Traffic flow intelligent balancing system and method based on deep reinforcement learning and congestion avoidance
By combining deep reinforcement learning and real-time response and prediction models of roadside unit RSU in an intelligent traffic flow balance system, global and local collaborative optimization is achieved, and the problems of traffic congestion and multi-objective optimization in the existing technology are solved, which significantly improves the The effectiveness and sustainability of traffic management systems.
Patent Information
- Application Number
- CN202411585921.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-07
- Publication Date
- 2025-02-18
AI Technical Summary
The existing route guidance system lacks a global perspective when optimizing traffic flow, and cannot effectively respond to the needs of concurrent paths of multiple vehicles, resulting in traffic congestion and lacks accurate congestion prediction capabilities, making it difficult to achieve multi-objective optimization.
The intelligent balance system of traffic flow based on deep reinforcement learning is adopted, and the global path optimization is performed through cloud agents, combined with the real-time response and prediction model of the roadside unit RSU, to achieve global and local collaborative optimization, and actively alleviate traffic congestion.
Effectively alleviate traffic congestion, shorten vehicle driving time, reduce energy consumption and pollution emissions, improve the efficiency of traffic management systems, and help cities achieve sustainable development.
Smart Images
Figure CN119479282A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a traffic flow intelligent balancing system and method based on deep reinforcement learning and congestion avoidance, and belongs to the field of intelligent transportation system control. Background Art
[0002] Against the backdrop of accelerating urbanization, the urban population has grown dramatically, and with it comes a significant increase in traffic demand and increasingly severe traffic congestion. This congestion not only affects citizens' travel experience and increases commuting time, but also brings about problems such as high energy consumption and environmental pollution. As the main source of energy consumption and greenhouse gas emissions, transportation accounts for 33% of global carbon dioxide emissions. Therefore, improving the energy efficiency of the transportation system, reducing emissions, and reducing the negative impact of the transportation system on the environment have become urgent needs for major cities to achieve sustainable development.
[0003] Although existing route guidance systems (RGS) have been used in intelligent transportation systems (ITS), they are mostly limited to single-vehicle path optimization and cannot fully consider the dynamic changes of the overall road network and the coordination of multi-vehicle behaviors. This approach often causes multiple vehicles to choose the same path, which in turn forms new congestion points, and ignores the balance of multiple factors such as fuel consumption and emission control when optimizing a single goal (such as the shortest path).
[0004] At the same time, the existing route guidance system (RGS) has limitations in dealing with the concurrent route demands of multiple vehicles in complex urban road networks. Traditional systems mainly optimize the route of a single vehicle and lack a global perspective, which easily leads to multiple vehicles choosing the same route and forming new congestion. In addition, RGS systems usually only focus on travel time and ignore the need for multi-objective optimization. The existing RGS also lacks accurate congestion prediction capabilities and usually relies on real-time data for passive response, making it difficult to predict potential congested areas and dynamic traffic demands in a timely manner. In the process of route allocation, how to achieve efficient multi-objective optimization in the competition among multiple vehicles and dynamic traffic changes has become a key technical challenge faced by RGS in urban road networks. Summary of the invention
[0005] The technical problem solved by the present invention is to provide a traffic flow intelligent balancing system and method based on deep reinforcement learning and congestion avoidance, which uses deep reinforcement learning to optimize traffic flow from both global and local levels. The cloud-based intelligent agent optimizes the global path through historical data to improve the overall traffic distribution efficiency, while the roadside unit RSU intelligent agent can quickly respond to local traffic conditions based on real-time feedback and effectively alleviate local traffic congestion. At the same time, the roadside unit RSU introduces a prediction model to identify potential congestion in advance and actively alleviate traffic congestion. Compared with the traditional passive response method, this active intervention can effectively reduce the possibility of congestion, shorten vehicle driving time, and reduce energy consumption and pollution emissions.
[0006] The technical solution adopted by the present invention is: an intelligent traffic flow balancing system based on deep reinforcement learning and congestion avoidance, comprising:
[0007] The cloud generates the road network weights at future times based on the real-time road network traffic conditions collected by the roadside units (RSUs). It generates a traffic allocation plan based on the road network weights at future times and the path planning requests and vehicle information of all vehicles entering the road network sent by the roadside units (RSUs). The plan is then returned to the roadside units. In addition, the cloud also provides the road section weights at future times based on the decision requirements sent by the roadside units (RSUs).
[0008] Roadside units (RSUs) collect route planning requests from vehicles entering the road network and monitor vehicle information within the coverage area in real time, and observe the real-time road network traffic conditions on the road sections in real time. They upload these data to the cloud, receive the traffic allocation plan sent by the cloud, and convey it to the vehicles that initiate route planning requests in the controlled road sections. At the same time, the roadside units (RSUs) build congestion formation trends in real time. When congestion risks are detected, they generate rerouting decisions and congested road section information, obtain road network weights at future times from the cloud, and send the road network weights at future times to the selected vehicles that need to be rerouted.
[0009] The vehicle unit is used to send a path planning request and vehicle information to the roadside unit RSU, and receive the traffic distribution plan forwarded by the roadside unit RSU to enter the road network. In the road network, if it is selected as a rerouting vehicle, it will recalculate the optimal path to the destination based on the congested road section information sent by the roadside unit RSU and the road section weight at the future time.
[0010] Specifically, the roadside unit RSU is fixedly deployed at the intersection of the road network, and the communication and observation range is limited to all road sections connected to the intersection. The coverage areas are independent of each other, and vehicles traveling in a certain area can only communicate with the roadside units RSU in the area.
[0011] Specifically, the real-time road network traffic conditions include average driving time, energy consumption, carbon dioxide emissions, and the number of vehicles on a road section.
[0012] Specifically, the vehicle information is a starting position and a destination.
[0013] Specifically, the roadside unit RSU generates a rerouting decision and congested road section information according to the HAA2C algorithm, and the rerouting decision is to select an optimal proportion of vehicles for rerouting.
[0014] Specifically, the cloud generates a traffic distribution plan based on the HAA2C algorithm. The traffic distribution plan is the number of paths for the vehicle that initiates the path planning request to reach the destination and the proportion of vehicles on these paths.
[0015] A method for intelligent balancing of traffic flow based on deep reinforcement learning and congestion avoidance. The specific steps are as follows:
[0016] Step 1: The vehicle that is about to enter the controlled road network sends a path planning request and vehicle information to the roadside unit RSU in the area;
[0017] Step 2: Roadside units (RSUs) collect real-time road network traffic conditions within the coverage area, and upload the real-time road network traffic conditions, path planning requests sent by vehicles, and vehicle information to the cloud;
[0018] Step 3: The cloud uses the real-time road network traffic conditions to calculate the road section weights at future times through the MTT model, and uses the KSP algorithm to calculate k optimal paths using the road section weights at future times and the path planning request and vehicle information sent by the vehicle. The HAA2C algorithm is used to generate a flow distribution plan that satisfies the optimal principle of the flow balance principle system and the distribution ratio of each path, and the flow distribution plan is fed back to the roadside unit RSU where the initiating vehicle is located;
[0019] Step 4: The roadside unit (RSU) to which the vehicle that initiates the route planning request belongs forwards the received traffic allocation plan to the vehicles within its jurisdiction to complete the route allocation instruction transmission;
[0020] Step 5: The vehicles that have obtained the traffic allocation plan enter the road network according to the allocated path;
[0021] Step 6: The roadside unit (RSU) builds the congestion formation trend in real time according to the CNN-LSTM-Attention model, identifies the impending traffic congestion based on the congestion formation trend, generates rerouting decisions and congested road section information when the congestion risk is detected, obtains the road network weight at the future time from the cloud, and sends the road network weight at the future time to the selected vehicle that needs to be rerouted;
[0022] Step 7: The vehicle that receives the congestion information uses the shortest path algorithm to recalculate the optimal path to the destination based on the congested road section information provided by the RSU and the road network weight at the future time;
[0023] Step8: Repeat Step1-Step7.
[0024] Specifically, the specific steps of Step 6 are:
[0025] The roadside unit RSU inputs its own stored historical traffic flow data into the CNN-LSTM-Attention model to predict the future traffic flow of the controlled section, builds a congestion formation trend based on the historical traffic flow data and the predicted traffic flow, and identifies the impending traffic congestion based on the congestion formation trend. When the congestion risk is detected, the roadside unit RSU selects the optimal proportion of vehicles from the vehicles about to enter the congested section for rerouting based on HAA2C, and passes the section information that may cause congestion and the section weights at future times to the relevant vehicles.
[0026] Specifically, the specific steps of the roadside unit RSU inputting the historical traffic flow data stored in itself into the CNN-LSTM-Attention model to predict the future traffic flow of the controlled section are:
[0027] The Attention-CNN-LSTM model uses historical traffic flow data as input to predict the traffic flow of the next time step of the road section. The congestion formation trend is constructed based on the predicted traffic flow, the traffic flow at the previous moment, and the traffic flow at the time of congestion. The trend is constructed through variable differences, including the difference between the predicted value and the value at the previous moment, and the congestion value and the predicted value. If the traffic flow changes accelerate or approaches the congestion value, the congestion trend accelerates.
[0028] Specifically:
[0029] First, CNN extracts local temporal features from historical traffic flow data through one-dimensional convolution and captures short-term dependencies at different scales.
[0030] Subsequently, BiLSTM further extracts long-term dependencies and transfers information bidirectionally to comprehensively characterize the complex temporal patterns of historical traffic flow data, with special attention paid to key time steps to cope with critical changes in traffic fluctuations;
[0031] Finally, the fully connected layer flattens the features processed by CNN, BiLSTM, and attention mechanism to generate the final traffic flow prediction results.
[0032] The beneficial effects of the present invention are as follows: through the combination of the HAA2C algorithm and the semi-distributed rerouting architecture, the system realizes global and local collaborative optimization, improves the accuracy and response speed of path allocation, can alleviate urban traffic pressure, improve citizens' travel experience, significantly reduce network congestion and carbon emissions, and improve the efficiency of the traffic management system. It will help cities optimize environmental protection, resource utilization efficiency and energy consumption control, provide strong support for global sustainable development, and is suitable for intelligent traffic management of large-scale complex urban road networks. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is a schematic diagram of the system scenario of the present invention;
[0034] Figure 2 It is a schematic diagram of the prediction method of the present invention;
[0035] Figure 3 It is a schematic diagram of the flow balancing decision method of the present invention;
[0036] Figure 4 It is a flow chart of the flow balancing method of the present invention. DETAILED DESCRIPTION
[0037] The present invention is further described below in conjunction with the accompanying drawings and specific implementation methods.
[0038] Example 1: Figure 1 As shown, a traffic flow intelligent balancing system based on deep reinforcement learning and congestion avoidance includes:
[0039] The cloud collects the path planning requests and vehicle information of all vehicles entering the road network. At the same time, the cloud-based intelligent agent receives the real-time road network traffic conditions collected by the roadside unit, and generates the road network weights for the future based on this. Based on the road network weights for the future and the path planning requests and vehicle information of all vehicles entering the road network sent by the roadside unit RSU, the cloud generates a traffic allocation plan through the HAA2C algorithm based on the system optimal goal based on the principle of traffic balancing. The plan is then sent by the cloud to the roadside unit RSU where the corresponding vehicle is located, and forwarded to the corresponding vehicle through the roadside unit RSU to implement path allocation. In addition, the cloud also provides the road section weights for the future based on the decision-making requirements sent by the roadside unit RSU to support local traffic optimization.
[0040] Roadside unit RSU is responsible for collecting the path planning requests of vehicles entering the road network and monitoring the vehicle information within the coverage area in real time, and observing the real-time road network traffic conditions on the road section in real time, uploading these data to the cloud, receiving the traffic allocation plan sent by the cloud, and conveying it to the vehicles that initiate the path planning request in the controlled road section. At the same time, the roadside unit RSU builds the congestion formation trend in real time. When the congestion risk is detected, it generates rerouting decisions and congested road section information, obtains the road network weight at the future time from the cloud, and sends the road network weight at the future time to the selected vehicle that needs to be rerouted;
[0041] The vehicle unit is used to send a path planning request and vehicle information to the roadside unit RSU, and receive the traffic distribution plan forwarded by the roadside unit RSU to enter the road network. In the road network, if it is selected as a rerouting vehicle, it will recalculate the optimal path to the destination based on the congested road section information sent by the roadside unit RSU and the road section weight at the future time.
[0042] Furthermore, the roadside unit RSU is fixedly deployed at the intersection of the road network, and the communication and observation range is limited to all road sections connected to the intersection. The coverage areas are independent of each other, and vehicles traveling in a certain area can only communicate with the roadside units RSU in the area.
[0043] Furthermore, the real-time road network traffic conditions include average travel time, energy consumption, carbon dioxide emissions, and the number of vehicles on the road section.
[0044] Furthermore, the vehicle information includes a starting location and a destination.
[0045] Furthermore, the roadside unit RSU generates a rerouting decision and congested road section information according to the HAA2C algorithm, and the rerouting decision is to select an optimal proportion of vehicles for rerouting.
[0046] Furthermore, the cloud generates a traffic distribution plan based on the HAA2C algorithm, and the traffic distribution plan is the number of paths for the vehicle initiating the path planning request to reach the destination and the proportion of vehicles on these paths.
[0047] A method for intelligent balancing of traffic flow based on deep reinforcement learning and congestion avoidance. The specific steps are as follows:
[0048] Step 1: The vehicle that is about to enter the controlled road network sends a path planning request and vehicle information to the roadside unit RSU in the area;
[0049] Step 2: Roadside units (RSUs) collect real-time road network traffic conditions within the coverage area, and upload the real-time road network traffic conditions, path planning requests sent by vehicles, and vehicle information to the cloud;
[0050] Step 3: The cloud uses the real-time road network traffic conditions to calculate the road section weights at future times through the MTT model, and uses the KSP algorithm to calculate k optimal paths using the road section weights at future times and the path planning request and vehicle information sent by the vehicle. The HAA2C algorithm is used to generate a flow distribution plan that satisfies the optimal principle of the flow balance principle system and the distribution ratio of each path, and the flow distribution plan is fed back to the roadside unit RSU where the initiating vehicle is located;
[0051] Step 4: The roadside unit (RSU) to which the vehicle that initiates the route planning request belongs forwards the received traffic allocation plan to the vehicles within its jurisdiction to complete the route allocation instruction transmission;
[0052] Step 5: The vehicles that have obtained the traffic allocation plan enter the road network according to the allocated path;
[0053] Step 6: The roadside unit (RSU) builds the congestion formation trend in real time according to the CNN-LSTM-Attention model, identifies the impending traffic congestion based on the congestion formation trend, generates rerouting decisions and congested road section information when the congestion risk is detected, obtains the road network weight at the future time from the cloud, and sends the road network weight at the future time to the selected vehicles that need to be rerouted, and performs local traffic optimization in advance to maintain the global optimal state;
[0054] Step 7: Vehicles that receive congestion information use the shortest path algorithm to recalculate the optimal path to the destination based on the congested road section information provided by the RSU and the road network weight at the future time. Under the premise of meeting the global equilibrium goal, the vehicle unit tries its best to achieve its own user optimal path selection and chooses a non-congested path to relieve traffic pressure and achieve global traffic equilibrium.
[0055] Step 8: Repeat Step 1-Step 7 to achieve real-time dynamic balance of traffic flow.
[0056] System optimization goal: Through real-time dynamic traffic adjustment and active congestion avoidance, the system achieves network-wide carbon dioxide emissions and energy consumption reduction, shortens vehicle travel time, and thus achieves efficient and sustainable management of intelligent transportation systems.
[0057] Furthermore, the specific steps of Step 6 are:
[0058] like Figure 2 As shown, the present invention uses the Attention-CNN-LSTM model to take historical traffic flow data as input to predict the traffic flow of the road section at the next time step, and constructs the congestion formation trend based on the predicted traffic flow, the traffic flow at the previous moment, and the traffic flow at the time of congestion. The trend is constructed through variable differences, including the difference between the predicted value and the value at the previous moment, and the congestion value and the predicted value. If the traffic flow changes accelerate or approaches the congestion value, the congestion trend accelerates. In order to make decisions before congestion. After the model is trained on the collected historical traffic flow data, it is deployed to the roadside unit (RSU) for real-time prediction. The prediction results are used as the state input of the roadside unit intelligent agent and the basis for the intelligent agent's decision.
[0059] The Attention-CNN-LSTM model combines a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism to improve prediction accuracy. First, CNN extracts local temporal features from historical traffic flow data through one-dimensional convolution, capturing short-term dependencies at different scales and enhancing the understanding of complex traffic patterns. Subsequently, BiLSTM further extracts long-term dependencies and transmits information bidirectionally to fully characterize the complex temporal patterns of traffic flow, paying special attention to key time steps to cope with critical changes in traffic fluctuations.
[0060] Finally, the fully connected layer flattens the features processed by CNN, BiLSTM and attention mechanism to generate the final traffic flow prediction results. The dropout layer reduces overfitting and improves the generalization ability of the model. The model integrates spatiotemporal feature extraction and attention to key time steps to achieve more accurate prediction of traffic flow changes, provide reliable input for reinforcement learning agents, and support intelligent decision-making.
[0061] like Figure 3 As shown in the figure, in the HAA2C algorithm, multiple neural networks are used: each agent contains an independent Actor network and Critic network. The Actor network is responsible for outputting actions and determining its interaction behavior with the environment based on the local observation information of each agent (i.e., the environmental state perceived by the agent); while the Critic network takes the observation and action of the agent as input and outputs the state-action value function (Q value) to evaluate the long-term reward of the action, thereby providing a basis for the strategy optimization of the Actor network. Compared with conventional multi-agent algorithms, HAA2C adopts a collaborative mechanism of heterogeneous agents in training. Each agent independently learns strategies according to its different tasks and observation spaces, but they are collaboratively optimized through shared environmental feedback.
[0062] The Joint Advantage Function introduced in HAA2C further enhances the policy optimization effect. The Joint Advantage Function evaluates the improvement of each action relative to other possible actions by calculating the difference between the state-action value function (Q value) and the state value function (V value) of each agent under the current policy.
[0063] The joint advantage function combines global and local feedback, allowing each agent to optimize its strategy based on a global perspective while making decentralized action decisions based on local observations. The introduction of this advantage function provides agents with more precise strategy improvement directions in complex multi-agent scenarios, thereby supporting HAA2C to achieve more efficient collaborative optimization and overall performance improvement during training.
[0064] Figure 3 The formulas and symbols are common knowledge and their meanings will not be repeated here.
[0065] like Figure 4As shown in the figure, a method for intelligent balancing of traffic flow based on deep reinforcement learning and congestion avoidance covers vehicle path planning requests, roadside unit information collection and rerouting decisions, and traffic allocation of cloud servers. At the beginning of system operation, the vehicle about to enter the controlled road network sends a path planning request and vehicle information to the roadside unit RSU in the area, including the current location and destination of the vehicle. The roadside unit RSU is responsible for collecting the real-time road network traffic conditions in the controlled area in real time, and uploading these data to the cloud server to obtain a path allocation strategy based on a global perspective. As two types of intelligent agents in the HAA2C algorithm, the cloud and roadside unit RSU, the cloud agent uses the global real-time road network traffic conditions as the state input of reinforcement learning, and the roadside unit RSU uses the local real-time road network traffic conditions and congestion formation trend of the controlled area as the state input of reinforcement learning and the congestion formation trend as the decision-making basis, respectively obtaining the traffic allocation scheme and local rerouting ratio that meet the optimal principle of the flow balancing principle system.
[0066] During the driving process, the vehicle will dynamically drive according to the traffic distribution plan forwarded by the roadside unit RSU. The roadside unit RSU will continuously adjust the route according to the real-time traffic conditions of the road section to ensure that the vehicles in the road network have the shortest driving time, the lowest energy consumption, and the lowest carbon dioxide emissions, thereby achieving efficient and green traffic management in the entire network.
[0067] The present invention realizes global and local collaborative optimization of intelligent transportation system (ITS) through HAA2C algorithm. The cloud-based agent uses historical data to perform global traffic distribution to balance the overall traffic flow and reduce carbon dioxide emissions and energy consumption. At the same time, the fog-based roadside unit (RSU) agent performs local optimization based on real-time traffic feedback and uses the rerouting ratio as a decision variable to adapt to complex dynamic traffic environments. By converting the traditional vehicle-by-vehicle path adjustment method into a flow-based optimization method, the computational burden is significantly reduced and the resource utilization and scalability of the system are improved.
[0068] In addition, the roadside unit RSU introduces the concept of "congestion formation trend", combining attention mechanism, convolutional neural network (CNN) and long short-term memory network (LSTM) to predict traffic flow changes and identify potential high-risk congestion areas. Before congestion forms, the roadside unit RSU agent can judge whether to take rerouting actions based on predictions to actively select the proportion of diverted vehicles and reduce traffic pressure. The overall optimization goal is to reduce carbon dioxide emissions and energy consumption of the entire network through active congestion avoidance measures, shorten vehicle travel time, thereby achieving efficient and low-carbon intelligent traffic management and promoting the sustainable development of urban transportation systems.
[0069] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.
Claims
1. A traffic flow intelligent balancing system based on deep reinforcement learning and congestion avoidance, characterized by: include: The cloud generates the road network weights at future times based on the real-time road network traffic conditions collected by the roadside units (RSUs). It generates a traffic allocation plan based on the road network weights at future times and the path planning requests and vehicle information of all vehicles entering the road network sent by the roadside units (RSUs). The plan is then returned to the roadside units. In addition, the cloud also provides the road section weights at future times based on the decision requirements sent by the roadside units (RSUs). Roadside units (RSUs) collect route planning requests from vehicles entering the road network and monitor vehicle information within the coverage area in real time, and observe the real-time road network traffic conditions on the road sections in real time. They upload these data to the cloud, receive the traffic allocation plan sent by the cloud, and convey it to the vehicles that initiate route planning requests in the controlled road sections. At the same time, the roadside units (RSUs) build congestion formation trends in real time. When congestion risks are detected, they generate rerouting decisions and congested road section information, obtain road network weights at future times from the cloud, and send the road network weights at future times to the selected vehicles that need to be rerouted. The vehicle unit is used to send a path planning request and vehicle information to the roadside unit RSU, and receive the traffic distribution plan forwarded by the roadside unit RSU to enter the road network. In the road network, if it is selected as a rerouting vehicle, it will recalculate the optimal path to the destination based on the congested road section information sent by the roadside unit RSU and the road section weight at the future time.
2. The intelligent traffic flow balancing system based on deep reinforcement learning and congestion avoidance according to claim 1 is characterized by: The roadside unit RSU is fixedly deployed at the intersection of the road network, and the communication and observation range is limited to all road sections connected to the intersection. The coverage areas are independent of each other, and vehicles traveling in a certain area can only communicate with the roadside unit RSU in the area.
3. The intelligent traffic flow balancing system based on deep reinforcement learning and congestion avoidance according to claim 1 is characterized by: The real-time road network traffic conditions include average travel time, energy consumption, carbon dioxide emissions, and the number of vehicles on the road section.
4. The intelligent traffic flow balancing system based on deep reinforcement learning and congestion avoidance according to claim 1 is characterized by: The vehicle information includes a starting location and a destination.
5. The intelligent traffic flow balancing system based on deep reinforcement learning and congestion avoidance according to claim 1 is characterized by: The roadside unit RSU generates a rerouting decision and congested road section information according to the HAA2C algorithm, and the rerouting decision is to select an optimal proportion of vehicles for rerouting.
6. The intelligent traffic flow balancing system based on deep reinforcement learning and congestion avoidance according to claim 1 is characterized by: The cloud generates a traffic distribution plan based on the HAA2C algorithm. The traffic distribution plan is the number of paths for the vehicle initiating the route planning request to reach the destination and the proportion of vehicles on these paths.
7. A method for intelligent balancing of traffic flow based on deep reinforcement learning and congestion avoidance, characterized by: The specific steps are as follows: Step 1: The vehicle about to enter the controlled road network sends a path planning request and vehicle information to the roadside unit RSU in the area; Step 2: Roadside units (RSUs) collect real-time road network traffic conditions within the coverage area, and upload the real-time road network traffic conditions, path planning requests sent by vehicles, and vehicle information to the cloud; Step 3: The cloud uses the real-time road network traffic conditions to calculate the road section weights at future times through the MTT model, and uses the KSP algorithm to calculate k optimal paths using the road section weights at future times and the path planning request and vehicle information sent by the vehicle. The HAA2C algorithm is used to generate a flow distribution plan that satisfies the optimal principle of the flow balance principle system and the distribution ratio of each path, and the flow distribution plan is fed back to the roadside unit RSU where the initiating vehicle is located; Step 4: The roadside unit (RSU) to which the vehicle that initiates the route planning request belongs forwards the received traffic allocation plan to the vehicles within its jurisdiction to complete the route allocation instruction transmission; Step 5: The vehicles that have obtained the traffic allocation plan enter the road network according to the allocated path; Step 6: The roadside unit (RSU) builds the congestion formation trend in real time according to the CNN-LSTM-Attention model, identifies the impending traffic congestion based on the congestion formation trend, generates rerouting decisions and congested road section information when the congestion risk is detected, obtains the road network weight at the future time from the cloud, and sends the road network weight at the future time to the selected vehicle that needs to be rerouted; Step 7: The vehicle that receives the congestion information uses the shortest path algorithm to recalculate the optimal path to the destination based on the congested road section information provided by the RSU and the road network weight at the future time; Step8: Repeat Step1-Step7.
8. The intelligent traffic flow balancing method based on deep reinforcement learning and congestion avoidance according to claim 7 is characterized by: The specific steps of Step 6 are: The roadside unit RSU inputs its own stored historical traffic flow data into the CNN-LSTM-Attention model to predict the future traffic flow of the controlled section, builds a congestion formation trend based on the historical traffic flow data and the predicted traffic flow, and identifies the impending traffic congestion based on the congestion formation trend. When the congestion risk is detected, the roadside unit RSU selects the optimal proportion of vehicles from the vehicles about to enter the congested section for rerouting based on HAA2C, and passes the section information that may cause congestion and the section weights at future times to the relevant vehicles.
9. The intelligent traffic flow balancing method based on deep reinforcement learning and congestion avoidance according to claim 8 is characterized by: The specific steps of the roadside unit RSU inputting its own stored historical traffic flow data into the CNN-LSTM-Attention model to predict the future traffic flow of the controlled section are: The Attention-CNN-LSTM model uses historical traffic flow data as input to predict the traffic flow of the next time step of the road section. The congestion formation trend is constructed based on the predicted traffic flow, the traffic flow at the previous moment, and the traffic flow at the time of congestion. The trend is constructed through variable differences, including the difference between the predicted value and the value at the previous moment, and the congestion value and the predicted value. If the traffic flow changes accelerate or approaches the congestion value, the congestion trend accelerates. Specifically: First, CNN extracts local temporal features from historical traffic flow data through one-dimensional convolution and captures short-term dependencies at different scales. Subsequently, BiLSTM further extracts long-term dependencies and transfers information bidirectionally to comprehensively characterize the complex temporal patterns of historical traffic flow data, with special attention paid to key time steps to cope with critical changes in traffic fluctuations; Finally, the fully connected layer flattens the features processed by CNN, BiLSTM, and attention mechanism to generate the final traffic flow prediction results.