersection. These AI agents are named, for example, DQN_select. Upon acquiring the traffic state data, the list of traffic signal timing candidates, and the experience replay pool described herein, the DQN_select agent can generate prompt words based on this data, including descriptions of the traffic state data (i.e., state space), the list of traffic signal timing candidates (i.e., action space), the experience replay pool, and the phase safety duration (i.e., the reasonable phase duration and switching rules as described above). These prompts are then fed into the AI ​​large model, thereby driving the AI ​​large model to generate Q-values ​​for each traffic signal timing candidate based on a deep Q-network. In this way, the method provided by this disclosure can perform comprehensive evaluation in real time based on a large amount of multidimensional data from different data sources.

[0035] In a preferred embodiment of this disclosure, based on an AI large model, the traffic signal control method provided by this disclosure can further optimize one or more of the aforementioned processes. In an exemplary embodiment, after selecting the optimal solution for each intersection in a pre-modeled road traffic network and performing intersection coordination, relevant data and historical coordination data can be fed into the AI ​​large model to comprehensively plan for the aforementioned intersection coordination process 300 (including adjacent intersection coordination process 310, key attention intersection coordination process 320, and spillover event intersection coordination process 330), thereby deriving further optimization of the optimal solution while considering all intersections. In another exemplary embodiment, during the internal table update process as described above, daily traffic condition data and experience replay pool can be fed into the AI ​​large model to further fine-tune the traffic signal timing candidate solutions in the internal table. In yet another exemplary embodiment of this disclosure, the AI ​​large model can also be used to calculate parameters such as the response time from receiving traffic condition features to calculating the Q value, and the model prediction accuracy of improved traffic conditions after the optimal solution is adopted, thereby performing an overall evaluation of the model. In yet another exemplary embodiment of this disclosure, during the process of generating anti-spillage event characteristics based on radar event data, an AI large model can also be used to collect radar event data within the most recent 10-minute time window and perform anti-spillage event prediction for each intersection, thereby informing the operators controlling the traffic signal timing scheme of the current intersection in advance of the prediction results.

[0036] refer to Figure 4 , Figure 4 A schematic diagram of the structure of a traffic signal control system 400 based on an AI large model, according to one embodiment of this disclosure, is shown. Figure 4 As shown, in a preferred embodiment of this disclosure, the traffic signal control system 400 includes a data receiving module 410 configured to cache traffic condition data from smart cameras, radar, and the Internet at different time windows; a data processing module 420 configured to preprocess the traffic condition data to generate traffic condition features; an AI decision-making module 430 configured to select the optimal solution from a pre-configured list of traffic signal timing candidate solutions based on the traffic condition features using deep reinforcement learning; and a result output module 440 configured to output the optimal solution. Figure 4 As shown, in a preferred embodiment of this disclosure, the traffic signal control system 400 further includes a timetable management module 450, which includes an inner table and an outer table. Both the inner table and the outer table are configured with a list of traffic signal timing candidate schemes. The inner table is updated daily based on traffic condition data, and the outer table supports manual intervention, thereby flexibly responding to special situations.

[0037] In a preferred embodiment of this disclosure, an AI agent is also provided, which is configured to perform the traffic signal control method described herein.

[0038] Example 1: Complete system deployment based on this disclosure

[0039] 1. Hardware Configuration Server: 8-core CPU, 32GB RAM, 500GB SSD Network: Gigabit LAN, supporting both Socket and HTTP protocols Detection equipment: smart cameras, radar equipment

[0040] 2. Software Environment Operating System: Ubuntu 20.04 LTS Python version: 3.11 Deep learning framework: PyTorch 2.0 Web framework: Flask 2.0 Task scheduling: APScheduler 3.9

[0041] 3. Deployment Steps Step 1: Install dependencies Step 2: Configure intersection information (Lambdas.py) Step 3: Load the DQN model Place the trained model file in the Agent directory. Model naming format: dqn_{intersection_id}.pth Model structure: Input layer (N-dimensional) → Hidden layer (64) → Output layer (1) Step 4: Start the main service After the service starts: The socket service listens on port 65432. HTTP service listens on port 8088 The data processing thread executes once every 50 seconds. Traffic / queue prediction task is executed daily at 3:00 AM. Step 5: Start the schedule service After the schedule service is launched: The internal table update task is executed daily at 0:00. External appearance is managed via HTTP interface

[0042] 4. Operation Process

[0043] This embodiment achieves the following beneficial effects: compared with fixed timing, dynamic timing reduces the average delay time by 25%; compared with traditional adaptive control, DQN decision improves traffic efficiency by 15%; it updates every 50 seconds, and the response speed is 3 times faster than the traditional SCOOT system; it coordinates the control of 139 intersections to achieve regional "green wave"; it reduces conflicts between adjacent intersections by 40% and the number of stops by 20%.

[0044] Various implementations of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed implementations. Many modifications and variations will be readily understood by those skilled in the art without departing from the scope and spirit of the described implementations. The terminology used herein is chosen to best explain the principles, practical applications, or improvements to technology in the market, or to enable others skilled in the art to better understand the various implementations disclosed herein.

Claims

1.A traffic signal control method based on an AI large model, characterized by, comprising: buffering traffic condition data from smart cameras, radars and the internet with different time windows; and for each intersection in a pre-modeled road traffic network, pre-processing the traffic condition data to generate traffic condition features; and based on the traffic condition features, employing deep reinforcement learning to select an optimal traffic signal timing plan from a pre-configured list of traffic signal timing plans. 2.The traffic signal control method of claim 1, wherein the traffic condition data from the smart cameras comprises flow data and queue data, the traffic condition data from the radars comprises counter-flow event data, and the traffic condition data from the internet comprises online data from a network map, wherein the network map comprises at least one of Baidu Map, Gaode Map, Google Map, Tencent Map; the pre-processing the traffic condition data to generate traffic condition features further comprises: pre-processing the flow data to generate flow features; pre-processing the queue data to generate queue features; pre-processing the counter-flow event data to generate counter-flow event features; and pre-processing the online data to generate online data features; the based on the traffic condition features, employing deep reinforcement learning to select an optimal traffic signal timing plan from a pre-configured list of traffic signal timing plans further comprises: establishing a state vector set comprising the flow features, the queue features, the counter-flow event features, the online data features, a current time feature, and a feature of each traffic signal timing plan in the list of traffic signal timing plans; feeding the state vector set into a deep Q-network model for model inference to generate a Q-value for each traffic signal timing plan, respectively; and selecting the traffic signal timing plan with the highest Q-value as the optimal traffic signal timing plan. 3.The traffic signal control method of claim 2, wherein the state vector set further comprises flow prediction features and queue prediction features for a current 10-minute time window; the flow prediction features and the queue prediction features are generated by: obtaining historical flow data and historical queue data for the past 10 days in response to the current day being a weekday, and obtaining historical flow data and historical queue data for the past 3 days in response to the current day being a non-weekday; segmenting the historical flow data and the historical queue data into 113 10-minute time windows with 10-minute time windows as intervals from 05:00 to 23:50 every day; and averaging and pre-processing the historical flow data and the historical queue data in each 10-minute time window to generate the flow prediction features and the queue prediction features for each 10-minute time window. 4.The traffic signal control method of claim 1, wherein the traffic condition data from the smart cameras is buffered with a 240-second to 600-second time window. cache the traffic condition data from the radar with a 500-700 second time window; and cache the traffic condition data from the internet with a 1600-2000 second time window. 5.The traffic light control method of claim 1, wherein after selecting the optimal solution for each intersection in the pre-modeled road traffic network, further comprising: identifying a pair of adjacent intersections in the pre-modeled road traffic network; detecting whether there is a conflict between the optimal solutions of the pair of adjacent intersections, if there is a conflict, adjusting the optimal solution of the intersection with larger traffic volume to avoid the intersection with smaller traffic volume according to the traffic condition data; otherwise, not adjusting. 6.The traffic light control method of claim 1, wherein after selecting the optimal solution for each intersection in the pre-modeled road traffic network, further comprising: identifying a key intersection and its adjacent intersections in the pre-modeled road traffic network; obtaining the traffic condition data and the optimal solution of the key intersection and its adjacent intersections; and adjusting the optimal solution of the key intersection and its adjacent intersections based on the traffic condition data. 7.The traffic light control method of claim 1, wherein the list of traffic light scheduling solutions is established based on the traffic condition data in the last 30 days. 8.The traffic light control method of claim 1, wherein for each intersection in the pre-modeled road traffic network, the decision is made and the optimal solution is generated every 50 seconds; the traffic light control method further comprises: saving the traffic condition data cached at the last decision, the optimal solution generated at the last decision, the traffic light scheduling solution adopted at the last actual execution, and the traffic condition data cached at the current decision in an experience replay pool. 9.A traffic signal control system based on an AI large model, characterized by, comprising a data receiving module configured to cache traffic condition data from smart cameras, radars, and the internet with different time windows; a data processing module configured to pre-process the traffic condition data to generate traffic condition features; an AI decision module configured to select an optimal solution from a pre-configured list of traffic light scheduling solutions based on the traffic condition features using deep reinforcement learning; and a result output module configured to output the optimal solution. 10.An AI agent configured to perform the traffic light control method of any one of claims 1-8.

Images