CN112926768A - Ground road lane-level traffic flow prediction method based on space-time attention mechanism - Google Patents
Ground road lane-level traffic flow prediction method based on space-time attention mechanism Download PDFInfo
- Publication number
- CN112926768A CN112926768A CN202110114317.8A CN202110114317A CN112926768A CN 112926768 A CN112926768 A CN 112926768A CN 202110114317 A CN202110114317 A CN 202110114317A CN 112926768 A CN112926768 A CN 112926768A
- Authority
- CN
- China
- Prior art keywords
- lane
- traffic flow
- time
- attention mechanism
- value
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 34
- 230000007246 mechanism Effects 0.000 title claims abstract description 24
- 238000012545 processing Methods 0.000 claims abstract description 20
- 230000004927 fusion Effects 0.000 claims abstract description 4
- 238000007781 pre-processing Methods 0.000 claims abstract description 3
- 238000013528 artificial neural network Methods 0.000 claims description 26
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000009466 transformation Effects 0.000 claims description 8
- 238000004140 cleaning Methods 0.000 claims description 4
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 230000009471 action Effects 0.000 claims description 2
- 238000005111 flow chemistry technique Methods 0.000 claims description 2
- 230000000644 propagated effect Effects 0.000 claims description 2
- 238000000844 transformation Methods 0.000 claims description 2
- 238000010219 correlation analysis Methods 0.000 abstract description 2
- 238000011160 research Methods 0.000 description 9
- 238000011144 upstream manufacturing Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 4
- 230000002123 temporal effect Effects 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000007726 management method Methods 0.000 description 2
- 230000002411 adverse Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000001413 cellular effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000006855 networking Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/26—Government or public services
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0125—Traffic data processing
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/065—Traffic control systems for road vehicles by counting the vehicles in a section of the road or in a parking area, i.e. comparing incoming count with outgoing count
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Data Mining & Analysis (AREA)
- Economics (AREA)
- Tourism & Hospitality (AREA)
- Development Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Educational Administration (AREA)
- Analytical Chemistry (AREA)
- Primary Health Care (AREA)
- Chemical & Material Sciences (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Traffic Control Systems (AREA)
Abstract
The method comprises the steps of preprocessing original traffic flow data based on a ground road lane level traffic flow prediction method of a space-time attention mechanism; making a lane selection strategy; carrying out speed processing and self-learning of a space encoder network; and obtaining a final prediction result by tensor fusion and time decoder network self-learning. The invention captures the deep level characteristics of the traffic flow through a space-time attention mechanism network of a coding and decoding structure and efficiently solves the problem of long-distance dependence in a parallel mode. Meanwhile, a specific ground lane selection strategy is verified by a grey correlation analysis method. The traffic flow prediction result of a refined lane level can be obtained, and the traffic management can be better served.
Description
Technical Field
The invention relates to a traffic flow prediction method, in particular to a ground road lane level traffic flow prediction method based on a space-time attention mechanism.
Background
Modern intelligent transportation system ITS combines various advanced information technologies, and applies the advanced information technologies to aspects of transportation, service control, coordination management and the like, thereby promoting the harmonious, efficient and stable development of the relationship among drivers, vehicles, roads and pedestrians. The short-term traffic flow prediction uses historical real data to analyze and learn, and after mastering the general traffic operation rule, the future condition is completed, and the prediction is usually carried out within one hour at time intervals. The intelligent ITS can well meet various requirements in ITS, and has attracted a great deal of attention in recent years, such as travel route guidance, congestion relief and urban road planning. Lane-level forecasts evolved from traditional traffic flow forecasts, which have lane-replacement road segments as the smallest units. The method not only can refine the prediction result, but also can lay a foundation for subsequent advanced applications such as unmanned driving, lane-level high-precision navigation, car networking coordination control and the like.
Lane-level predictions are often overlooked, although one of the future areas of key development in ITS. The main reason is that traffic operation modes between lanes are considered to be similar except that lane occupancy, average speed, flow and the like are difficult to obtain due to lane-level traffic flow characteristic parameters. A great deal of theoretical research and methods prove that the running conditions of all lanes at the same intersection are possibly greatly different, and compared with a road section or a road network, the lanes have weaker anti-interference capability and are more prone to presenting different modes when being influenced by factors such as accidents, severe weather and the like. After fully reading the existing research and method, we find that the existing lane prediction method has certain limitations: firstly, compared with overhead and high speed, the ground lane has smaller traffic flow and relatively slower vehicle running speed, but due to the influence of the complicated adjacency relation among roads in a road network and various timing schemes of densely distributed traffic signal machines, the traffic running mode is more diversified, and timely and effective lane-level fine prediction is urgently needed for better regulation and control management; secondly, during the prediction research, the state of a certain lane node at the next moment is related to the self history situation and the adjacent nodes, and the processing of the time sequence and the space sequence is not a simple parallel or linear synthetic relation. How to deal with the temporal and spatial dependency relationship is still a big problem in research.
Disclosure of Invention
In order to solve the defects of the prior art, a ground road lane-level traffic flow prediction method based on a space-time attention mechanism is provided. The invention builds the coding and decoding neural network embedded with the space-time attention module, and completes the prediction of the future lane running condition through the strong self-learning capability of the network.
In order to achieve the purpose, the invention adopts the following technical scheme:
the ground road lane traffic flow prediction method based on the space-time attention mechanism comprises the following steps:
s1: preprocessing original traffic flow data;
s2: making a lane selection strategy according to the data obtained in the step S1;
s3: carrying out speed processing neural network self-learning on the data obtained in the step S1;
s4: carrying out self-learning of the spatial encoder network on the data obtained in the step S1;
s5: and fusing the tensors obtained in the step S3 and the step S4, and performing self-learning of the time decoder network to obtain a final prediction result.
In the above ground road lane-level traffic flow prediction method based on the space-time attention mechanism, S1 specifically includes the following steps:
s1.1: cleaning lane-level traffic flow parameters, such as missing value removal, outlier data and the like, and carrying out standardized processing on basic data;
s1.2: and drawing a curve according to the cleaned characteristic parameters, and repairing the obvious burr noise points by using a sliding window method with a set length to smooth the burr noise points.
In the above ground road lane-level traffic flow prediction method based on the space-time attention mechanism, S2 specifically includes the following steps:
s2.1: in the method, each urban ground road intersection which is tightly connected is used as a lane dividing and selecting basis, and is verified by using an association degree analysis method, and a corresponding grey association degree calculation formula is as follows:
wherein:
characteristic parameters of the target lane at the time t-k,
characteristic parameters of the participating lanes at the time t-k, u is the value of each item in the sequence, ΔmaxIs the maximum value of the global traffic characteristic parameter, DeltaminIs the minimum value of the global traffic characteristic parameter, rho is an empirical coefficient, and then the obtained correlation value gamma
Normalization processing is carried out to obtain the final result p t-k, u
In the above ground road lane-level traffic flow prediction method based on the space-time attention mechanism, S3 specifically includes the following steps:
s3.1: the lane speed data is arranged into a matrix format required by a speed processing neural network:
wherein:
is the average speed of the i number lane in the j time period, n is the total number of lanes, m is the total number of time periods
S3.2: the speed processing neural network subunit is Bi-LSTM, in the scheme, in order to better capture the deep level characteristics of the vehicle speed sequence, a deep Bi-LSTM network is constructed in a head-to-tail series connection mode:
wherein:
is the average speed value of the lanes in the unit of number l,
and
for the hidden layer and the cell state layer after the unit is propagated in the forward direction, the e parameter represents an encoder used for distinguishing from a decoder; in the same way
And
the hidden layer and the cell state layer after back propagation are obtained by using the positive and negative results of the previous unit, namely the unit l-1 after propagation,
is the hidden layer obtained after the unit is fused.
In the above ground road lane-level traffic flow prediction method based on the space-time attention mechanism, S4 specifically includes the following steps:
s4.1: with built-in spatial attention mechanismThe coding neural network is composed of another Bi-LSTM unit; the spatial attention in the ith cell in a coded neural network can be summarized as: firstly, the initial Query value in the ith unit
Initial Key value
And Value
After linear transformation, respectively obtain
Then will be
And
making dot product, calculating every weight, finally making
Multiplied by the weights of the terms:
wherein: the hidden layer of the i-1 subunit in the neural network of the decoder is
The cell state layer is
For this unit of lane flow data,
for the weight term of the unit of interest,
the parameter is an offset item, q is an identifier of a flow processing parameter, and a is an identifier of an attention mechanism parameter; the intermediate term of the weight can be obtained after the action of the activation function tanh
Weighting and averaging the middle items in the m time periods to obtain a weighting result
In the above ground road lane-level traffic flow prediction method based on the space-time attention mechanism, S5 specifically includes the following steps:
s5.1: fusing the tensors obtained by calculation in S3.2 and S4.1 to obtain input data theta in the time decoder network;
s5.2: and (3) taking the deep space-time feature tensor obtained by fusion in the S5.1 as an input value in a time decoder network, and obtaining a prediction result through a final full-connection layer:
wherein:
for the weight terms in the i' unit of the decoder,
for the offset term within a cell,
and
the hidden layer and the cell state layer of the previous unit are respectively, and the parameter d represents an encoder and is used for distinguishing the encoder;
the merged tensor for the encoder network and the speed processing neural network,
for each timing weight, assign it to
Can obtain the result after distribution
For the historical true value at time t, the value is compared with
Splicing and linear transformation are carried out to obtain
Obtaining the final prediction result after the processing of the full connection layer
Tau is the prediction duration, and the short-time prediction is generally set within one hour;
and
for the weight terms in the two transformations,
and
for the offset term in the transform, f is used to identify various types of parameters in this fully-connected layer.
The innovation of the invention is that:
(1) a space-time attention mechanism network with an encoding and decoding structure is designed for the first time, the deep level characteristics of traffic flow are captured, and the problem of long-distance dependence is efficiently solved in a parallel mode.
(2) Specific ground lane selection strategies are proposed for the network firstly and verified by a grey correlation analysis method.
(3) And obtaining a refined traffic flow prediction result of the lane level.
Drawings
Fig. 1 is a flow chart of a ground road lane-level traffic flow prediction method based on a space-time attention mechanism.
Fig. 2 is a diagram of the overall architecture of a spatiotemporal attention neural network of a codec structure.
Fig. 3 is a plan view of a ground road specific lane selection strategy.
FIG. 4 is a diagram of a speed processing neural network subunit.
Fig. 5 is a diagram of a spatial sequence encoder neural network subunit.
FIG. 6 is a diagram of a time series decoder neural network subunit.
FIG. 7 shows an embodiment of the present invention: 7-month Hangzhou city Xiaoshan road traffic flow lane-level traffic flow prediction result chart in 2017.
FIG. 8 shows an embodiment of the present invention: a prediction result graph of the heart road traffic flow in the Xiaoshan district of Hangzhou city in 7 months in 2017.
Detailed Description
The following steps are further introduced to the embodiment of the present invention in combination with the example of predicting the traffic flow at the level of the lanes in the shore region of Hangzhou city, and the system flow is shown in FIG. 1:
s1: data are extracted on the basis of an original intersection vehicle passing data set, and the steps are as follows:
s1.1: the cleaning of the intersection passing vehicle video data aims at 123 days of Hangzhou city Xiaoshan region data from 1 month 7 in 2017 to 31 month 10 in 2017. The attributes that data possess include:
TABLE 1
| Numbering | Name (R) | |
| 1 | vehicleID | |
| 2 | entryTime | Time to enter |
| 3 | leaveTime | Time of departure from intersection |
| 4 | vehicleType | Type of |
| 5 | cameraID | Camera id |
| 6 | cameraPosition | Location of camera |
| 7 | roadID | Located lane id |
| 8 | turnID | Steering id |
| 9 | memo | Remarks for note |
Because the original data are scattered, the required vehicles are screened out according to the attributes of the two time and the place of the camera and the leaving time of the vehicles, and the wrong data are removed (the leaving time is earlier than the entering time, the license plate number is empty, and the like). This step significantly reduces the amount of data, facilitating subsequent analytical calculations. The further extracted data-related attributes are as follows:
TABLE 2
| Numbering | Name (R) | Note | |
| 1 | vehicleID | |
|
| 2 | entryTime | Time to enter |
|
| 3 | leaveTime | Time of departure from intersection | |
| 4 | | Camera id | |
| 5 | roadID | Located lane id |
After the data are preliminarily extracted, the calculation method of cleaning the average speed by using the average speed is used for calculating the crossing interval distance/driving time length;
the driving time is calculated by the time of entering and leaving the intersection, and then the following attributes are added to the data:
TABLE 3
| Numbering | Name (R) | Note |
| 6 | interval | Length of travel |
| 7 | speed | Average velocity |
S1.2: and (4) slicing the data cleaned in the S1.1 at a time interval of 5 minutes, and counting the average flow and the average driving speed of each lane under each intersection in the time slice. And drawing a curve according to the obtained characteristic parameters, and repairing the obvious burr noise points by using a sliding window lattice method with a set length (12 intervals) of one hour to ensure that the burr noise points are as smooth as possible, thereby improving the accuracy of final prediction.
S2: making a ground lane selection strategy according to the data obtained in the step S1;
in previous lane-level traffic flow prediction studies, the experimental data came from high speed or overhead. In these two types of roads, the traffic flow is large and the vehicle traveling speed is high. Under the condition of the road, the corresponding scheme is generally as follows: when the road is on a high-speed or elevated road, the sensors in the main road and the ramp port are used as separation bases, the long and straight road is divided into a plurality of subsections, and the lanes in the subsections are recorded. The above research background and corresponding solutions often have some disadvantages: firstly, the ramp road surface is narrow in width and is easy to become a bottleneck of smooth traffic in the morning and evening peak with dense traffic flow. According to the scheme, a large amount of incoming and outgoing traffic flows in the ramp are not considered, the traffic running states at all times are considered in an oversimplification mode, and the final experimental result is low in accuracy due to the fact that a large amount of key characteristic parameters are lacked. Secondly, due to the limitation of the research background, the experiment only can use a single straight-going flow lane as a research object, the expansibility is poor, and the actual application range is narrow. In comparison, the ground road has the characteristics of higher road total mileage occupation ratio, wider coverage area, more complex traffic running state and the like. The partitioning scheme shown in fig. 3 is provided herein by comprehensively considering the research strategies for high-speed and overhead lanes in the existing research and the above-mentioned various characteristics of urban road networks. In urban areas, the intervals of all the intersections are short, the association degree is high, and the intersections with close connection are used as the basis for lane division and selection to support the subsequent prediction. The scheme not only takes the adjacent lanes under the same intersection into consideration, but also selects each flow direction lane in the upstream and downstream intersections, wherein the flow direction lanes comprise straight lanes and left-turn lanes and right-turn lanes which are often ignored. The lane division by taking the intersection as an interval is more suitable for a traffic operation mode judging method by taking the intersection as a basic unit in the traditional traffic flow theory. The selection of the upstream and downstream and the participated lanes starts from a physical road connection structure, so that the information of each flow direction lane is perfectly captured, and the experimental prediction precision is improved. Meanwhile, the prediction of multiple flow direction lanes (straight-going, left-turning and right-turning) widens the practical application range of the lane-level traffic flow prediction compared with the prediction of a single straight-going lane.
In order to prove the relevance between the selected lane and the target lane, the method uses the index of grey relevance for analysis:
wherein:
characteristic parameters of the target lane at the time t-k,
for the characteristic parameters of the participating lanes at time t-k,
Δmaxis the maximum value of the global traffic characteristic parameter, DeltaminIs a global minimum, p is an empirical coefficient, and then the correlation value gamma is obtained
Normalization processing is carried out to obtain the final result p t-k, u
The calculation result shows that the relevance of all the selected lanes and the target lane is greater than the previously established threshold value of 0.85, the extremely high relevance is obtained, and the effectiveness of the lane selection scheme is proved.
S3: and (5) carrying out speed processing neural network self-learning on the data obtained in the step S1:
the traffic operation condition of the current road section is closely related to the upstream and downstream road sections. The upstream traffic flows travel toward the downstream link, and the state (for example, congestion) of the downstream link gradually accumulates and spreads in the reverse direction, and adversely affects the upstream. To address this phenomenon, we introduce a BilSTM network. In the direction of time dimension, the BilSTM is formed by overlapping and fusing a forward LSTM sequence and a backward LSTM sequence, and a series of behaviors of forward propagation and backward propagation of sequence data are considered while the problems of gradient disappearance and gradient explosion are solved. Such characteristics are more closely related to the traffic phenomenon that the upstream and downstream road sections mutually influence each other. In order to better capture the deep features of the vehicle speed sequence, a deep neural network is constructed in a head-to-tail series connection mode:
wherein
Is composed of
And
the result of the concatenation, meaning the l subunitA hidden layer formed by combining all the hidden layers
Splicing and transposing the product of l-1, 2, … m to obtain the product
Finally, the linear change is carried out on the hidden layer to obtain a final hidden layer
And
as weight terms and offset terms in the cell
S4: and (4) carrying out self-learning of the spatial encoder network on the data obtained in the step S1:
when traffic flow is predicted, the accuracy of the final experimental result is directly influenced by the processing method of the time dependency relationship and the space dependency relationship. The spatial and temporal dependency relationship cannot be completely and dynamically processed only by simple self-learning of the neural network or a method of manually setting a fixed value. By means of the introduction of the attention mechanism, the dependency relationship among lanes under each intersection at each moment can be analyzed accurately in real time. After the target lane is selected, the target lane gives more attention to the lane with high relevance degree, the weight value of the irrelevant lane is reduced, and the weight is efficiently and dynamically and optimally distributed in a parallel mode.
Encoder neural network for processing flow sequence of each lane in road network
The input value of the ith coding unit is
The meaning of this is the traffic information in each lane at time i. According to formula (9):
wherein: the hidden layer of the i-1 subunit in the neural network of the decoder is
The cell state layer is
For this unit of lane flow data,
for the weight term of the unit of interest,
for the offset term, the intermediate term of the weight can be obtained after the function of the activation function tanh
Weighting and averaging the middle items in the m time periods to obtain a weighting result
S5: fusing the tensors obtained in S3 and S4, and carrying out self-learning of a time decoder network to obtain a final prediction result:
and according to the tensors obtained by calculation in S3.2 and S4.1, fusing the tensors to obtain input data in the time decoder network:
wherein Hυ，eAnd Hq，eThe matrix is formed by splicing hidden layers of each unit in the speed processing neural network and the encoder network. Wq，vAs weight terms, bq，vFor the offset term, the theta is obtained after linear transformation
The input data in the ith' unit of the decoder is thetai'according to the formula (9), let the hidden layer of the i' -1 th subunit in the decoder neural network be
Layer of cellular state
The formula for temporal attention is as follows:
wherein
Is formed by connecting a hidden layer in a previous decoder unit, a cell state layer and mixed characteristic data in a current unit,
are self-learned weight terms and offset terms within the cell.
Weighting the weighted intermediate items to obtain the in-cellTime of day attention weight
Further weighted to obtain
Historical true flow value and completed timing sequence weight distribution of target lane
Splicing and linear transformation are carried out to obtain
Wherein
Is the historical true flow value for the target lane,
and
is a self-learning weight term and an offset term
Each decoding unit performs timing weight assignment in the above manner. And continuously iterating until the Nth unit. And N is the size of the other dimension of the deep speed and flow characteristic tensor theta except the time dimension. Will be provided with
And
splicing and inputting the data into a full connection layer to obtain the traffic flow predicted by the network
Wherein
And
for the weight term and the offset term in the fully-connected layer,
is the final prediction result.
Claims (6)
1. The ground road lane-level traffic flow prediction method based on the space-time attention mechanism is characterized by comprising the following steps of:
s1: preprocessing original traffic flow data;
s2: making a lane selection strategy according to the data obtained in the step S1;
s3: carrying out speed processing neural network self-learning on the data obtained in the step S1;
s4: carrying out self-learning of the spatial encoder network on the data obtained in the step S1;
s5: and fusing the tensors obtained in the step S3 and the step S4, and performing self-learning of the time decoder network to obtain a final prediction result.
2. The ground road lane-level traffic flow prediction method based on the spatio-temporal attention mechanism as claimed in claim 1, wherein S1 specifically comprises the steps of:
s1.1: cleaning lane-level traffic flow parameters, such as missing value removal, outlier data and the like, and carrying out standardized processing on basic data;
s1.2: and drawing a curve according to the cleaned characteristic parameters, and repairing the obvious burr noise points by using a sliding window method with a set length to smooth the burr noise points.
3. The ground road lane-level traffic flow prediction method based on the spatio-temporal attention mechanism as claimed in claim 1, wherein S2 specifically comprises the steps of:
s2.1: in the method, each urban ground road intersection which is tightly connected is used as a lane dividing and selecting basis, and is verified by using an association degree analysis method, and a corresponding grey association degree calculation formula is as follows:
wherein:
characteristic parameters of the target lane at the time t-k,
the characteristic parameters of the participating lanes at the time t-k, u is the value of each item in the sequence, Δ max is the maximum value of the global traffic characteristic parameter, Δ min is the minimum value of the global traffic characteristic parameter, ρ is an empirical coefficient, and the obtained association value is obtained
And (5) carrying out normalization processing to obtain a final result pt-k, u.
4. The ground road lane-level traffic flow prediction method based on the spatio-temporal attention mechanism as claimed in claim 1, wherein S3 specifically comprises the steps of:
s3.1: the lane speed data is arranged into a matrix format required by a speed processing neural network:
wherein:
is the average speed of the i number lane in the j time period, n is the total number of lanes, m is the total number of time periods
S3.2: the speed processing neural network subunit is Bi-LSTM, in the scheme, in order to better capture the deep level characteristics of the vehicle speed sequence, a deep Bi-LSTM network is constructed in a head-to-tail series connection mode:
wherein:
is the average speed value of the lanes in the unit of number l,
and
for the hidden layer and the cell state layer after the unit is propagated in the forward direction, the e parameter represents an encoder used for distinguishing from a decoder; in the same way
And
the hidden layer and the cell state layer after back propagation are obtained by using the positive and negative results of the previous unit, namely the unit l-1 after propagation,
is the hidden layer obtained after the unit is fused.
5. The ground road lane-level traffic flow prediction method based on the spatio-temporal attention mechanism as claimed in claim 1, wherein S4 specifically comprises the steps of:
s4.1: the coding neural network embedded with the spatial attention mechanism consists of another Bi-LSTM unit; the spatial attention in the ith cell in a coded neural network can be summarized as: firstly, the initial Query value in the ith unit
Initial Key value
And Value
After linear transformation, respectively obtain
Then will be
And
making dot product, calculating every weight, finally making
Multiplied by the weights of the terms:
wherein: the hidden layer of the i-1 subunit in the neural network of the decoder is
The cell state layer is
For this unit of lane flow data,
for the weight term of the unit of interest,
the parameter is an offset item, q is an identifier of a flow processing parameter, and a is an identifier of an attention mechanism parameter; the intermediate term of the weight can be obtained after the action of the activation function tanh
Weighting and averaging the middle items in the m time periods to obtain a weighting result
6. The ground road lane-level traffic flow prediction method based on the spatio-temporal attention mechanism as claimed in claim 1, wherein S5 specifically comprises the steps of:
s5.1: fusing the tensors obtained by calculation in S3.2 and S4.1 to obtain input data theta in the time decoder network;
s5.2: and (3) taking the deep space-time feature tensor obtained by fusion in the S5.1 as an input value in a time decoder network, and obtaining a prediction result through a final full-connection layer:
wherein:
for the weight terms in the i' unit of the decoder,
for the offset term within a cell,
and
the hidden layer and the cell state layer of the previous unit are respectively, and the parameter d represents an encoder and is used for distinguishing the encoder;
for encoder networks and speedsThe tensor after neural network fusion is processed in degrees,
for each timing weight, assign it to
Can obtain the result after distribution
For the historical true value at time t, the value is compared with
Splicing and linear transformation are carried out to obtain
Obtaining the final prediction result after the processing of the full connection layer
Tau is the prediction duration, and the short-time prediction is generally set within one hour;
and
for the weight terms in the two transformations,
and
for the offset term in the transform, f is used to identify various types of parameters in this fully-connected layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110114317.8A CN112926768A (en) | 2021-01-27 | 2021-01-27 | Ground road lane-level traffic flow prediction method based on space-time attention mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110114317.8A CN112926768A (en) | 2021-01-27 | 2021-01-27 | Ground road lane-level traffic flow prediction method based on space-time attention mechanism |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112926768A true CN112926768A (en) | 2021-06-08 |
Family
ID=76167421
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110114317.8A Pending CN112926768A (en) | 2021-01-27 | 2021-01-27 | Ground road lane-level traffic flow prediction method based on space-time attention mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112926768A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114399901A (en) * | 2022-01-04 | 2022-04-26 | 青岛海信网络科技股份有限公司 | Method and equipment for controlling traffic system |
| CN114692941A (en) * | 2021-12-30 | 2022-07-01 | 江南大学 | Multi-attention-based company financial prediction method |
| CN114822019A (en) * | 2022-03-31 | 2022-07-29 | 阿里巴巴（中国）有限公司 | Traffic information processing method and device |
| CN117854279A (en) * | 2024-01-09 | 2024-04-09 | 南京清正源信息技术有限公司 | Road condition prediction method and system based on edge calculation |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9760806B1 (en) * | 2016-05-11 | 2017-09-12 | TCL Research America Inc. | Method and system for vision-centric deep-learning-based road situation analysis |
| CN109242140A (en) * | 2018-07-24 | 2019-01-18 | 浙江工业大学 | A kind of traffic flow forecasting method based on LSTM_Attention network |
| CN110570651A (en) * | 2019-07-15 | 2019-12-13 | 浙江工业大学 | Road network traffic situation prediction method and system based on deep learning |
| CN110619430A (en) * | 2019-09-03 | 2019-12-27 | 大连理工大学 | Space-time attention mechanism method for traffic prediction |
| CN110827546A (en) * | 2019-11-21 | 2020-02-21 | 银江股份有限公司 | Signalized intersection road section short-term speed prediction method |
| US20200082248A1 (en) * | 2018-09-11 | 2020-03-12 | Nvidia Corporation | Future object trajectory predictions for autonomous machine applications |
| US20200139973A1 (en) * | 2018-11-01 | 2020-05-07 | GM Global Technology Operations LLC | Spatial and temporal attention-based deep reinforcement learning of hierarchical lane-change policies for controlling an autonomous vehicle |
| CN111639791A (en) * | 2020-05-11 | 2020-09-08 | 同济大学 | Traffic flow prediction method, system, storage medium and terminal |
| CN112215337A (en) * | 2020-09-30 | 2021-01-12 | 江苏大学 | Vehicle trajectory prediction method based on environment attention neural network model |
-
2021
- 2021-01-27 CN CN202110114317.8A patent/CN112926768A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9760806B1 (en) * | 2016-05-11 | 2017-09-12 | TCL Research America Inc. | Method and system for vision-centric deep-learning-based road situation analysis |
| CN109242140A (en) * | 2018-07-24 | 2019-01-18 | 浙江工业大学 | A kind of traffic flow forecasting method based on LSTM_Attention network |
| US20200082248A1 (en) * | 2018-09-11 | 2020-03-12 | Nvidia Corporation | Future object trajectory predictions for autonomous machine applications |
| US20200139973A1 (en) * | 2018-11-01 | 2020-05-07 | GM Global Technology Operations LLC | Spatial and temporal attention-based deep reinforcement learning of hierarchical lane-change policies for controlling an autonomous vehicle |
| CN110570651A (en) * | 2019-07-15 | 2019-12-13 | 浙江工业大学 | Road network traffic situation prediction method and system based on deep learning |
| CN110619430A (en) * | 2019-09-03 | 2019-12-27 | 大连理工大学 | Space-time attention mechanism method for traffic prediction |
| CN110827546A (en) * | 2019-11-21 | 2020-02-21 | 银江股份有限公司 | Signalized intersection road section short-term speed prediction method |
| CN111639791A (en) * | 2020-05-11 | 2020-09-08 | 同济大学 | Traffic flow prediction method, system, storage medium and terminal |
| CN112215337A (en) * | 2020-09-30 | 2021-01-12 | 江苏大学 | Vehicle trajectory prediction method based on environment attention neural network model |
Non-Patent Citations (3)
| Title |
|---|
| DO, LNN ETC.: "An effective spatial-temporal attention based neural network for traffic flow prediction", TRANSPORTATION RESEARCH PART C-EMERGING TECHNOLOGIES, vol. 108, 17 September 2019 (2019-09-17), pages 12 - 28, XP085888874, DOI: 10.1016/j.trc.2019.09.008 * |
| 杜圣东;李天瑞;杨燕;王浩;谢鹏;洪西进;: "一种基于序列到序列时空注意力学习的交通流预测模型", 计算机研究与发展, no. 08, 6 August 2020 (2020-08-06) * |
| 陆文琦: "基于微波数据的快速路交通流数据修复及预测方法研究", 中国硕士学位论文全文数据库工程科技Ⅱ辑, 15 January 2020 (2020-01-15) * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114692941A (en) * | 2021-12-30 | 2022-07-01 | 江南大学 | Multi-attention-based company financial prediction method |
| CN114399901A (en) * | 2022-01-04 | 2022-04-26 | 青岛海信网络科技股份有限公司 | Method and equipment for controlling traffic system |
| CN114399901B (en) * | 2022-01-04 | 2022-12-13 | 青岛海信网络科技股份有限公司 | Method and equipment for controlling traffic system |
| CN114822019A (en) * | 2022-03-31 | 2022-07-29 | 阿里巴巴（中国）有限公司 | Traffic information processing method and device |
| CN117854279A (en) * | 2024-01-09 | 2024-04-09 | 南京清正源信息技术有限公司 | Road condition prediction method and system based on edge calculation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112926768A (en) | Ground road lane-level traffic flow prediction method based on space-time attention mechanism | |
| CN104778834B (en) | Urban road traffic jam judging method based on vehicle GPS data | |
| CN100444210C (en) | Mixed controlling method of single dot signal controlling crossing | |
| CN109191849B (en) | Traffic jam duration prediction method based on multi-source data feature extraction | |
| CN107919014B (en) | Taxi running route optimization method for multiple passenger mileage | |
| CN110390349A (en) | Bus passenger flow volume based on XGBoost model predicts modeling method | |
| CN107025788B (en) | Travel time prediction method and device | |
| Monteiro et al. | On-street parking prediction using real-time data | |
| CN109191852B (en) | Vehicle-road-cloud cooperative traffic flow situation prediction method | |
| CN112150207A (en) | Online taxi appointment order demand prediction method based on space-time context attention network | |
| CN110288198B (en) | Method for measuring bearing capacity of rental bicycle transportation facility based on traffic subareas | |
| CN113837446B (en) | Airport land side area traffic situation prediction method based on multi-source heterogeneous data | |
| CN109489679B (en) | Arrival time calculation method in navigation path | |
| CN110836675A (en) | Decision tree-based automatic driving search decision method | |
| CN115687709A (en) | Traffic dynamic control method based on traffic data dimension reduction reconstruction and multidimensional analysis | |
| Glushkov et al. | Analysis of the intersection throughput at changes in the traffic flow structure | |
| CN113140108A (en) | Cloud traffic situation prediction method in internet-connected intelligent traffic system | |
| Izadi et al. | Accident analysis of bus rapid transit system: Before and after construction | |
| Chandra et al. | Headway modelling under mixed traffic on urban roads | |
| CN105608315A (en) | Quantitative evaluation method and system of safety service level of bus stop | |
| CN112509357B (en) | Bus departure time interval optimization method based on genetic algorithm | |
| Mussone et al. | A neural network approach to motorway OD matrix estimation from loop counts | |
| Ravish et al. | Software Implementation of an Automatic Movable Road Barrier | |
| Wang et al. | Intelligent Driving Strategy of Expressway Based on Big Data of Road Network and Driving Time | |
| CN106530689B (en) | A kind of real-time predictor method of bus arrival time based on genetic algorithm and running data |
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 |