CN110570672B - Regional traffic signal lamp control method based on graph neural network - Google Patents
Regional traffic signal lamp control method based on graph neural network Download PDFInfo
- Publication number
- CN110570672B CN110570672B CN201910881544.6A CN201910881544A CN110570672B CN 110570672 B CN110570672 B CN 110570672B CN 201910881544 A CN201910881544 A CN 201910881544A CN 110570672 B CN110570672 B CN 110570672B
- Authority
- CN
- China
- Prior art keywords
- traffic
- network
- flow
- road network
- graph
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 58
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 20
- 230000009471 action Effects 0.000 claims abstract description 18
- 230000008901 benefit Effects 0.000 claims abstract description 11
- 238000010586 diagram Methods 0.000 claims description 10
- 230000002776 aggregation Effects 0.000 claims description 7
- 238000004220 aggregation Methods 0.000 claims description 7
- 230000001276 controlling effect Effects 0.000 claims description 6
- 230000014509 gene expression Effects 0.000 claims description 5
- 230000033228 biological regulation Effects 0.000 claims description 2
- 230000001105 regulatory effect Effects 0.000 claims description 2
- 239000010881 fly ash Substances 0.000 claims 1
- 230000008859 change Effects 0.000 abstract description 7
- 230000007774 longterm Effects 0.000 abstract description 3
- 238000011156 evaluation Methods 0.000 abstract 1
- 238000004088 simulation Methods 0.000 description 10
- 230000000694 effects Effects 0.000 description 8
- 230000006870 function Effects 0.000 description 8
- 230000008569 process Effects 0.000 description 5
- 238000011217 control strategy Methods 0.000 description 4
- 230000008447 perception Effects 0.000 description 3
- 230000002787 reinforcement Effects 0.000 description 3
- 206010039203 Road traffic accident Diseases 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000006698 induction Effects 0.000 description 2
- 238000012800 visualization Methods 0.000 description 2
- 208000004451 Membranoproliferative Glomerulonephritis Diseases 0.000 description 1
- 241001417517 Scatophagidae Species 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000007792 addition Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 208000011511 primary membranoproliferative glomerulonephritis Diseases 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 239000000523 sample Substances 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000036962 time dependent Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000009423 ventilation Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/07—Controlling traffic signals
- G08G1/08—Controlling traffic signals according to detected number or speed of vehicles
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention provides a regional traffic light control method based on a graph neural network, which is characterized in that a traffic flow predictor and a traffic light controller are trained at the same time, a traffic light controller is helped to generate a new control scheme by using a predicted value of future traffic flow change under a current intervention action of the traffic flow predictor, and evaluation information of the current action value is used for assisting in training the traffic light controller to maximize long-term and short-term benefits of the traffic light control scheme. The traffic flow predictor and the signal lamp controller are both built on the basis of a deep message propagation graph network. The invention can continuously optimize the system to adapt to the changing traffic flow, and improve the smoothness degree of the road network and the traffic efficiency.
Description
Technical Field
The invention belongs to the field of traffic signal lamp control, and particularly relates to a regional traffic signal lamp control method based on a graph neural network.
Background
Traffic light control is a key and challenging reality with the goal of maximizing traffic efficiency in road networks and avoiding possible traffic conflicts within intersections. In recent years, signalized intersections have become one of the most significant bottlenecks in traffic efficiency improvement in urban traffic networks. Therefore, a feasible traffic signal control method capable of automatically learning and adjusting according to the current and future traffic flow conditions is found, the traffic jam can be remarkably relieved, and remarkable economic, environmental and social benefits are brought.
At present, traffic signal control systems widely used in many modern cities, such as SCATS, SCOOT and other systems, traffic signal schemes of these systems are mainly designed by means of traffic signal control algorithms based on statistical historical traffic data. The control strategy of these methods (predefined phase based control strategy) is to change the parameters of each predefined phase, such as the release time, etc. The concept of phase refers to the combination of signal states of each lane in each direction during the phase release time. These methods are neither flexible in changing, time-dependent traffic scenarios, nor are they capable of dynamically adjusting the signal control scheme based on real-time traffic. In addition to this, there are other less used adaptive traffic signal control methods that determine which lanes (phases) should be cleared next second (or time period) based on the signals of the induction coils near the intersection. However, these methods also only intervene in the traffic flow according to the current traffic flow condition, and do not fully utilize the historical traffic data to help optimize and design the control scheme. Meanwhile, the methods are also influenced by the problems that the induction coils depending on the methods are easy to damage, the online rate is unstable and the like.
In order to solve the problems in the conventional traffic control methods, many reinforcement learning-based signal light control methods have been recently proposed. Most of these methods still use a phase switching strategy that determines the next second (or time period) of the passable lane combination for each second (or time period). These methods are more flexible than control strategies using predefined phases, but the frequent, abrupt phase switching of such methods is prone to traffic accidents and severely impacts the driving experience. Therefore, in order to avoid the problems of poor driving experience, traffic accidents and the like caused by sudden phase switching, some reinforcement learning methods based on predefined phases are recently proposed.
However, both the conventional method and the reinforcement learning-based method focus on the independent control of one or a small number of traffic lights, and the possible influence of traffic control behaviors on the surrounding road network is ignored. Meanwhile, the current method is still limited to the learning of historical traffic data and the timely response to the current traffic flow, and the influence on the traffic flow in the future area, which may be brought by traffic control behaviors, is ignored.
Disclosure of Invention
The invention provides a regional traffic signal lamp control method based on a graph neural network, which can continuously optimize a system to adapt to changing traffic flows and improve the smoothness degree and traffic efficiency of a road network.
A regional traffic signal lamp control method based on a graph neural network is characterized by comprising the following steps:
(1) acquiring a current signal control scheme and flow data of a target area road network in a plurality of past periods from a signal lamp control system, wherein the signal control scheme comprises a period length, a phase scheme and release time of each phase; the flow data of the past five cycles is generally selected, and the specific number of cycles can be modified according to the application requirements (each cycle refers to the total time of once execution of each phase);
(2) inputting current phase timing, flow data of a plurality of past periods of each target intersection and a road network connection graph into a traffic flow predictor MPTF constructed based on a depth message propagation graph network MPGNN to obtain traffic flow prediction data of each intersection in each direction in the current period;
(3) inputting a current signal control scheme, flow data of a plurality of past periods of each target intersection, a road network connection graph and traffic flow prediction data obtained in the step (2) into a traffic light controller RTSC constructed based on a depth message propagation graph network MPGNN, and taking each generated phase timing of the current period as an adjusted control scheme; the RTSC constructs a control sub-network for each control intersection, and each control sub-network uses the same MPGNN output as input;
(4) inputting the control scheme regulated in the step (3), flow data of past cycles of each target intersection and a road network connection diagram by using a traffic flow predictor MPTF, and evaluating the value of the regulation action in the step (3);
(5) controlling the road network for one period time by using the control scheme adjusted in the step (3);
(6) collecting current road network flow data from a signal lamp control system, and calculating the benefit of the adjusting scheme in the step (3) by combining the road network flow data before the period starts;
(7) training a traffic flow predictor MPTF by using the road network flow data and the adjustment scheme income collected in the step (6) and combining the value estimation obtained in the step (4);
(8) training a traffic signal lamp controller (RTSC) by using the adjustment scheme gains obtained in the step (6) and the value estimation obtained in the step (4);
(9) and (5) starting the next period, and repeating the steps (1) to (8) every period.
The invention uses the current and historical flow data and the timing scheme to generate the signal lamp phase timing scheme. Simultaneously training a flow predictor and a signal lamp controller, simultaneously controlling a plurality of traffic signal lamps, and cooperatively optimizing each traffic signal control scheme; the traffic signal lamp controller is assisted in training to maximize long-term and short-term benefits of the traffic signal lamp control scheme by predicting a future traffic flow change prediction value under a current intervention action by using a traffic flow predictor to help the traffic signal lamp controller generate a new control scheme and evaluating the value of the new control scheme by using an action value predictor of the traffic flow predictor in an online training mode. The overall algorithm optimizes each step with the goal of minimizing intersection latency.
The traffic flow predictor and the signal lamp controller are both built based on the deep message propagation graph network (MPGNN) provided by the invention.
The MPGNN consists of a plurality of graph neural network layers, and the network input is an input graph consisting of flow values of all nodes on a road network and a road network connection graph representing the connection relation among all nodes; at each layer in the MPGNN, two operations of information propagation and information aggregation are performed on each node on the input graph, and mathematical expressions of the two operations are respectively:
wherein,
is the output of node v undergoing a k-th layer information propagation operation,
is the output of node v undergoing a k-th layer information aggregation operation,
is the flow value of a node v on an input graph, N (v) represents the set of all nodes directly connected to the node v, and MLP represents a multilayer perceptron consisting of three fully-connected neural network layers.
In the step (2), the traffic flow predictor MPTF obtaining the traffic flow prediction data comprises the steps of: firstly, extracting each phase timing in the current signal control scheme into a feature code by using a multilayer fully-connected neural network; then inputting the flow data, road network connection graph and phase timing feature codes of each target intersection in past periods into MPGNN, and extracting to obtain a feature vector of the road network traffic condition of the current region; and finally, inputting the characteristic vector into a future flow predictor to obtain a predicted value of the future flow.
In the step (3), the step of generating the row phase timing by the traffic signal lamp controller RTSC is as follows:
(3-1) inputting the traffic flow prediction data obtained in the step (2), the current phase timing, the flow data of a plurality of past periods of each target intersection and the road network connection diagram into an MPGNN to obtain a feature vector of the road network traffic condition of the current region;
and (3-2) respectively constructing a crossing control sub-network according to each crossing which needs to be controlled, wherein the phase timing output by each sub-network is related to the phase number of the crossing. Specifically, if an intersection has 6 phases, then the sub-network controlling the intersection will generate 6-phase timing, while another sub-network controlling only 2 phases will generate 2-phase timing. Each sub-network uses the feature vector of the road network traffic condition of the current region as input and outputs the mean value and the square difference value of the high-dimensional continuous distribution of the control action of each road junction;
and (3-3) sampling each phase timing from the control action high-dimensional continuous distribution of the corresponding intersection by each sub-network, normalizing each phase timing by using a Softmax function to obtain each phase timing proportion, and multiplying the phase timing proportion by the cycle length to obtain each phase timing length.
In the step (4), the traffic flow predictor MPTF is used in a manner similar to that in the step (2), but the current phase timing is replaced by the new timing generated in the step (3). The method comprises the following specific steps: and (4) after inputting the control scheme adjusted in the step (3), the flow data of each target intersection in a plurality of past periods and the road network connection graph, inputting the obtained current regional road network traffic condition feature vector into an action value predictor to predict the new timing value generated in the step (3).
In the step (6), the benefit expression of the adjusting scheme is as follows:
R(O(t-1),O(t),A(t))＝O(t-1)-O(t)
wherein, O(i)The queuing length of vehicles in all directions at all road junctions of the regional road network in the ith period; a. the(t)The phase timing generated in step (3) before the start of the ith period. The concrete meaning of the income is a queuing length change value on each node of the regional road network, and the finally obtained R is a vector.
In step (7), the loss function used for training the traffic flow predictor is as follows:
wherein, Vθ(O(t-1),A(t)) For the new timing value predicted in step (4),
calculating the predicted flow (the queuing length of the vehicles) in the step (2) | + |L1Is mean absolute valueFor the error function.
In step (8), the objective function used by the training signal lamp controller is:
and N is a set of all intersections, an online training strategy is adopted for training, and all networks are optimized once after each period.
Compared with the prior art, the invention has the following beneficial effects:
1. the method is based on relevant characteristics of regional road network traffic flow mined by a graph neural network, models the evolution process of the road network traffic flow, and can improve the perception capability of a traffic signal lamp controller and a flow predictor on the road network traffic flow change trend.
2. All networks based on the graph neural network provided by the invention only need to use the road network connection graph to indicate the road network topological structure, and the traffic flow transfer weight on the connection edge of each road junction is obtained by network dynamic learning. The method overcomes the defect that other methods based on the convolutional graph network need to use Laplace eigenbase to construct a graph network structure.
3. The traffic predictor provided by the invention can better depict the dynamic change of traffic flow by dynamically modeling the evolution process of the traffic flow of the road network. Tested on a convincing public data set, the effect exceeds other current world best-shown flow prediction methods.
4. The traffic signal lamp controller provided by the invention can control a large number of intersection signal lamps simultaneously. Compared with the prior optimal method, the method can be used for synergistically optimizing all signal lamps so as to achieve the effects of balancing the traffic of the road network and effectively improving the unobstructed degree of the road network. The current traffic flow characteristics extracted based on the graph neural network improve the global perception capability of the traffic signal lamp controller.
5. The invention organically combines flow prediction and traffic signal lamp control, continuously optimizes the system to adapt to the changing traffic flow in an on-line training mode, and improves the smoothness degree and traffic efficiency of a road network. The traffic signal lamp controller improves the perception capability of action future income by using the flow prediction value, and helps to improve the action value of the traffic signal lamp controller. Meanwhile, the prediction of the action value can help to improve the long-term benefit of the action generated by the traffic light controller.
Drawings
FIG. 1 is a schematic flow chart of a regional traffic signal lamp control method based on a graph neural network according to the present invention
FIG. 2 is a schematic structural diagram of a traffic flow predictor MPTF in the present invention;
FIG. 3 is a schematic structural diagram of a traffic signal controller RTSC according to the present invention;
FIG. 4 is a schematic diagram of a simulation road network constructed based on an SUMO simulator in the embodiment of the present invention;
fig. 5 is a road network average speed visualization graph tested under the simulated flow configuration 1 according to the embodiment of the present invention;
fig. 6 is a visualization graph of the average queuing length of the road network tested under the simulated flow configuration 1 according to the embodiment of the present invention.
Detailed Description
The invention will be described in further detail below with reference to the drawings and examples, which are intended to facilitate the understanding of the invention without limiting it in any way.
As shown in FIG. 1, a regional traffic signal lamp control method based on a graph neural network uses a control strategy of pre-defining phases and adjusting timing length of each phase. The present invention uses an on-line training method that learns continuously during operation, comprising the following steps during each cycle (total length of time for each phase to be performed once):
s01 obtains the current signal control scheme and flow indicator data from the signal light control system. The signal control scheme comprises cycle length, phase scheme, release time of each phase and signal lamp structured static data such as GPS positioning, version number and the like. The flow index data comprises the vehicle queue length of each intersection in each direction in the past several periods of the target area road network, and can be replaced by other indexes capable of representing the flow condition according to the application requirement, such as the number of passing vehicles and the like.
S02, inputting the current phase timing, the past flow data of each target crossing in several periods and the road network connection diagram into a traffic flow predictor MPTF constructed based on MPGNN, and obtaining the predicted value of the vehicle queue length of each road crossing in each direction in the current period, wherein the structure of the traffic flow predictor MPTF is shown in figure 2. The MPGNN is composed of a plurality of graph neural network layers, and the network input is an input graph composed of flow values of nodes on a network and a network connection graph representing the connection relation between the nodes. At each layer in the MPGNN, two operations of 'information propagation' and 'information aggregation' are carried out on each node on the input graph, and the mathematical expressions of the two operations are as follows:
wherein,
is the output of node v in the 'information propagation' operation through the k-th layer,
is the output of node v undergoing a k-th layer 'information aggregation' operation,
is the flow value of a node v on an input graph, N (v) represents the set of all nodes directly connected to the node v, and MLP represents a multilayer perceptron consisting of three fully-connected neural network layers. In the 'information propagation' operation, the information of all nodes directly connected to the node v is linearly summed. In 'information aggregation' operation, the output of 'information propagation' operation is added with the node information output by the previous layer, and then the input MLP is obtainedOutput graph of the previous layer.
S03, inputting the current signal control scheme, the flow data of the past several periods of each target intersection, the flow prediction data obtained in S02 and the road network connection diagram into the traffic light controller RTSC constructed based on MPGNN, and generating each phase timing of the current period. The structure of the traffic signal controller RTSC is shown in fig. 3. The traffic light controller RTSC individually constructs a control sub-network for each control intersection, and each control sub-network uses the output of the same MPGNN as input. And the phase timing of each sub-network output is sampled from an independent high-dimensional continuous distribution.
S04 uses the traffic flow predictor to input the adjusted control scheme in S03, the past several periods of flow data of each target crossing and the road network connection graph, obtains the characteristic vector of the current road network condition after passing through the depth propagation graph network MPGN, and then inputs the characteristic vector into the action value predictor composed of three full connection layers to obtain the value prediction value of the adjustment action. There is a ReLU activation function between each of the three fully connected layers.
S05 controls the road network for a period of time using the adjusted control scheme of S03.
S06 collects the current road network flow data from the signal lamp control system again, and combines the road network flow data before the period starts to calculate the benefit of the adjusting scheme in S03. The earnings are calculated as follows:
R(O(t-1),O(t),A(t))＝O(t-1)-O(t)
wherein O is(i)The queuing length of vehicles in each direction at each road junction of the regional road network in the ith period, A(t)
The phase timing generated in S03 before the start of the i-th cycle. The concrete meaning of the income is a queuing length change value on each node of the regional road network, and the finally obtained R is a vector.
S07 trains a traffic flow predictor by using the road network flow data collected in S06 and the calculated profit and combining the value estimation obtained in S04. The loss function used to train the traffic flow predictor is:
wherein Vθ(O(t-1),A(t)) For the new dispensing price predicted in S04,
(ii) predicted flow rate (vehicle queue length) obtained in S02 | + | ventilationL1As a function of the mean absolute error.
S08 trains the traffic light controller using the profit obtained in S06 and the value estimate obtained in S04. The objective function used by the training signal controller is:
where N is the set of all intersections. The present invention uses an online training strategy, and all networks are optimized once after each cycle.
S09 begins the next cycle, repeating S01 through S09.
In order to verify the effectiveness of the invention in improving the operation efficiency of a traffic network, a simulation road network with 21 intersections and 72 roads is constructed on an SUMO simulator, and a schematic diagram of the simulation road network is shown in fig. 4. A challenging simulation traffic flow is generated according to a real traffic flow rule, and the effect of the method is compared with the effect of the most effective similar method when the method is used for simultaneously controlling 21 intersections. The SUMO simulator is generally called Simulation of Urban Mobility, and is a traffic Simulation software developed by Institute of Transportation Systems at the German Aerospace Center, which is currently the most commonly used in the traffic field, and is known for its Simulation effect close to reality. We generated three configured traffic flows in the SUMO simulator to verify the effect of the method in different scenarios. The three configured flows are shown in table 1 below:
TABLE 1
Wherein the main trend of the traffic flow represents the trend of the traffic flow on the route in the road network during the period of time, for example, the west to the east represents that the traffic flow mostly moves from the west to the east in the road network, thereby simulating a tidal traffic flow scene. The vehicle arrival rate determines the magnitude of the traffic flow, and the larger the value, the larger the traffic pressure.
The results of the present invention compared to the most effective of the same type of current methods in three traffic flow configurations are shown in table 2 below:
TABLE 2
Wherein, Traffic Configuration represents the Configuration number in table 1, avg.speed is the average value of the average speed of the road network in the whole simulation process, avg.queue is the average value of the average intersection queuing length of the road network in the whole simulation process, avg.waiting is the average Time of waiting for Traffic (including congestion and waiting for Traffic lights) in the road network of the vehicle in the whole simulation process, and Time Duration is the Time taken for completing the whole simulation. From the results, compared with all selected similar methods, the method (GraphRTSC) has the advantage that each index exceeds the similar method under three traffic flow configurations. Meanwhile, we tested the effect without the predicted flow rate provided by MPTF (graphtrtc-nomppf), which is inferior to the predicted flow rate provided by MPTF (graphtsc) as shown in table 2.
Meanwhile, the average speed and the average queue length of the road network per second under configuration 1 are recorded, and the effects of the method are compared with those of the similar method. As shown in fig. 5 and fig. 6, the present invention (graphtscs) performs better than the same methods in both the average vehicle speed and the average queue length of the road network.
In addition, in order to verify the accuracy and effectiveness of the flow prediction of the flow predictor, a comparison experiment is carried out between a METR-LA data set and the most excellent flow prediction method in the world at present. The METR-LA dataset published by the university of southern california contains traffic information collected from the highway ring probe in los angeles county, including 207 sensors' traffic data from 3/1/2012 to 6/30/2012.
The test set of the experiment is compared with the most effective similar methods DCRNN, STGCN and ST-UNet at present in terms of prediction accuracy rates of 15 minutes, 30 minutes and 60 minutes, and the comparison result is shown in Table 3.
TABLE 3
It is seen from the results that the method of the invention (MPTF) has a significantly higher accuracy on the data set than all of the chosen homogeneous methods.
The embodiments described above are intended to illustrate the technical solutions and advantages of the present invention, and it should be understood that the above-mentioned embodiments are only specific embodiments of the present invention, and are not intended to limit the present invention, and any modifications, additions and equivalents made within the scope of the principles of the present invention should be included in the scope of the present invention.
Claims (7)
1. A regional traffic signal lamp control method based on a graph neural network is characterized by comprising the following steps:
(1) acquiring a current signal control scheme and flow data of a target area road network in a plurality of past periods from a signal lamp control system, wherein the signal control scheme comprises a period length, a phase scheme and release time of each phase;
(2) inputting current phase timing, flow data of a plurality of past periods of each target intersection and a road network connection graph into a traffic flow predictor MPTF constructed based on a depth message propagation graph network MPGNN to obtain traffic flow prediction data of each intersection in each direction in the current period;
(3) inputting a current signal control scheme, flow data of a plurality of past periods of each target intersection, a road network connection graph and traffic flow prediction data obtained in the step (2) into a traffic light controller RTSC constructed based on a depth message propagation graph network MPGNN, and taking each generated phase timing of the current period as an adjusted control scheme; the RTSC constructs a control sub-network for each control intersection, and each control sub-network uses the same MPGNN output as input; the traffic signal lamp controller RTSC generates the row phase timing steps as follows:
(3-1) inputting the traffic flow prediction data obtained in the step (2), the current phase timing, the flow data of a plurality of past periods of each target intersection and the road network connection diagram into an MPGNN to obtain a feature vector of the road network traffic condition of the current region;
(3-2) respectively constructing a crossing control sub-network according to each crossing required to be controlled, wherein the phase timing output by each sub-network is related to the number of phases of the crossing, each sub-network uses the feature vector of the traffic condition of the current regional road network as input, and the output is the mean value and the square difference value of the high-dimensional continuous distribution of the control action of each crossing;
(3-3) each sub-network samples each phase timing from the control action high-dimensional continuous distribution of the corresponding intersection, normalizes each phase timing by using a Softmax function to obtain each phase timing proportion, and multiplies the phase timing proportion by a period length to obtain each phase timing length;
(4) inputting the control scheme regulated in the step (3), flow data of past cycles of each target intersection and a road network connection diagram by using a traffic flow predictor MPTF, and evaluating the value of the regulation action in the step (3);
(5) controlling the road network for one period time by using the control scheme adjusted in the step (3);
(6) collecting current road network flow data from a signal lamp control system, and calculating the benefit of the adjusting scheme in the step (3) by combining the road network flow data before the period starts;
(7) training a traffic flow predictor MPTF by using the road network flow data and the adjustment scheme income collected in the step (6) and combining the value estimation obtained in the step (4);
(8) training a traffic signal lamp controller (RTSC) by using the adjustment scheme gains obtained in the step (6) and the value estimation obtained in the step (4);
(9) and (5) starting the next period, and repeating the steps (1) to (8) every period.
2. The method according to claim 1, wherein the MPGNN comprises a plurality of neural network layers, and the network input is an input graph comprising traffic values of nodes on a network and a network connection graph representing connection relationships between the nodes; at each layer in the MPGNN, two operations of information propagation and information aggregation are performed on each node on the input graph, and mathematical expressions of the two operations are respectively:
wherein,
is the output of node v undergoing a k-th layer information propagation operation,
is the output of node v undergoing a k-th layer information aggregation operation,
is the flow value of a node v on an input graph, N (v) represents the set of all nodes directly connected to the node v, and MLP represents a multilayer perceptron consisting of three fully-connected neural network layers.
3. The regional traffic signal lamp control method based on the graph neural network according to claim 1, wherein in the step (2), the step of obtaining the traffic flow prediction data by the traffic flow predictor MPTF is: firstly, extracting each phase timing in the current signal control scheme into a feature code by using a multilayer fully-connected neural network; then inputting the flow data, road network connection graph and phase timing feature codes of each target intersection in past periods into MPGNN, and extracting to obtain a feature vector of the road network traffic condition of the current region; and finally, inputting the characteristic vector into a future flow predictor to obtain a predicted value of the future flow.
4. The regional traffic signal lamp control method based on the graph neural network as claimed in claim 1, wherein the specific steps of step (4) are as follows: and (4) after inputting the control scheme adjusted in the step (3), the flow data of each target intersection in a plurality of past periods and the road network connection graph, inputting the obtained current regional road network traffic condition feature vector into an action value predictor to predict the new timing value generated in the step (3).
5. The regional traffic signal light control method based on the graph neural network as claimed in claim 1, wherein in the step (6), the benefit expression of the adjusting scheme is as follows:
R(O(t-1),O(t),A(t))＝O(t-1)-O(t)
wherein, O(t)The length of vehicle queue in each direction of each road junction of the regional road network in the t-th period; a. the(t)Timing the phase generated in step (3) before the start of the t-th period.
6. The regional traffic signal lamp control method based on the graph neural network as claimed in claim 1, wherein in the step (7), the loss function used for training the traffic flow predictor is as follows:
wherein, Vθ(O(t-1),A(t)) For the novelty predicted in step (4)The value of the timing is high,
calculating a predicted flow value (|) of non-woven fly ash in step (2)L1As a function of the mean absolute error.
7. The regional traffic signal control method based on the graph neural network as claimed in claim 1, wherein in step (8), the objective function used by the training signal controller is:
and N is a set of all intersections, an online training strategy is adopted for training, and all networks are optimized once after each period.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910881544.6A CN110570672B (en) | 2019-09-18 | 2019-09-18 | Regional traffic signal lamp control method based on graph neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910881544.6A CN110570672B (en) | 2019-09-18 | 2019-09-18 | Regional traffic signal lamp control method based on graph neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110570672A CN110570672A (en) | 2019-12-13 |
| CN110570672B true CN110570672B (en) | 2020-12-01 |
Family
ID=68780815
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910881544.6A Active CN110570672B (en) | 2019-09-18 | 2019-09-18 | Regional traffic signal lamp control method based on graph neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110570672B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110969871B (en) * | 2019-12-18 | 2020-11-24 | 浙江大学 | Intelligent traffic light control system and control method |
| CN111341109B (en) * | 2020-05-19 | 2020-09-11 | 银江股份有限公司 | City-level signal recommendation system based on space-time similarity |
| US20220058944A1 (en) * | 2020-08-24 | 2022-02-24 | Quantela Inc | Computer-based method and system for traffic congestion forecasting |
| CN112330962B (en) * | 2020-11-04 | 2022-03-08 | 杭州海康威视数字技术股份有限公司 | Traffic signal lamp control method and device, electronic equipment and computer storage medium |
| CN112734139B (en) * | 2021-01-28 | 2023-09-29 | 腾讯科技（深圳）有限公司 | Method and device for predicting passage duration, storage medium and electronic equipment |
| CN113053122B (en) * | 2021-03-23 | 2022-02-18 | 成都信息工程大学 | WMGIRL algorithm-based regional flow distribution prediction method in variable traffic control scheme |
| GB2607880A (en) * | 2021-06-11 | 2022-12-21 | Vivacity Labs Ltd | Traffic control system |
| CN114694112B (en) * | 2022-02-22 | 2024-06-21 | 广州文远知行科技有限公司 | Traffic signal lamp identification method and device and electronic equipment |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006331002A (en) * | 2005-05-25 | 2006-12-07 | Omron Corp | Signal controller |
| CN102906800A (en) * | 2010-02-01 | 2013-01-30 | 影视技术集成公司 | System and method for modeling and optimizing the performance of transportation networks |
| US9363783B2 (en) * | 2010-07-09 | 2016-06-07 | Digimarc Corporation | Mobile device positioning in dynamic groupings of communication devices |
| CN109830102A (en) * | 2019-02-14 | 2019-05-31 | 重庆邮电大学 | A kind of short-term traffic flow forecast method towards complicated urban traffic network |
| CN110073301A (en) * | 2017-08-02 | 2019-07-30 | 强力物联网投资组合2016有限公司 | The detection method and system under data collection environment in industrial Internet of Things with large data sets |
-
2019
- 2019-09-18 CN CN201910881544.6A patent/CN110570672B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006331002A (en) * | 2005-05-25 | 2006-12-07 | Omron Corp | Signal controller |
| CN102906800A (en) * | 2010-02-01 | 2013-01-30 | 影视技术集成公司 | System and method for modeling and optimizing the performance of transportation networks |
| US9363783B2 (en) * | 2010-07-09 | 2016-06-07 | Digimarc Corporation | Mobile device positioning in dynamic groupings of communication devices |
| CN110073301A (en) * | 2017-08-02 | 2019-07-30 | 强力物联网投资组合2016有限公司 | The detection method and system under data collection environment in industrial Internet of Things with large data sets |
| CN109830102A (en) * | 2019-02-14 | 2019-05-31 | 重庆邮电大学 | A kind of short-term traffic flow forecast method towards complicated urban traffic network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110570672A (en) | 2019-12-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110570672B (en) | Regional traffic signal lamp control method based on graph neural network | |
| CN111696370B (en) | Traffic light control method based on heuristic deep Q network | |
| CN108847037B (en) | Non-global information oriented urban road network path planning method | |
| CN108197739B (en) | Urban rail transit passenger flow prediction method | |
| CN110032782B (en) | City-level intelligent traffic signal control system and method | |
| CN112216124B (en) | Traffic signal control method based on deep reinforcement learning | |
| CN112419726B (en) | Urban traffic signal control system based on traffic flow prediction | |
| CN109272157A (en) | A kind of freeway traffic flow parameter prediction method and system based on gate neural network | |
| EP3035314A1 (en) | A traffic data fusion system and the related method for providing a traffic state for a network of roads | |
| CN109215355A (en) | A kind of single-point intersection signal timing optimization method based on deeply study | |
| CN109269516B (en) | Dynamic path induction method based on multi-target Sarsa learning | |
| CN112907970B (en) | Variable lane steering control method based on vehicle queuing length change rate | |
| CN113643528A (en) | Signal lamp control method, model training method, system, device and storage medium | |
| Eriksen et al. | Uppaal stratego for intelligent traffic lights | |
| CN113516277B (en) | Internet intelligent traffic path planning method based on road network dynamic pricing | |
| CN112863182A (en) | Cross-modal data prediction method based on transfer learning | |
| CN115171408B (en) | Traffic signal optimization control method | |
| CN116895158A (en) | Urban road network traffic signal control method based on multi-agent Actor-Critic and GRU | |
| Kao et al. | A self-organizing map-based adaptive traffic light control system with reinforcement learning | |
| CN113392577B (en) | Regional boundary main intersection signal control method based on deep reinforcement learning | |
| CN114120670A (en) | Method and system for traffic signal control | |
| Li et al. | Construction of Intelligent Transportation Information Management System Based on Artificial Intelligence Technology | |
| CN113724507A (en) | Traffic control and vehicle induction cooperation method and system based on deep reinforcement learning | |
| CN110021168B (en) | Grading decision method for realizing real-time intelligent traffic management under Internet of vehicles | |
| CN116758767A (en) | Traffic signal lamp control method based on multi-strategy reinforcement learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |