Signal timing optimization method and system based on vehicle-road cooperation and traffic flow prediction

Abstract

The invention relates to the technical field of traffic signal control, and discloses a signal timing optimization method and system based on vehicle-road cooperation and traffic flow prediction, and the method comprises the steps: obtaining the track and road side detection data of a network-connected vehicle, dynamically estimating the permeability through Kalman filtering, and reconstructing the full traffic state of an intersection; predicting future traffic flow parameters by using a long-short-term memory network based on the total state historical sequence; establishing an optimization model aiming at minimizing delay and queuing, and solving by adopting a deep reinforcement learning algorithm to obtain a signal timing scheme; and issuing and executing the scheme, broadcasting a green light speed guide message, and feeding back online update model parameters based on an execution result. According to the method, the problem of state perception in a low permeability environment is solved, active prediction control of traffic signals and vehicle-road collaborative closed-loop optimization are realized, and the traffic efficiency of the intersection is effectively improved.

Description

A signal timing optimization method and system based on vehicle-road cooperation and traffic flow prediction Technical Field

[0001] This invention relates to the field of traffic signal control technology, specifically to a signal timing optimization method and system based on vehicle-road cooperation and traffic flow prediction. Background Technology

[0002] Urban road intersections are key nodes in the transportation network, and their traffic efficiency directly affects the overall operation of the road network. As the core means of regulating the allocation of time and space resources at intersections, the scientific nature of their timing schemes is crucial for alleviating traffic congestion.

[0003] Existing traffic signal control systems primarily rely on induction coils buried in the road surface or video cameras mounted on poles to collect traffic flow data. These fixed detection devices typically only acquire throughput or occupancy at specific cross-sections, resulting in limited sensing range, blind spots, and high maintenance costs. They also struggle to capture the microscopic driving status of vehicles across the entire intersection. With the development of vehicle-road cooperative technology, connected vehicles can provide high-precision real-time trajectory data. However, due to the current stage of market promotion, the penetration rate of connected vehicles in mixed traffic flows is generally low. Directly using sparse connected vehicle data for signal control can lead to significant deviations in the estimation of overall traffic demand. Furthermore, existing data fusion technologies often struggle to accurately calculate the full traffic status in low-penetration environments, causing the input data of the control system to fail to accurately reflect the actual load at the intersection.

[0004] In terms of control strategies, traditional timing control or sensor-based control methods typically rely on historical statistical data or current detection data for passive responses. This reactive control logic assumes that traffic flow is stable or changes slowly, lacking the ability to predict short-term traffic flow trends. When traffic flow experiences nonlinear abrupt changes or queue overflow risks, the control system often exhibits lag, failing to adjust signal timing in advance to avoid congestion. Furthermore, existing optimization models often employ offline parameter calibration, making it difficult to adapt to the dynamic drift of traffic flow distribution over time. Moreover, the lack of an effective coordination mechanism between the signal control unit and the driver's end means that the signal controller cannot not only fail to correct the control strategy online based on actual vehicle operation feedback but also cannot provide effective speed guidance, leading to frequent starts and stops at intersections, increasing delays and energy consumption. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a signal timing optimization method and system based on vehicle-road cooperation and traffic flow prediction. It solves the problems of limited sensing range of fixed detection methods, difficulty in accurately reconstructing the full traffic state in environments with low penetration of connected vehicles, lack of predictive ability for future traffic flow changes in traditional passive control strategies leading to response lag, and lack of online closed-loop coordination between the signal control end and the vehicle operation end.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a signal timing optimization method based on vehicle-road cooperation and traffic flow prediction, comprising the following steps: acquiring connected vehicle trajectory data and traffic flow statistics data from roadside detection equipment at the target intersection; performing spatiotemporal alignment and fusion of the connected vehicle trajectory data and the traffic flow statistics data to dynamically estimate the connected vehicle penetration rate, and reconstructing the full traffic state of the intersection based on the connected vehicle penetration rate; predicting traffic flow parameters of the target intersection within a future preset time period using a long short-term memory network model based on the historical sequence of the full traffic state; establishing a signal timing optimization model with the objective of minimizing vehicle delays and queue lengths based on the traffic flow parameters; transforming the signal timing optimization model into a Markov decision process, solving the signal timing optimization model using a deep reinforcement learning algorithm to obtain the signal timing scheme for the next control cycle of the target intersection; issuing the signal timing scheme to the signal controller for execution, and feeding back the execution result to the deep reinforcement learning algorithm for model updates.

[0007] The process of spatiotemporally aligning and fusing the connected vehicle trajectory data and traffic flow statistics to dynamically estimate the connected vehicle penetration rate includes: establishing a spatial mapping relationship between the connected vehicle trajectory data and lane topology; statistically analyzing the number of connected vehicles matched to each entrance lane within a preset time window and the total number of vehicles detected by roadside detection equipment; constructing a state-space model for penetration rate estimation, the state-space model including a state equation and an observation equation; the state equation characterizing the current penetration rate as the sum of the previous penetration rate and process noise; the observation equation characterizing the current number of connected vehicles as the sum of the current penetration rate and the total number of vehicles plus observation noise; using a Kalman filter algorithm to predict the current prior penetration rate based on the posterior estimate of the previous time, and correcting the prior penetration rate using the observation bias of the current time to obtain the optimal estimate of the current connected vehicle penetration rate.

[0008] The method of reconstructing the full traffic state of an intersection based on the connected vehicle penetration rate includes: identifying connected vehicles currently queuing in the lane, obtaining the position of the connected vehicle farthest from the stop line, and calculating the queue length of connected vehicles; calculating the ratio of the queue length to the connected vehicle penetration rate to obtain the total queue length; statistically analyzing the connected vehicle traffic flow within a time window, calculating the ratio of the traffic flow to the connected vehicle penetration rate to obtain the total vehicle arrival rate; and combining the total queue length, the total vehicle arrival rate, and the average vehicle speed to generate a full traffic state vector.

[0009] The method of predicting traffic flow parameters of a target intersection within a predetermined future time period using a long short-term memory network model includes: slicing the historical sequence of the full traffic state using a sliding window technique to generate an input sequence vector; constructing a long short-term memory network model, which includes an input layer, several long short-term memory hidden layers, and an output layer; the storage units in the long short-term memory hidden layers are configured with forget gates, input gates, and output gates; the forget gates are used to control the proportion of information retained in the cell state at the previous time step; the input gates are used to control the degree to which new information at the current time step is updated to the cell state; the output gates are used to calculate the hidden state output at the current time step based on the updated cell state; and inputting the input sequence vector into the trained long short-term memory network model to output the predicted vehicle arrival rate and predicted queue length within the predetermined future time period.

[0010] The establishment of a signal timing optimization model aimed at minimizing vehicle delays and queue lengths includes: using the signal cycle duration of the intersection signal controller and the effective green light time of each phase as decision variables; constructing an objective function, which is the sum of the weighted sum of the total vehicle delays of all approach lanes at the intersection and the queue overflow penalty term; constructing a queue overflow penalty function, whereby the queue overflow penalty term increases exponentially with the increase of queue length when the predicted queue length exceeds a preset warning threshold; and setting constraints, including green light time constraints, cycle duration constraints, and saturation constraints; the saturation constraint requires that the ratio of the flow rate ratio to the green light ratio of each phase be less than a preset saturation upper limit.

[0011] The calculation of total vehicle delay employs an improved Webster delay model, specifically including: calculating the flow ratio based on the predicted vehicle arrival rate and saturation flow rate; calculating the green light ratio based on the effective green light time and the signal cycle duration; calculating a first delay term, which represents the uniform delay caused by the red light and is related to the signal cycle duration and the green light ratio; calculating a second delay term, which represents the queuing delay caused by random fluctuations in traffic flow and is related to the flow ratio and the predicted vehicle arrival rate; and using the sum of the first delay term and the second delay term as the average vehicle delay for each phase.

[0012] The method of using a deep reinforcement learning algorithm to solve the signal timing optimization model includes: defining a state space, which includes the current queue length of each approach lane, the vehicle arrival rate in the future preset time period, the current phase index, and the duration of the current phase; defining an action space, which includes actions such as maintaining the current phase and switching to the next phase; defining a reward function, which uses the negative value of the sum of the queue lengths of each approach lane at the intersection at the next moment after the action is executed as the basic reward, and deducts a penalty value when the queue length exceeds a warning threshold; constructing a deep Q-network to fit the state-action value function, using a greedy strategy to select actions, and generating experience tuples through interaction with the environment and storing them in an experience replay pool.

[0013] The method of using a deep reinforcement learning algorithm to solve the signal timing optimization model further includes: setting up a main network and a target network with identical structures; randomly extracting a small batch of sample data from the experience replay pool; calculating a target Q value, wherein the target Q value is the sum of the maximum Q value corresponding to the immediate reward and the next time state in the target network after being weighted by a discount factor; calculating a loss function, wherein the loss function is the mean square error between the estimated Q value calculated by the main network and the target Q value; updating the weight parameters of the main network using a gradient descent algorithm, and periodically synchronizing the weight parameters of the main network to the target network.

[0014] The step of sending the signal timing scheme to the signal controller for execution and feeding back the execution result to the deep reinforcement learning algorithm for model update includes: broadcasting the latest signal phase and timing message to connected vehicles within the communication range through roadside units; calculating the optimal speed guidance interval for green lights: if the current phase is red, calculating the suggested lower speed limit based on the distance from the vehicle to the stop line and the remaining red light time; if the current phase is green, determining whether the time required for a vehicle to pass through the intersection at the current speed is greater than the remaining green light time, and if so, sending a deceleration and stopping suggestion; statistically analyzing the actual cumulative vehicle delay time and the actual maximum queue length after the control cycle ends, and calculating the true reward value; and fine-tuning the network parameters in the deep reinforcement learning algorithm online based on the deviation between the true reward value and the model's predicted Q value.

[0015] A signal timing optimization system based on vehicle-road cooperation and traffic flow prediction includes: a data acquisition module for acquiring connected vehicle trajectory data and traffic flow statistics data from roadside detection equipment at the target intersection; a data processing module for spatiotemporal alignment and fusion of the connected vehicle trajectory data and the traffic flow statistics data, dynamically estimating the connected vehicle penetration rate, and reconstructing the full traffic state of the intersection based on the connected vehicle penetration rate; a traffic flow prediction module for predicting traffic flow parameters of the target intersection within a future preset time period based on the historical sequence of the full traffic state using a long short-term memory network model; a decision optimization module for establishing a signal timing optimization model with the objective of minimizing vehicle delays and queue lengths based on the traffic flow parameters, and solving the signal timing optimization model using a deep reinforcement learning algorithm to obtain the signal timing scheme for the next control cycle of the target intersection; and an execution control module for distributing the signal timing scheme to the signal controller for execution and feeding back the execution results to the decision optimization module for model updates.

[0016] This invention provides a signal timing optimization method and system based on vehicle-road cooperation and traffic flow prediction. It offers the following advantages: 1. By employing multi-source data fusion technology based on Kalman filtering, this invention establishes a spatiotemporal mapping relationship between connected vehicle trajectory data and roadside cross-section detection data. This enables dynamic estimation of the market penetration rate of connected vehicles. Even in mixed traffic environments with low connected vehicle penetration, the system can still accurately reconstruct the total queue length and vehicle arrival rate at intersections using partial observation data. This effectively solves the problem of state perception bias caused by data gaps or limited coverage of single detection methods, improving the comprehensiveness and accuracy of traffic state perception.

[0017] 2. This invention combines the sequence prediction capability of Long Short-Term Memory (LSTM) networks with the decision optimization capability of deep reinforcement learning, realizing the transformation from passive response to active predictive control. By predicting future short-term traffic flow parameters through LSM networks and combining a reward function that includes a queue overflow penalty term, deep Q-networks can adjust signal timing schemes in advance before congestion occurs, effectively avoiding the risk of queue overflow. At the same time, this method has the ability to handle nonlinear and high-dimensional traffic flow characteristics, and is more adaptable to complex and ever-changing traffic flow dynamics compared to traditional modeling methods.

[0018] 3. This invention establishes a closed-loop feedback mechanism that includes optimal green light speed guidance and online model updates. At the execution end, the vehicle-road cooperative network guides vehicles to adjust their speed to match the signal timing, reducing the number of times vehicles stop at intersections and idling fuel consumption, thus achieving longitudinal control optimization at the vehicle-road cooperative level. At the feedback end, the actual reward value is calculated using traffic operation indicators after actual execution, and the network parameters are fine-tuned online. This overcomes the performance degradation problem caused by environmental distribution drift in practical applications of offline trained models, ensuring the long-term robustness of the control strategy. Attached Figure Description

[0019] Figure 1 is a flowchart of the method of the present invention; Figure 2 is a system architecture diagram of the present invention; Figure 3 is a flowchart of the multi-source heterogeneous traffic data acquisition and preprocessing of the present invention.

[0020] The module includes: 10. Data acquisition module; 20. Data processing module; 30. Traffic prediction module; 40. Decision optimization module; and 50. Execution control module. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Embodiments

[0022] Please refer to Figures 1-3. This embodiment of the invention provides a signal timing optimization method based on vehicle-road cooperation and traffic flow prediction, comprising the following steps: S100, acquiring connected vehicle trajectory data and traffic flow statistics data from roadside detection equipment at the target intersection; S200, performing spatiotemporal alignment and fusion of the connected vehicle trajectory data and traffic flow statistics data to estimate the connected vehicle penetration rate and reconstruct the full traffic state of the intersection; S300, based on the historical sequence of the full traffic state, using a long short-term memory network model to predict the traffic flow parameters of the target intersection within a future preset time period; S400, establishing a signal timing optimization model based on the traffic flow parameters, with the objective of minimizing vehicle delays and queue lengths; S500, using a deep reinforcement learning algorithm to solve the signal timing optimization model, obtaining the signal timing scheme for the next control cycle of the target intersection; S600, sending the signal timing scheme to the signal controller for execution, and feeding back the execution results to the deep reinforcement learning algorithm for model updates.

[0023] Referring to Figure 2, the intelligent signal timing optimization system based on vehicle-road cooperation and dynamic traffic flow prediction includes a data acquisition module 10, a data processing module 20, a traffic flow prediction module 30, a decision optimization module 40, and an execution control module 50.

[0024] In step S100, the process of acquiring the trajectory data of connected vehicles at the target intersection and the traffic flow statistics data of the roadside detection equipment may specifically include the following steps: receiving basic safety messages sent by connected vehicles entering the communication range of the target intersection through the vehicle-road cooperative communication network. The trajectory data of connected vehicles comes from the on-board unit installed on the vehicle, and the on-board unit broadcasts the basic safety messages to the roadside unit according to a preset communication frequency through the C-V2X or DSRC communication protocol.

[0025] The basic security message specifically includes the following data fields: the vehicle's temporary identifier, the vehicle's current GPS coordinates, the vehicle's instantaneous speed, the vehicle's heading angle, the vehicle's longitudinal acceleration, the vehicle's lateral acceleration, and vehicle size information. The GPS coordinates include longitude and latitude values; the heading angle indicates the vehicle's current direction of travel; and the vehicle size information is used to subsequently calculate the vehicle's projected area on the road and its headway. The specific encoding format and transmission mechanism of the above communication protocol are well-known technologies in the field and will not be elaborated upon here.

[0026] Traffic flow statistics are collected by roadside sensing devices deployed at the approach lanes of the target intersection. These roadside sensing devices specifically include one or more combinations of LiDAR, millimeter-wave radar, video surveillance cameras, or integrated radar-visual systems. In this embodiment, the roadside sensing devices are installed behind the stop lines and upstream of each approach lane at the intersection.

[0027] Roadside sensing equipment identifies and counts vehicles passing through the detection area, generating traffic flow statistics including: cross-sectional traffic volume, average lane speed, lane occupancy rate, and vehicle queue length within the detection area. Cross-sectional traffic volume refers to the number of vehicles passing through the detection cross-section per unit time; lane occupancy rate is the ratio of the area occupied by vehicles within the detection area to the total area of ​​the detection area, or the ratio of the time vehicles spend passing through the detector to the total time within the detection period.

[0028] Outlier removal is performed on the received connected vehicle trajectory data and traffic flow statistics. Due to positioning drift, sensor noise, or signal transmission interference, outliers may exist in the raw data. For connected vehicle trajectory data, the rate of change of vehicle speed within consecutive time stamps is calculated. If the rate of change exceeds a preset threshold for vehicle physical acceleration, or if the vehicle's position coordinates exceed the road boundary, the data point is considered an outlier and removed. For roadside detection data, if the detected vehicle speed is negative or exceeds a preset multiple of the road's maximum design speed limit, it is considered invalid data.

[0029] Missing values ​​are imputed in the data after outlier removal. When packet loss during data transmission causes discontinuity in the data sequence, if the continuous duration of the missing data is less than a preset time window threshold, linear interpolation or Lagrange interpolation is used to estimate and fill the missing values ​​using valid data points before and after the missing time. If the continuous duration of the missing data exceeds the preset time window threshold, the time period is marked as invalid and will not be included in subsequent calculations.

[0030] Coordinate system transformation is required for the trajectory data of connected vehicles. Since the location information uploaded by the vehicle unit usually adopts the WGS-84 geodetic coordinate system, while signal control and roadside detection are usually based on the local Cartesian coordinate system, a unified spatial reference benchmark needs to be established.

[0031] The geometric center of the target intersection is selected as the origin, with the east direction as the positive X-axis and the north direction as the positive Y-axis. A local Cartesian coordinate system is established. Using Gauss-Kruger projection or Mercator projection algorithms, the longitude and latitude coordinates of the connected vehicles are converted into longitudinal and lateral distance coordinates relative to the origin. Through the above transformation, the trajectory data of all vehicles are mapped to a unified intersection road network topology model, providing a spatial consistency basis for subsequent data fusion and state reconstruction.

[0032] In step S200, the process of spatiotemporal alignment and fusion of connected vehicle trajectory data and traffic flow statistics, estimating connected vehicle penetration rate, and reconstructing the full traffic state of the intersection can specifically include the following sub-steps: establishing a spatial mapping and temporal synchronization relationship between connected vehicle trajectory data and lane topology; based on the local Cartesian coordinate system transformed in the step, combined with the lane centerline data provided by the high-precision map, using the minimum distance projection algorithm from point to curve, matching the real-time position coordinates of connected vehicles to specific approach lanes.

[0033] At the same time, a uniform data aggregation time window length is set. The continuous time axis is divided into discrete time steps. For each time step k and each entrance lane i, the number of connected vehicles matched to that lane within that time period is counted and denoted as m. i (k); Simultaneously, count the total number of all vehicles detected by the roadside sensing equipment in this lane, denoted as N. i (k) If the data update frequency of the roadside sensing device and the vehicle unit is inconsistent, the connected vehicle data is resampled or interpolated based on the detection cycle of the roadside sensing device to complete the time alignment of the data.

[0034] The Kalman filter algorithm is used to dynamically estimate the penetration rate of connected vehicles at each entrance lane. The penetration rate of connected vehicles is defined as the proportion of vehicles equipped with on-board communication units in the current traffic flow. Due to the randomness of traffic flow, the penetration rate fluctuates dynamically over time. A state-space model for penetration rate estimation is constructed, including state equations and observation equations.

[0035] set up To determine the true permeability of the i-th inlet channel at the k-th time step, the state equation is established as follows:

[0036] in, The penetration rate at the previous moment. This is the process noise, used to characterize the random drift of permeability over time, and is assumed to follow a Gaussian distribution with a mean of 0 and a variance of Q.

[0037] The observation equations are established as follows:

[0038] Where, m i (k) represents the number of connected vehicle observations obtained through vehicle-road cooperative communication, N i (k) represents the total number of vehicles detected by the roadside equipment, and v(k) represents the observation noise, which is used to characterize the counting error. It is assumed that v(k) follows a Gaussian distribution with a mean of 0 and a variance of R.

[0039] Based on the above model, the prediction and update steps of Kalman filtering are iteratively calculated. First, the prior permeability at the current time is predicted based on the posterior estimate of the previous time step. Then, the Kalman gain is calculated, and the prior value is corrected using the observation bias at the current time step to obtain the optimal estimate of the permeability at the current time step. This process can effectively smooth out fluctuations in instantaneous data and provide more stable penetration rate parameters.

[0040] Based on the estimated penetration rate and connected vehicle status data, the full traffic status of the intersection is reconstructed, including the real-time queue length, vehicle arrival rate and average speed of each approach lane.

[0041] To reconstruct the queue length, the first step is to identify connected vehicles currently queuing in the lane. Queuing is defined as a vehicle speed being less than a preset queuing speed threshold. The position of the connected vehicle furthest from the stop line among all queuing vehicles is then obtained, and the distance from that position to the stop line is calculated and denoted as the connected vehicle queue length. .

[0042] Using estimated penetration rate Adjust the queue length and estimate the total queue length. The calculation formula is as follows:

[0043] When calculated When the confidence level is below the preset threshold, it indicates that the number of connected vehicle samples is too small. In this case, lane occupancy data detected by roadside sensing devices is used, and a historically calibrated occupancy-queue length conversion model is used to assist in correction. .

[0044] For the reconstruction of vehicle arrival rate, the arrival traffic of connected vehicles is amplified and restored using penetration rate. For the connected vehicle traffic detected within the time window, the total vehicle arrival rate is... The calculation is as follows:

[0045] Through the above steps, the sparse network data observed in some areas are fused with roadside statistical data to generate a full traffic state vector containing queue length and flow information for all approach lanes. This vector will serve as the input feature for subsequent deep learning models.

[0046] In step S300, based on the historical sequence of the full traffic state, the process of predicting the traffic flow parameters of the target intersection in the future preset time period using a long short-term memory network model may specifically include the following steps: constructing a time series sample set for model training and inference; extracting key feature data from the full traffic state reconstructed in step S200; the key feature data includes the full vehicle arrival rate, full queue length and average driving speed of each approach lane; and using the sliding window technique to slice the continuous time series data to generate an input sequence vector and a target label vector.

[0047] Given a historical time window length of T and a prediction time step of P, construct the input feature matrix X for any time t. t This matrix contains historical traffic state data from time t-T+1 to time t, and the corresponding target label vector Y. t It contains the true values ​​of future traffic flow parameters from time t+1 to time t+P. Before being fed into the network model, the input data is mapped to a preset numerical range, such as the [0,1] range, using the max-min normalization method to eliminate the influence of data with different dimensions on the model convergence speed.

[0048] The network topology of the Long Short-Term Memory (LSTM) network model is constructed. The model consists of an input layer, several stacked LSM hidden layers, a fully connected layer, and an output layer connected in sequence. The input layer is responsible for receiving the processed time series samples. The LSM hidden layers are used to capture the long-term dependencies and nonlinear time-varying features in the traffic flow data. The fully connected layers are used to map the high-dimensional feature space of the hidden layers to the dimension space of the target prediction value. The output layer outputs the predicted traffic flow parameters for the future preset time period.

[0049] The Long Short-Term Memory (LSTM) hidden layer consists of several LSM storage units. Each LSM storage unit is equipped with a cell state transmission channel and three gating mechanisms: a forget gate, an input gate, and an output gate. During the forward propagation of the model, the LSM storage unit performs the following logical operations: the forget gate receives the hidden state from the previous time step and the input data from the current time step, performs weighted summation and activation function operations, and generates a forgetting coefficient between 0 and 1. This coefficient is used to control the proportion of information retained in the cell state from the previous time step, thereby filtering out non-critical historical interference information.

[0050] The input gate also receives the hidden state from the previous time step and the input data from the current time step. It calculates the degree to which the new information from the current time step updates the cell state and generates candidate cell states for the current time step. The old cell state after being processed by the forget gate is superimposed with the candidate cell state after being weighted by the input gate to complete the update of the cell state for the current time step.

[0051] The output gate calculates the hidden state output at the current time step based on the updated cell state, the hidden state at the previous time step, and the input data at the current time step. This hidden state is passed to the next layer of the network as the output feature of the current time step, and also participates in the recursive calculation as the input state of the next time step. Through the synergistic effect of the above gating mechanism, the model can effectively memorize the long-term traffic flow evolution pattern and suppress the gradient vanishing problem.

[0052] The Long Short-Term Memory (LSTM) network model is trained offline and inferred online. The loss function is defined as the mean square error between the predicted traffic flow parameters and the actual observed values. An adaptive moment estimation optimization algorithm or a stochastic gradient descent algorithm is used to calculate the gradient of the loss function with respect to the network weights through the backpropagation mechanism, and the network weight parameters are iteratively updated until the loss function value converges to below the preset threshold.

[0053] During the online inference phase, the traffic state data from the most recent T time steps, which are collected and processed in real time, are input into the trained model. The model directly outputs the predicted vehicle arrival rate and predicted queue length for the next P time steps. The prediction results can reflect the traffic flow change trend in the short term and provide a forward-looking state input for subsequent signal timing optimization, thereby overcoming the lag problem caused by traditional signal control relying solely on current detection data.

[0054] In step S400, the process of establishing a signal timing optimization model with the goal of minimizing vehicle delays and queue lengths based on the predicted traffic flow parameters can specifically include the following steps: defining the decision variables and objective function for signal timing optimization, and using the timing parameters of the intersection signal controller as decision variables, specifically including the signal cycle duration and the effective green light time of each phase within the cycle.

[0055] An objective function is established to improve the traffic efficiency of intersections and prevent deadlocks. The objective function consists of two parts: the first part is the weighted sum of the total vehicle delays of all approach lanes of the intersection, which is used to characterize the average time cost of vehicles passing through the intersection; the second part is the queue overflow penalty term, which is used to characterize the risk of the queue length exceeding the lane capacity. The optimization objective is to find a set of optimal decision variables that minimizes the calculated value of the above objective function.

[0056] A vehicle delay calculation model based on predicted traffic flow is constructed. Using the predicted traffic flow output in step S300, the average vehicle delay for each phase is calculated using an improved Webster delay model.

[0057] A queue overflow penalty function is constructed to prevent vehicles queuing at the entrance lane from extending to the upstream intersection or blocking the entrance of the auxiliary lane. The penalty function is defined to apply a nonlinear penalty to the long queue state.

[0058] The system obtains the lane length capacity of the corresponding entrance lane and the current predicted queue length. When the predicted queue length is less than the preset warning threshold, the penalty term is set to zero. When the predicted queue length exceeds the warning threshold, the penalty term increases exponentially with the increase of queue length. This mechanism ensures that green light time is prioritized for phases with high queuing risk during the optimization process, thereby avoiding queue overflow.

[0059] To ensure the feasibility and safety of the timing scheme in practical engineering applications, the decision variables must meet the following physical and traffic engineering constraints: First, green light time constraints. The effective green light time of each phase must be between the minimum green light time and the maximum green light time. The minimum green light time is determined based on the shortest time required for pedestrians to cross the street safely, and the maximum green light time is set based on the driver's psychological tolerance to prevent excessively long waiting times for red lights in the same phase.

[0060] Second, the cycle duration constraint: the sum of the green light times of all phases plus the green light interval time of each phase should be equal to the signal cycle duration. At the same time, the cycle duration must be limited to the minimum and maximum cycle allowed by the signal controller.

[0061] Third, saturation constraints: to prevent the intersection from entering an oversaturated state and causing control failure, the actual saturation of each phase must be controlled within the preset upper limit of saturation.

[0062] In step S500, a deep reinforcement learning algorithm is used to solve the signal timing optimization model to obtain the signal timing scheme for the next control cycle of the target intersection. Specifically, this may include the following steps: Step S501, the signal timing optimization problem is transformed into a Markov decision process, defining the state space, action space, and reward function. For the state space, a combination of variables that can characterize the real-time operation characteristics of the intersection is selected, specifically including: the current queue length of each approach lane reconstructed in step S200, the vehicle arrival rate in the future preset time period predicted in step S300, the phase state index of the traffic light at the current moment, and the duration of the green light that has been sustained for the current phase. The above variables together constitute a high-dimensional state vector, which serves as the environmental perception input for the agent.

[0063] For the action space, a discretized set of control operations is defined. The action set includes two basic actions: maintaining the current phase and switching to the next phase. The action of maintaining the current phase means extending the green light time of the current phase by a preset time step. The action of switching to the next phase means ending the green light of the current phase, starting the yellow light and all-red time, and then executing the next phase in the phase sequence.

[0064] Regarding the reward function, in order to achieve the optimization goal of minimizing vehicle delays and queue lengths, the reward value is set as a feedback of the overall operational efficiency of the intersection. The reward function calculation logic is as follows: obtain the sum of the queue lengths of each approach lane of the intersection at the next moment after performing a certain action, take the negative value of the sum of the queue lengths as the basic reward, and if the queue length of any approach lane exceeds the preset warning threshold at the next moment, a fixed penalty value is deducted in addition to the basic reward, thereby guiding the agent to avoid the high-risk state of queue overflow.

[0065] A deep Q-network is constructed as the decision-making core of the agent. The deep Q-network is used to fit the state-action value function, that is, to evaluate the long-term cumulative expected return that can be obtained by taking a specific timing action under a specific traffic state. The network structure includes an input layer, several fully connected hidden layers, and an output layer. The number of nodes in the input layer is consistent with the dimension of the state vector. The fully connected hidden layers use rectified linear units as activation functions to extract the nonlinear features of the traffic state. The number of nodes in the output layer is consistent with the dimension of the action space, and each node outputs the Q value of the corresponding action.

[0066] A greedy strategy is adopted for action selection and interaction. At each decision time, the agent generates a random number between 0 and 1. If the random number is less than the preset exploration rate, the agent randomly selects an action from the action space to maintain the ability to explore unknown strategies. If the random number is greater than or equal to the exploration rate, the agent inputs the current state vector into the deep Q network and selects the action with the largest output Q to execute, so as to utilize the currently learned optimal strategy. After the selected action is executed, the environmental state changes and the corresponding immediate reward is fed back.

[0067] An experience replay mechanism and a dual-network training architecture are established. An experience replay pool with a fixed capacity is initialized to store historical transformation samples. Each sample contains the current state, the action performed, the immediate reward, and the state at the next moment. During training, a small batch of sample data is randomly drawn from the experience replay pool to eliminate the temporal correlation between sample sequences.

[0068] Meanwhile, two identical neural networks are set up: the main network is used to calculate the Q value of the current action in real time, and the target network is used to calculate the target Q value. The calculation logic of the target Q value is: the immediate reward and the maximum Q value corresponding to the next state in the target network are weighted by a discount factor and then added together. The discount factor is used to adjust the degree of importance attached to the future reward.

[0069] The network parameters are iteratively updated using the backpropagation algorithm. The loss function is defined as the mean square error between the estimated Q value calculated by the main network and the target Q value calculated by the target network in a mini-batch of samples. The gradient descent algorithm is used to calculate the gradient of the loss function with respect to the weight parameters of the main network, and the weights of the main network are adjusted along the gradient descent direction to minimize the prediction error. Every preset fixed number of iterations, the weight parameters of the main network are copied and synchronized to the target network to ensure the stability of the training process. Through continuous interaction and iteration, the agent gradually converges to the optimal signal control strategy and outputs a signal timing scheme that maximizes long-term returns.

[0070] In step S600, the signal timing scheme is sent to the signal controller for execution, and the execution result is fed back to the deep reinforcement learning algorithm for model update. Specifically, this may include the following steps: converting the optimized signal timing scheme into a standard instruction format that conforms to the communication protocol of the traffic signal controller. Since the output of the deep reinforcement learning algorithm is usually an abstract action index or a numerical green ratio parameter, the execution control module needs to consult the pre-stored communication protocol dictionary and map the above parameters to specific control object identifiers.

[0071] If the target signal controller supports the NTCIP standard protocol or the GB / T20999 standard protocol, the timing scheme will be encapsulated into the corresponding phase control object data packet, specifically setting the maximum green light time, minimum green light time and forced switching time point for each phase. If the signal controller only supports pulse control, the action command will be converted into the level trigger signal of the corresponding phase pin.

[0072] The converted commands are sent to the signal controller at the target intersection via industrial Ethernet, serial communication, or wireless communication links. Upon receiving the commands, the signal controller overwrites or temporarily stores the currently running signal timing scheme and officially loads the new timing scheme at the next phase switching point or cycle start point. The system monitors the signal controller's operation log in real time to confirm the command response status. If a communication timeout or command rejection occurs, it automatically rolls back to the locally preset timing control or sensor control fallback scheme to ensure basic traffic safety at the intersection.

[0073] By broadcasting the latest signal phase and timing information, along with map topology information, to connected vehicles within the communication range via roadside units, the upcoming signal timing scheme is resolved into the remaining red light time or remaining green light time for each approach lane. Combined with the green wave speed suggestion algorithm, optimal green light speed guidance information is generated. After receiving this information, the on-board terminal of the connected vehicle calculates and prompts the driver to drive at a specific speed, thereby enabling the vehicle to pass through the intersection without stopping. This achieves longitudinal speed control at the vehicle-road cooperative level, further reducing fuel consumption and emissions.

[0074] The system monitors the actual traffic operation indicators after the implementation of the timing scheme in real time, calculates the actual reward value, and continuously collects the actual vehicle trajectory and cross-sectional flow at the intersection through the data fusion module during and after the implementation of the timing scheme. It calculates the actual average vehicle delay time and actual queue length within the control cycle, and substitutes them into the reward function defined in the step to calculate the actual environmental feedback reward value corresponding to the action.

[0075] Based on the deviation between the actual reward value and the expected Q value, the deep reinforcement learning model is fine-tuned online. The current state vector, the executed action, the calculated real reward value, and the new state at the next time step are combined to form a new experience tuple, which is then stored in the experience replay pool.

[0076] The system periodically or when it detects that the prediction error exceeds a preset threshold, it starts an online learning program to extract a training set containing the latest samples from the experience replay pool and perform backpropagation updates. This process enables the agent to adapt to the long-term changing trend of traffic flow, correct the model bias caused by the inconsistency between the distribution of offline training data and the actual scene, and ensure that the signal control strategy is always maintained at the optimal level.

Claims

1. A signal timing optimization method based on vehicle-road cooperation and traffic flow prediction, characterized in that, Includes the following steps: Obtain the trajectory data of connected vehicles at the target intersection and the traffic flow statistics data of roadside detection equipment; The connected vehicle trajectory data and the traffic flow statistics are spatiotemporally aligned and fused to dynamically estimate the connected vehicle penetration rate, and the full traffic state of the intersection is reconstructed based on the connected vehicle penetration rate. Based on the historical sequence of the full traffic conditions, a long short-term memory network model is used to predict the traffic flow parameters of the target intersection in the future preset time period. According to the traffic flow parameters, a signal timing optimization model is established with the goal of minimizing vehicle delays and queue lengths. The signal timing optimization model is transformed into a Markov decision process, and a deep reinforcement learning algorithm is used to solve the signal timing optimization model to obtain the signal timing scheme for the next control cycle of the target intersection. The signal timing scheme is sent to the signal controller for execution, and the execution result is fed back to the deep reinforcement learning algorithm for model updates.

2. The signal timing optimization method based on vehicle-road cooperation and traffic flow prediction according to claim 1, characterized in that, The process of spatiotemporally aligning and fusing the connected vehicle trajectory data and traffic flow statistics to dynamically estimate the connected vehicle penetration rate includes: establishing a spatial mapping relationship between the connected vehicle trajectory data and lane topology; statistically analyzing the number of connected vehicles matched to each entrance lane within a preset time window and the total number of vehicles detected by roadside detection equipment; constructing a state-space model for penetration rate estimation, the state-space model including a state equation and an observation equation; the state equation characterizing the current penetration rate as the sum of the previous penetration rate and process noise; the observation equation characterizing the current number of connected vehicles as the sum of the current penetration rate and the total number of vehicles plus observation noise; using a Kalman filter algorithm to predict the current prior penetration rate based on the posterior estimate of the previous time, and correcting the prior penetration rate using the observation bias of the current time to obtain the optimal estimate of the current connected vehicle penetration rate.

3. The signal timing optimization method based on vehicle-road cooperation and traffic flow prediction according to claim 2, characterized in that, The method of reconstructing the full traffic state of an intersection based on the connected vehicle penetration rate includes: identifying connected vehicles currently queuing in the lane, obtaining the position of the connected vehicle farthest from the stop line, and calculating the queue length of connected vehicles; calculating the ratio of the queue length to the connected vehicle penetration rate to obtain the total queue length; statistically analyzing the connected vehicle traffic flow within a time window, calculating the ratio of the traffic flow to the connected vehicle penetration rate to obtain the total vehicle arrival rate; and combining the total queue length, the total vehicle arrival rate, and the average vehicle speed to generate a full traffic state vector.

4. The signal timing optimization method based on vehicle-road cooperation and traffic flow prediction according to claim 1, characterized in that, The method of predicting traffic flow parameters of a target intersection within a predetermined future time period using a long short-term memory network model includes: slicing the historical sequence of the full traffic state using a sliding window technique to generate an input sequence vector; constructing a long short-term memory network model, which includes an input layer, several long short-term memory hidden layers, and an output layer; the storage units in the long short-term memory hidden layers are configured with forget gates, input gates, and output gates; the forget gates are used to control the proportion of information retained in the cell state at the previous time step; the input gates are used to control the degree to which new information at the current time step is updated to the cell state; the output gates are used to calculate the hidden state output at the current time step based on the updated cell state; and inputting the input sequence vector into the trained long short-term memory network model to output the predicted vehicle arrival rate and predicted queue length within the predetermined future time period.

5. The signal timing optimization method based on vehicle-road cooperation and traffic flow prediction according to claim 1, characterized in that, The establishment of a signal timing optimization model aimed at minimizing vehicle delays and queue lengths includes: using the signal cycle duration of the intersection signal controller and the effective green light time of each phase as decision variables; constructing an objective function, which is the sum of the weighted sum of the total vehicle delays of all approach lanes at the intersection and the queue overflow penalty term; constructing a queue overflow penalty function, whereby the queue overflow penalty term increases exponentially with the increase of queue length when the predicted queue length exceeds a preset warning threshold; and setting constraints, including green light time constraints, cycle duration constraints, and saturation constraints; the saturation constraint requires that the ratio of the flow rate ratio to the green light ratio of each phase be less than a preset saturation upper limit.

6. The signal timing optimization method based on vehicle-road cooperation and traffic flow prediction according to claim 5, characterized in that, The calculation of total vehicle delay employs an improved Webster delay model, specifically including: calculating the flow ratio based on the predicted vehicle arrival rate and saturation flow rate; calculating the green light ratio based on the effective green light time and the signal cycle duration; calculating a first delay term, which represents the uniform delay caused by the red light and is related to the signal cycle duration and the green light ratio; calculating a second delay term, which represents the queuing delay caused by random fluctuations in traffic flow and is related to the flow ratio and the predicted vehicle arrival rate; and using the sum of the first delay term and the second delay term as the average vehicle delay for each phase.

7. The signal timing optimization method based on vehicle-road cooperation and traffic flow prediction according to claim 1, characterized in that, The method of using a deep reinforcement learning algorithm to solve the signal timing optimization model includes: defining a state space, which includes the current queue length of each approach lane, the vehicle arrival rate in the future preset time period, the current phase index, and the duration of the current phase; defining an action space, which includes actions such as maintaining the current phase and switching to the next phase; defining a reward function, which uses the negative value of the sum of the queue lengths of each approach lane at the intersection at the next moment after the action is executed as the basic reward, and deducts a penalty value when the queue length exceeds a warning threshold; constructing a deep Q-network to fit the state-action value function, using a greedy strategy to select actions, and generating experience tuples through interaction with the environment and storing them in an experience replay pool.

8. The signal timing optimization method based on vehicle-road cooperation and traffic flow prediction according to claim 1, characterized in that, The method of using a deep reinforcement learning algorithm to solve the signal timing optimization model further includes: setting up a main network and a target network with identical structures; randomly extracting a small batch of sample data from the experience replay pool; calculating a target Q value, wherein the target Q value is the sum of the maximum Q value corresponding to the immediate reward and the next time state in the target network after being weighted by a discount factor; calculating a loss function, wherein the loss function is the mean square error between the estimated Q value calculated by the main network and the target Q value; updating the weight parameters of the main network using a gradient descent algorithm, and periodically synchronizing the weight parameters of the main network to the target network.

9. The signal timing optimization method based on vehicle-road cooperation and traffic flow prediction according to claim 1, characterized in that, The step of sending the signal timing scheme to the signal controller for execution and feeding back the execution result to the deep reinforcement learning algorithm for model update includes: broadcasting the latest signal phase and timing message to connected vehicles within the communication range through roadside units; calculating the optimal speed guidance interval for green lights: if the current phase is red, calculating the suggested lower speed limit based on the distance from the vehicle to the stop line and the remaining red light time; if the current phase is green, determining whether the time required for a vehicle to pass through the intersection at the current speed is greater than the remaining green light time, and if so, sending a deceleration and stopping suggestion; statistically analyzing the actual cumulative vehicle delay time and the actual maximum queue length after the control cycle ends, and calculating the true reward value; and fine-tuning the network parameters in the deep reinforcement learning algorithm online based on the deviation between the true reward value and the model's predicted Q value.

10. A signal timing optimization system based on vehicle-road cooperation and traffic flow prediction, wherein the signal timing optimization method based on vehicle-road cooperation and traffic flow prediction according to any one of claims 1-9 is characterized in that, include: The data acquisition module is used to acquire the trajectory data of connected vehicles at the target intersection and the traffic flow statistics data of the roadside detection equipment; The data processing module is used to perform spatiotemporal alignment and fusion of the connected vehicle trajectory data and the traffic flow statistics data, dynamically estimate the connected vehicle penetration rate, and reconstruct the full traffic state of the intersection based on the connected vehicle penetration rate. The traffic flow prediction module is used to predict the traffic flow parameters of the target intersection within a future preset time period based on the historical sequence of the full traffic state using a long short-term memory network model; the decision optimization module is used to establish a signal timing optimization model with the objective of minimizing vehicle delays and queue lengths based on the traffic flow parameters, and to solve the signal timing optimization model using a deep reinforcement learning algorithm to obtain the signal timing scheme for the next control cycle of the target intersection; the execution control module is used to send the signal timing scheme to the signal controller for execution, and to feed back the execution results to the decision optimization module for model updates.

Images