CN114372627A - Urban vehicle travel time estimation method based on hybrid deep learning framework - Google Patents
Urban vehicle travel time estimation method based on hybrid deep learning framework Download PDFInfo
- Publication number
- CN114372627A CN114372627A CN202210017292.4A CN202210017292A CN114372627A CN 114372627 A CN114372627 A CN 114372627A CN 202210017292 A CN202210017292 A CN 202210017292A CN 114372627 A CN114372627 A CN 114372627A
- Authority
- CN
- China
- Prior art keywords
- data set
- vector
- path
- sparse
- feature vector
- 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
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
- G06N3/045—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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Traffic Control Systems (AREA)
Abstract
The embodiment of the disclosure provides an urban vehicle travel time estimation method based on a hybrid deep learning framework, which belongs to the technical field of data processing and specifically comprises the following steps: constructing an undirected graph; forming a set of embedded vectors; collecting an original data set, and dividing the original data set into a path data set, a time data set, a traffic data set, a weather data set and a personal data set according to a data structure of the original data set; obtaining a vector sequence corresponding to each data set; obtaining multi-modal features; obtaining a comprehensive characteristic vector; calculating an estimated value of travel time; iteratively updating the learnable parameters according to the error values of the estimated values and the real values until the error values are smaller than a threshold value to obtain a target model; and collecting a path data set, a time data set, a traffic data set, a weather data set and a personal data set of the target vehicle, and inputting the data sets into a target model to obtain a travel time predicted value. By the scheme, the interactive relation among the features is fitted, the multi-mode features are effectively fused, and the vehicle travel time is comprehensively and accurately estimated.
Description
Technical Field
The embodiment of the disclosure relates to the technical field of data processing, in particular to a method for estimating the travel time of an urban vehicle based on a hybrid deep learning framework.
Background
At present, with the acceleration of urbanization and the improvement of living standard of people, the travel demand of urban residents is increasing day by day. In order to improve the travel experience of residents and realize effective regulation and intelligent management of an urban traffic system in the travel process of the residents, an accurate Time of Arrival (ETA) technology becomes the key point of academic research, and large-scale wide deployment is successfully carried out in the industry, and the ETA technology is a core technology vane for landing operation and market competition of a network booking platform including dripping, T3, Cao and the like.
The existing vehicle travel time estimation method has certain problems in the model design level facing to the actual urban traffic scene and application requirements, and is mainly embodied in the following points:
data reflecting the travel law of vehicles in the big data era should be heterogeneous from multiple sources, such as characters (weather, drivers), numbers (time), topological graphs (road network), sequences (vehicle path), and the like. Most of the existing methods can only estimate the travel time based on single type data or a small amount of type data, and the problems of joint learning and multi-mode feature extraction of multi-source heterogeneous data in a real traffic scene cannot be solved, so that effective information is lost.
Most of the existing methods are based on the assumption of complete data, and the problem of data sparsity in real traffic scenes cannot be considered, for example, in the data, only a few vehicles pass through part of roads, or only a few strokes are generated by part of drivers. Such as sparseness issues that do not account for data in model design, can lead to the introduction of null noise.
In a complex traffic scene, the travel time of a vehicle should be determined by environmental factors and path information, and a complex interaction mechanism exists between the environmental factors and the path information to influence the final travel time result. Most of the existing methods only simply map regression relations between the features and the travel time, and do not consider complex action mechanisms among the features, so that the travel time estimation result is influenced.
Therefore, an urban vehicle travel time estimation method based on a hybrid deep learning framework with high prediction accuracy and high adaptability is needed.
Disclosure of Invention
In view of the above, the disclosed embodiments provide an urban vehicle travel time estimation method based on a hybrid deep learning framework, which at least partially solves the problems of poor prediction efficiency and environmental adaptability in the prior art.
The embodiment of the disclosure provides an urban vehicle travel time estimation method based on a hybrid deep learning framework, which comprises the following steps:
dividing urban road networks based on road intersections to construct an undirected graph;
learning the vector expression of each road section in the undirected graph by adopting a preset algorithm to form an embedded vector set of the undirected graph;
collecting an original data set related to vehicle travel time in a complex traffic scene, and dividing the original data set into a path data set, a time data set, a traffic data set, a weather data set and a personal data set according to a data structure of the original data set;
respectively inputting different data sets into a multi-view and multi-scale coupled digital representation model for coding to obtain a vector sequence corresponding to each data set;
inputting all the vector sequences into a mixed deep neural network learning framework to obtain multi-modal features, wherein the multi-modal features comprise dense feature vectors, sparse feature vectors and path feature vectors;
fusing the dense feature vector, the sparse feature vector and the path feature vector to obtain a comprehensive feature vector;
calculating an estimated value of travel time according to the comprehensive feature vector and learnable parameters of the hybrid deep neural network learning framework;
iteratively updating the learnable parameters according to the error values of the estimated values and the real values until the error values are smaller than a threshold value to obtain a target model;
and collecting a path data set, a time data set, a traffic data set, a weather data set and a personal data set of the target vehicle, and inputting the path data set, the time data set, the traffic data set, the weather data set and the personal data set into the target model to obtain a predicted travel time value.
According to a specific implementation manner of the embodiment of the present disclosure, the step of learning the vector expression of each road segment in the undirected graph by using a preset algorithm to form an embedded vector set of the undirected graph includes:
calculating a transition matrix based on connectivity of the road segment;
calculating a random walk sequence of each road section according to the transfer matrix to form a sequence set;
and solving a mapping function according to the sequence set and the preset algorithm to form the embedded vector set.
According to a specific implementation manner of the embodiment of the present disclosure, the step of inputting different data sets into the multi-view and multi-scale coupled digital representation model respectively for encoding to obtain a vector sequence corresponding to each data set includes:
mapping the path data set to a vector space according to the embedded vector set to obtain a path vector sequence;
characterizing the time data set by utilizing multi-scale periodic position coding to obtain a real-value vector sequence;
and performing multi-mode parallel representation on the traffic data set and the weather data set and the personal data set to obtain a sparse vector sequence.
According to a specific implementation manner of the embodiment of the present disclosure, the step of performing multi-modal parallel representation on the traffic data set and the personal data set of the weather data set to obtain a sparse vector sequence includes:
encoding each sparse data in the traffic data set and the personal data set of the weather data set by using a real-value vector generated randomly;
and coding each sparse data by adopting a sparse vector to obtain a sparse vector sequence, setting the corresponding element of the type dimension represented by the sparse data in the sparse vector sequence as 1, and setting the elements of other dimensions as zero.
According to a specific implementation manner of the embodiment of the disclosure, the hybrid deep neural network learning framework comprises a gating cycle unit, a multilayer perceptron, an integrated factorization module and a feature fusion module.
According to a specific implementation manner of the embodiment of the present disclosure, the step of inputting all the vector sequences into a hybrid deep neural network learning framework to obtain multi-modal features includes:
learning the dependency characteristics of the path vector sequence by using the gating cycle unit to obtain a path characteristic vector;
utilizing the multilayer perceptron to learn the dense features of the real-valued vector sequence to obtain dense feature vectors;
extracting the features of the sparse vector sequence by using the integration factor module to obtain a sparse feature vector;
and forming the multi-modal features according to the path feature vector, the dense feature vector and the sparse feature vector.
According to a specific implementation manner of the embodiment of the present disclosure, the step of learning the dependency feature of the path vector sequence by using the gated loop unit to obtain the path feature vector includes:
and iteratively calculating the characteristic quantity corresponding to each position vector in the path vector sequence according to the learnable parameters of the gating cycle unit to form the path characteristic vector.
According to a specific implementation manner of the embodiment of the present disclosure, the step of fusing the dense feature vector, the sparse feature vector, and the path feature vector to obtain a comprehensive feature vector includes:
the feature fusion module splices the dense feature vector and the sparse feature vector to serve as an environment feature vector;
performing linear mapping on the path characteristic vector and the environment characteristic vector through the learning parameters of the characteristic fusion module to obtain interaction condition probability;
and calculating an expected value of the path vector sequence according to the interaction conditional probability to serve as the comprehensive characteristic vector.
According to a specific implementation manner of the embodiment of the present disclosure, the step of iteratively updating the learnable parameter according to the error value between the estimated value and the true value until the error value is smaller than the threshold value to obtain the target model includes:
updating the learnable parameters by using an Adam optimizer according to the error values and calculating the error values corresponding to the learnable parameters updated this time;
transmitting the optimized error value to the learnable parameter updated last time by using a back propagation algorithm, and updating;
and when the error value is smaller than the threshold value, obtaining the target model according to the updated learnable parameters.
The urban vehicle travel time estimation scheme based on the hybrid deep learning framework in the embodiment of the disclosure comprises the following steps: dividing urban road networks based on road intersections to construct an undirected graph; learning the vector expression of each road section in the undirected graph by adopting a preset algorithm to form an embedded vector set of the undirected graph; collecting an original data set related to vehicle travel time in a complex traffic scene, and dividing the original data set into a path data set, a time data set, a traffic data set, a weather data set and a personal data set according to a data structure of the original data set; respectively inputting different data sets into a multi-view and multi-scale coupled digital representation model for coding to obtain a vector sequence corresponding to each data set; inputting all the vector sequences into a mixed deep neural network learning framework to obtain multi-modal features, wherein the multi-modal features comprise dense feature vectors, sparse feature vectors and path feature vectors; fusing the dense feature vector, the sparse feature vector and the path feature vector to obtain a comprehensive feature vector; calculating an estimated value of travel time according to the comprehensive feature vector and learnable parameters of the hybrid deep neural network learning framework; iteratively updating the learnable parameters according to the error values of the estimated values and the real values until the error values are smaller than a threshold value to obtain a target model; and collecting a path data set, a time data set, a traffic data set, a weather data set and a personal data set of the target vehicle, and inputting the path data set, the time data set, the traffic data set, the weather data set and the personal data set into the target model to obtain a predicted travel time value.
The beneficial effects of the embodiment of the disclosure are: by the scheme, a data vector characterization method under multi-view and multi-scale coupling is designed in a targeted manner, data type characteristics under a real traffic scene are fully considered, the problem of applicability characterization of sparse and multi-source heterogeneous data is solved, self-adaptive extraction of the multi-source heterogeneous data characteristics is realized while the problem of data sparsity is considered, interactive relations among the characteristics are fitted under a complex traffic scene, and effective fusion of multi-mode characteristics is realized, so that the travel time of a vehicle is estimated comprehensively and accurately.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed to be used in the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present disclosure, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a schematic flow chart diagram of a method for estimating the travel time of an urban vehicle based on a hybrid deep learning framework according to an embodiment of the present disclosure;
FIG. 2 is a schematic view of a vehicle travel time estimation data processing flow provided by an embodiment of the present disclosure;
fig. 3 is a schematic diagram of a result of comparing accuracy with that of a conventional method under different stroke length distributions according to an embodiment of the present disclosure;
fig. 4 is a diagram illustrating a result of comparing precision with that of a conventional method in different time periods according to an embodiment of the present disclosure.
Detailed Description
The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.
The embodiments of the present disclosure are described below with specific examples, and other advantages and effects of the present disclosure will be readily apparent to those skilled in the art from the disclosure in the specification. It is to be understood that the described embodiments are merely illustrative of some, and not restrictive, of the embodiments of the disclosure. The disclosure may be embodied or carried out in various other specific embodiments, and various modifications and changes may be made in the details within the description without departing from the spirit of the disclosure. It is to be noted that the features in the following embodiments and examples may be combined with each other without conflict. All other embodiments, which can be derived by a person skilled in the art from the embodiments disclosed herein without making any creative effort, shall fall within the protection scope of the present disclosure.
It is noted that various aspects of the embodiments are described below within the scope of the appended claims. It should be apparent that the aspects described herein may be embodied in a wide variety of forms and that any specific structure and/or function described herein is merely illustrative. Based on the disclosure, one skilled in the art should appreciate that one aspect described herein may be implemented independently of any other aspects and that two or more of these aspects may be combined in various ways. For example, an apparatus may be implemented and/or a method practiced using any number of the aspects set forth herein. Additionally, such an apparatus may be implemented and/or such a method may be practiced using other structure and/or functionality in addition to one or more of the aspects set forth herein.
It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present disclosure, and the drawings only show the components related to the present disclosure rather than the number, shape and size of the components in actual implementation, and the type, amount and ratio of the components in actual implementation may be changed arbitrarily, and the layout of the components may be more complicated.
In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects may be practiced without these specific details.
The embodiment of the disclosure provides an urban vehicle travel time estimation method based on a hybrid deep learning framework, and the method can be applied to a vehicle travel time prediction process in an urban traffic management or personal occurrence planning scene.
Referring to fig. 1, a schematic flow chart of an urban vehicle travel time estimation method based on a hybrid deep learning framework provided by an embodiment of the present disclosure is shown. As shown in fig. 1, the method mainly comprises the following steps:
s101, dividing an urban road network based on road intersections to construct an undirected graph;
for example, road network data composed of 8651005 road segments in Shenzhen city whole city and full-volume network joke journey data from 8/1/2020 to 8/31/2020 can be combined as an actual data set implemented by the specific invention, the implementation process of the disclosure is explained, and an undirected graph structure G ═ V, E is constructed for the topological relation of 8651005 road segments in Shenzhen city whole city, where V ═ Ei,1≤i≤NVDenotes all N in a cityVSet of links, E ═ Ei,1≤i≤NEDenotes connectivity of a road segment, if a road segment viAnd vjConnected through an intersection, then eij＝1。
S102, learning the vector expression of each road section in the undirected graph by adopting a preset algorithm to form an embedded vector set of the undirected graph;
optionally, in step S102, learning a vector expression of each road segment in the undirected graph by using a preset algorithm to form an embedded vector set of the undirected graph, where the method includes:
calculating a transition matrix based on connectivity of the road segment;
calculating a random walk sequence of each road section according to the transfer matrix to form a sequence set;
and solving a mapping function according to the sequence set and the preset algorithm to form the embedded vector set.
In specific implementation, based on the undirected graph G ═ V, E, each road section V is learned by using a node2vec methodiVector expression of (v)i) And form an embedded vector set J ═ Φ (v) of Gi),i∈[1,NV]}. Where Φ (·) is the mapping function. The method mainly comprises the following steps:
calculating a transition matrix P based on connectivity of road segmentsM＝diag(E1)-1E, wherein E ═ Eij]For the matrix formed by E, 1 is a vector with all elements 1. Element p in the transfer matrixijRepresenting a slave section viTo road section vjThe probability of (c). Then for each section viE.g. V, are all based on the transfer matrix PMGenerating a length of NSRandom walk sequence RS ofiForming sequence set RSS ═ RSi,i∈[1,NV]}. Then pass through
The mapping function Φ is solved. Wherein, NNk(v) Representing k road segments in topological distance from the road segment v in the graph G, forming the set of embedded vectors.
S103, collecting an original data set related to vehicle travel time in a complex traffic scene, and dividing the original data set into a path data set, a time data set, a traffic data set, a weather data set and a personal data set according to a data structure of the original data set;
in specific implementation, the implementation data set is arranged as the original data set, and then divided into path data such as a road section through which a vehicle passes, time data such as departure time, traffic data such as congestion condition, weather data such as temperature, and personal data such as a driver number according to a data structure of the implementation data set.
For example, path data
Time data time(i)Traffic data state(i)Weather data weather(i)Temperature data temperature(i)And personal data driver ID(i)Wherein N isTRepresenting the total number of strokes.
S104, respectively inputting different data sets into a multi-view and multi-scale coupled digital representation model for coding to obtain a vector sequence corresponding to each data set;
further, in step S104, inputting different data sets into the multi-view and multi-scale coupled digital representation model respectively for encoding, so as to obtain a vector sequence corresponding to each data set, including:
mapping the path data set to a vector space according to the embedded vector set to obtain a path vector sequence;
characterizing the time data set by utilizing multi-scale periodic position coding to obtain a real-value vector sequence;
and performing multi-mode parallel representation on the traffic data set and the weather data set and the personal data set to obtain a sparse vector sequence.
On the basis of the above embodiment, the step of performing multi-modal parallel representation on the traffic data set and the weather data set and the personal data set to obtain a sparse vector sequence includes:
encoding each sparse data in the traffic data set and the personal data set of the weather data set by using a real-value vector generated randomly;
and coding each sparse data by adopting a sparse vector to obtain a sparse vector sequence, setting the corresponding element of the type dimension represented by the sparse data in the sparse vector sequence as 1, and setting the elements of other dimensions as zero.
In specific implementation, a multi-view and multi-scale coupled digital characterization model can be designed by considering that the data structures of different data sets are different
The method is used for adaptively and pertinently coding multi-source heterogeneous and sparse data. In particular, for path datasets
Wherein v isjJ-th route section l passing through the route of i-th traveliThe number of road sections included in the ith travel path is indicated. The set of embedded vectors J ═ Φ (v) constructed by step 3)j),j∈[1,li]}，Mapping path data r(i)To vector space, composing said sequence of path vectors
Time for the time data(i)Can be characterized by a multi-scale periodic position code
Obtaining the real value vector sequence, wherein cat (·,) represents the splicing operation of the vector. E.g. computing
And
and the results are respectively used as vectors
The values of odd and even elements, where h represents the vector
When scale is d, posdRepresenting time(i)At the hour of the day, and when scale ═ w, poswRepresenting time(i)On the week of the week.
The method adopts a multi-mode parallel characterization method for sparse data including traffic data, weather data and personal data. The method mainly comprises the following steps:
and coding each sparse data in the traffic data set and the personal data set by adopting a real-value vector generated randomly, then coding each sparse data by adopting a sparse vector to obtain a sparse vector sequence, setting the corresponding element of the type dimension represented by the sparse data in the sparse vector sequence as 1, and setting the elements of other dimensions as zero.
S105, inputting all the vector sequences into a hybrid deep neural network learning framework to obtain multi-modal features, wherein the multi-modal features comprise dense feature vectors, sparse feature vectors and path feature vectors;
optionally, the hybrid deep neural network learning framework includes a gate control cycle unit, a multilayer perceptron, an integrated factorization module, and a feature fusion module.
Further, in step S105, inputting all the vector sequences into a hybrid deep neural network learning framework to obtain multi-modal features, including:
learning the dependency characteristics of the path vector sequence by using the gating cycle unit to obtain a path characteristic vector;
utilizing the multilayer perceptron to learn the dense features of the real-valued vector sequence to obtain dense feature vectors;
extracting the features of the sparse vector sequence by using the integration factor module to obtain a sparse feature vector;
and forming the multi-modal features according to the path feature vector, the dense feature vector and the sparse feature vector.
Optionally, the step of learning the dependency feature of the path vector sequence by using the gated loop unit to obtain a path feature vector includes:
and iteratively calculating the characteristic quantity corresponding to each position vector in the path vector sequence according to the learnable parameters of the gating cycle unit to form the path characteristic vector.
In specific implementation, considering that multi-modal features H (-) in data need to be extracted in a targeted and adaptive manner to comprehensively characterize complex urban traffic scenes, a hybrid deep neural network learning framework can be designed, specifically, a path vector sequence is targeted
Learning by using gated cyclic unit GRUNon-linear time-dependent characteristics
The method mainly comprises the following steps:
characterised by the path sequence
Of a certain position t, and a vector phi (v)t) By way of example, via learnable network parameters
Calculating the feature vector corresponding to the position vector
The formula is as follows:
then from 1 to l according to tiIteratively calculating in a loop until the path feature vector is generated
Aiming at the real-valued vector sequence, a multilayer perceptron is adopted to learn the dense feature h of the real-valued vector sequenced(i) The formula is as follows:
wherein l is the number of layers,
and
is a learnable parameter of the corresponding layer. Taking the output of the last layer of formula (4) as the dense feature vector hd(i)。
Aiming at the sparse vector sequence, designing an integrated factor decomposition module (EFMB) to carry out feature vector hs(i) Obtaining the sparse feature vector by the following formula:
where x denotes the input sparse vector, V is a learnable parameter tensor, as a Hadamard product, and Δ is the bulk matrix product.
And then forming the multi-modal features according to the path feature vector, the dense feature vector and the sparse feature vector.
S106, fusing the dense feature vector, the sparse feature vector and the path feature vector to obtain a comprehensive feature vector;
on the basis of the foregoing embodiment, the step S106 of fusing the dense feature vector, the sparse feature vector, and the path feature vector to obtain a comprehensive feature vector includes:
the feature fusion module splices the dense feature vector and the sparse feature vector to serve as an environment feature vector;
performing linear mapping on the path characteristic vector and the environment characteristic vector through the learning parameters of the characteristic fusion module to obtain interaction condition probability;
and calculating an expected value of the path vector sequence according to the interaction conditional probability to serve as the comprehensive characteristic vector.
In specific implementation, the feature fusion module can consider complex interaction relation and fit the environmental factor features (h)d(i),hs(i) And path characteristics H (r)(i)) The fused comprehensive characteristic vector h (i) ═ MCLB (h) is obtainedd(i),hs(i),H(r(i))). The method mainly comprises the following steps:
splicing dense feature vector hd(i) And sparse feature vector hs(i) As an environmental feature vector gf(i)＝cat(hd(i),hs(i) ). And then passes the learnable parameters
And
respectively for path characteristics H (r)(i)) Each feature vector Φ (v) inj) And an environmental feature vector gf(i)Linear mapping is carried out to obtain phi (v)j)cAnd gf(i) qThe mapping formula is:
based on the above, the interaction mechanism and interaction relation between the modeling environmental factor characteristic and the path characteristic are the interaction conditional probability p (phi (v)j)c|gf(i) q)。
Then calculating the expected value of the path vector sequence according to the interaction conditional probability
As a combination of features(Vector)
S107, calculating an estimated value of travel time according to the comprehensive feature vector and learnable parameters of the hybrid deep neural network learning framework;
in specific implementation, the estimated value of the travel time can be calculated according to the comprehensive characteristic vector and the learnable parameters of the hybrid deep neural network learning framework, and a specific calculation formula can be
S108, iteratively updating the learnable parameters according to the error values of the estimated values and the true values until the error values are smaller than a threshold value to obtain a target model;
further, in step S109, iteratively updating the learnable parameter according to an error value between the estimated value and the true value until the error value is smaller than a threshold value, to obtain a target model, including:
updating the learnable parameters by using an Adam optimizer according to the error values and calculating the error values corresponding to the learnable parameters updated this time;
transmitting the optimized error value to the learnable parameter updated last time by using a back propagation algorithm, and updating;
and when the error value is smaller than the threshold value, obtaining the target model according to the updated learnable parameters.
In specific implementation, considering that the learnable parameters are initialized randomly at first, if a prediction model is obtained by directly embedding the learnable parameters into a neural network and prediction accuracy has a large error, the prediction model can be obtained according to the estimated value
With true value τ(i)Iteratively updating the learnable parameter
V，
WoAnd obtaining the target model until the error value is smaller than a threshold value.
For example, the mean absolute percentage error may be used
Calculating an estimate
With true value τ(i)Then the error of each operation is transferred to the learnable parameters of the model by using a back propagation algorithm
V，
WoThe value of each learnable parameter is updated based on the mean absolute percentage error loss using an Adam optimizer.
And then, repeating the steps until the error value is smaller than the threshold value, and selecting the learnable parameter at the moment
V，
WoAnd embedding the parameters into a model as final parameters to obtain the target model.
And S109, collecting a path data set, a time data set, a traffic data set, a weather data set and a personal data set of the target vehicle, inputting the data sets into the target model, and obtaining a travel time predicted value.
After the target model is obtained, when the traffic condition needs to be managed, or when an individual needs to go out, a path data set, a time data set, a traffic data set, a weather data set and an individual data set of a target vehicle can be collected and input into the target model, and a travel time prediction value is obtained.
According to the urban vehicle travel time estimation method based on the hybrid deep learning framework, the data type characteristics under the real traffic scene are fully considered by designing a data vector characterization method under the coupling of multiple views and multiple scales in a targeted manner, the applicability characterization problem of sparse and multi-source heterogeneous data is solved, the self-adaptive extraction of the multi-source heterogeneous data characteristics is realized while the data sparsity problem is considered based on the vector characterization of data, the interactive relation among the characteristics is fitted under the complex traffic scene, the effective fusion of the multi-mode characteristics is realized, and therefore the travel time of the vehicle is comprehensively and accurately estimated.
The present solution will be described with reference to a specific embodiment, and when prediction is performed according to corresponding data of a certain city vehicle, the overall flow is shown in fig. 2. Then, the accuracy comparison and verification are carried out by using various existing travel time estimation methods and the invention, and the effectiveness of the invention is checked. The comparison method comprises the following steps of: AVG and RealTimeAVG; the widely applied travel time estimation method comprises the following steps: GBDT, MlpETA and DeeptTE; the existing advanced travel time estimation method comprises the following steps: WDR and FMA-ETA. The method mainly comprises the following steps:
selecting a plurality of evaluation indexes of the estimation precision of the travel time, including average absolute percentage error (MAPE); root Mean Square Error (RMSE); ③ Mean Absolute Error (MAE), the calculation formula is as follows:
wherein, tau(i)And
respectively representing the real travel time of the ith journey and the estimated travel time of the method model. The smaller the values of the above three evaluation indexes are, the higher the accuracy of the method model is.
Table 1 shows the precision comparison results of the urban vehicle travel time estimation method and the comparison method based on the hybrid deep learning framework in the embodiment of the present disclosure, and it can be seen from analyzing the estimation precision comparison tables of different methods that the method of the present disclosure obtains the best precision on three evaluation indexes.
TABLE 1
The method of the present invention and the international advanced method are analyzed for the accuracy comparison results under different stroke length distributions, as shown in fig. 3, wherein (a), (b), and (c) are the accuracy comparison results under different stroke length distributions, respectively. The method of the invention can realize the lowest estimation error and keep the highest precision under any path length.
The method of the present invention and the international advanced method are analyzed for the accuracy comparison result when dealing with the travel data in different time periods, as shown in fig. 4, wherein (a), (b), and (c) are the accuracy comparison results when dealing with the travel data in different time periods, respectively. Also, the method of the present invention can achieve the lowest estimation error and maintain the highest accuracy in any time period.
The contents of table 1, fig. 3 and fig. 4 can explain that: the method provided by the invention is applied to an actual traffic scene, can obtain the best travel time estimation effect, and can meet the requirements of more refined traffic management and decision.
The units described in the embodiments of the present disclosure may be implemented by software or hardware.
It should be understood that portions of the present disclosure may be implemented in hardware, software, firmware, or a combination thereof.
The above description is only for the specific embodiments of the present disclosure, but the scope of the present disclosure is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present disclosure should be covered within the scope of the present disclosure. Therefore, the protection scope of the present disclosure shall be subject to the protection scope of the claims.
Claims (9)
1. A method for estimating the travel time of an urban vehicle based on a hybrid deep learning framework is characterized by comprising the following steps:
dividing urban road networks based on road intersections to construct an undirected graph;
learning the vector expression of each road section in the undirected graph by adopting a preset algorithm to form an embedded vector set of the undirected graph;
collecting an original data set related to vehicle travel time in a complex traffic scene, and dividing the original data set into a path data set, a time data set, a traffic data set, a weather data set and a personal data set according to a data structure of the original data set;
respectively inputting different data sets into a multi-view and multi-scale coupled digital representation model for coding to obtain a vector sequence corresponding to each data set;
inputting all the vector sequences into a mixed deep neural network learning framework to obtain multi-modal features, wherein the multi-modal features comprise dense feature vectors, sparse feature vectors and path feature vectors;
fusing the dense feature vector, the sparse feature vector and the path feature vector to obtain a comprehensive feature vector;
calculating an estimated value of travel time according to the comprehensive feature vector and learnable parameters of the hybrid deep neural network learning framework;
iteratively updating the learnable parameters according to the error values of the estimated values and the real values until the error values are smaller than a threshold value to obtain a target model;
and collecting a path data set, a time data set, a traffic data set, a weather data set and a personal data set of the target vehicle, and inputting the path data set, the time data set, the traffic data set, the weather data set and the personal data set into the target model to obtain a predicted travel time value.
2. The method according to claim 1, wherein the step of learning the vector expression of each road segment in the undirected graph by using a preset algorithm to form the embedded vector set of the undirected graph comprises:
calculating a transition matrix based on connectivity of the road segment;
calculating a random walk sequence of each road section according to the transfer matrix to form a sequence set;
and solving a mapping function according to the sequence set and the preset algorithm to form the embedded vector set.
3. The method according to claim 1, wherein the step of inputting different data sets into the multi-view and multi-scale coupled digital representation model respectively for encoding to obtain the vector sequence corresponding to each data set comprises:
mapping the path data set to a vector space according to the embedded vector set to obtain a path vector sequence;
characterizing the time data set by utilizing multi-scale periodic position coding to obtain a real-value vector sequence;
and performing multi-mode parallel representation on the traffic data set and the weather data set and the personal data set to obtain a sparse vector sequence.
4. The method of claim 3, wherein the step of performing multi-modal parallel characterization on the traffic data set and the weather data set and the personal data set to obtain a sparse vector sequence comprises:
encoding each sparse data in the traffic data set and the personal data set of the weather data set by using a real-value vector generated randomly;
and coding each sparse data by adopting a sparse vector to obtain a sparse vector sequence, setting the corresponding element of the type dimension represented by the sparse data in the sparse vector sequence as 1, and setting the elements of other dimensions as zero.
5. The method of claim 1, wherein the hybrid deep neural network learning framework comprises a gated loop unit, a multi-layered perceptron, an integrated factorization module, and a feature fusion module.
6. The method of claim 5, wherein the step of inputting all the vector sequences into a hybrid deep neural network learning framework to obtain multi-modal features comprises:
learning the dependency characteristics of the path vector sequence by using the gating cycle unit to obtain a path characteristic vector;
utilizing the multilayer perceptron to learn the dense features of the real-valued vector sequence to obtain dense feature vectors;
extracting the features of the sparse vector sequence by using the integration factor module to obtain a sparse feature vector;
and forming the multi-modal features according to the path feature vector, the dense feature vector and the sparse feature vector.
7. The method according to claim 6, wherein the step of learning the dependency features of the path vector sequence by using the gated loop unit to obtain a path feature vector comprises:
and iteratively calculating the characteristic quantity corresponding to each position vector in the path vector sequence according to the learnable parameters of the gating cycle unit to form the path characteristic vector.
8. The method of claim 5, wherein the step of fusing the dense feature vector, the sparse feature vector, and the path feature vector to obtain a composite feature vector comprises:
the feature fusion module splices the dense feature vector and the sparse feature vector to serve as an environment feature vector;
performing linear mapping on the path characteristic vector and the environment characteristic vector through the learning parameters of the characteristic fusion module to obtain interaction condition probability;
and calculating an expected value of the path vector sequence according to the interaction conditional probability to serve as the comprehensive characteristic vector.
9. The method of claim 1, wherein iteratively updating the learnable parameter based on an error value between the estimated value and a true value until the error value is less than a threshold value to obtain a target model comprises:
updating the learnable parameters by using an Adam optimizer according to the error values and calculating the error values corresponding to the learnable parameters updated this time;
transmitting the optimized error value to the learnable parameter updated last time by using a back propagation algorithm, and updating;
and when the error value is smaller than the threshold value, obtaining the target model according to the updated learnable parameters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210017292.4A CN114372627A (en) | 2022-01-07 | 2022-01-07 | Urban vehicle travel time estimation method based on hybrid deep learning framework |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210017292.4A CN114372627A (en) | 2022-01-07 | 2022-01-07 | Urban vehicle travel time estimation method based on hybrid deep learning framework |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114372627A true CN114372627A (en) | 2022-04-19 |
Family
ID=81143282
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210017292.4A Pending CN114372627A (en) | 2022-01-07 | 2022-01-07 | Urban vehicle travel time estimation method based on hybrid deep learning framework |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114372627A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114973668A (en) * | 2022-05-18 | 2022-08-30 | 广州市交通规划研究院有限公司 | Urban road traffic weak link identification method based on topological step number analysis |
-
2022
- 2022-01-07 CN CN202210017292.4A patent/CN114372627A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114973668A (en) * | 2022-05-18 | 2022-08-30 | 广州市交通规划研究院有限公司 | Urban road traffic weak link identification method based on topological step number analysis |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bi et al. | Daily tourism volume forecasting for tourist attractions | |
| CN112216108B (en) | Traffic prediction method based on attribute-enhanced space-time graph convolution model | |
| CN109887282B (en) | Road network traffic flow prediction method based on hierarchical timing diagram convolutional network | |
| CN112382082B (en) | Method and system for predicting traffic running state in congested area | |
| CN111612243B (en) | Traffic speed prediction method, system and storage medium | |
| CN114299723B (en) | Traffic flow prediction method | |
| Li et al. | Graph CNNs for urban traffic passenger flows prediction | |
| CN113313947A (en) | Road condition evaluation method of short-term traffic prediction graph convolution network | |
| CN112766551B (en) | Traffic prediction method, intelligent terminal and computer readable storage medium | |
| CN113762595B (en) | Traffic time prediction model training method, traffic time prediction method and equipment | |
| CN111242395B (en) | Method and device for constructing prediction model for OD (origin-destination) data | |
| CN113762338B (en) | Traffic flow prediction method, equipment and medium based on multiple graph attention mechanism | |
| CN112071062A (en) | Driving time estimation method based on graph convolution network and graph attention network | |
| Jin et al. | HetGAT: a heterogeneous graph attention network for freeway traffic speed prediction | |
| Huo et al. | Cooperative control for multi-intersection traffic signal based on deep reinforcement learning and imitation learning | |
| Xing et al. | A data fusion powered bi-directional long short term memory model for predicting multi-lane short term traffic flow | |
| CN115206092A (en) | Traffic prediction method of BiLSTM and LightGBM model based on attention mechanism | |
| CN114780739A (en) | Time sequence knowledge graph completion method and system based on time graph convolution network | |
| Xu et al. | Short‐term traffic flow prediction based on whale optimization algorithm optimized BiLSTM_Attention | |
| Xiao et al. | Parking prediction in smart cities: A survey | |
| CN114372627A (en) | Urban vehicle travel time estimation method based on hybrid deep learning framework | |
| Chen et al. | Prediction of Public Bus Passenger Flow Using Spatial–Temporal Hybrid Model of Deep Learning | |
| CN113159371A (en) | Unknown target feature modeling and demand prediction method based on cross-modal data fusion | |
| Huang et al. | Multistep coupled graph convolution with temporal-attention for traffic flow prediction | |
| CN116612633A (en) | Self-adaptive dynamic path planning method based on vehicle-road cooperative sensing |
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 |