CN109190795B - Inter-area travel demand prediction method and device - Google Patents
Inter-area travel demand prediction method and device Download PDFInfo
- Publication number
- CN109190795B CN109190795B CN201810864027.3A CN201810864027A CN109190795B CN 109190795 B CN109190795 B CN 109190795B CN 201810864027 A CN201810864027 A CN 201810864027A CN 109190795 B CN109190795 B CN 109190795B
- Authority
- CN
- China
- Prior art keywords
- network
- travel demand
- spatial
- time
- matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 28
- 239000011159 matrix material Substances 0.000 claims abstract description 136
- 238000013528 artificial neural network Methods 0.000 claims abstract description 18
- 238000012549 training Methods 0.000 claims abstract description 15
- 238000000605 extraction Methods 0.000 claims abstract description 12
- 230000004927 fusion Effects 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 10
- 230000008859 change Effects 0.000 claims description 8
- 239000013598 vector Substances 0.000 claims description 8
- 230000007787 long-term memory Effects 0.000 claims description 7
- 238000013507 mapping Methods 0.000 claims description 7
- 230000006403 short-term memory Effects 0.000 claims description 7
- 230000006835 compression Effects 0.000 claims description 4
- 238000007906 compression Methods 0.000 claims description 4
- 238000002407 reforming Methods 0.000 claims description 4
- 241000764238 Isis Species 0.000 claims 1
- 230000000875 corresponding effect Effects 0.000 description 26
- 238000010586 diagram Methods 0.000 description 12
- 230000002123 temporal effect Effects 0.000 description 10
- 230000015654 memory Effects 0.000 description 6
- 238000010276 construction Methods 0.000 description 5
- 238000011160 research Methods 0.000 description 4
- 230000017105 transposition Effects 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000010076 replication Effects 0.000 description 2
- 230000003068 static effect Effects 0.000 description 2
- 230000002596 correlated effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000010030 laminating Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000007670 refining 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
- 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
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
- G06Q10/06315—Needs-based resource requirements planning or analysis
-
- G06Q50/40—
Abstract
The invention discloses a method and a device for predicting inter-area travel demands, wherein the method comprises the following steps: step S1, constructing a depth model with multiple context information extraction, using the traffic travel demand matrix sequence of multiple historical time periods as input and the actual traffic travel demand matrix of the next time period of the corresponding sequence as target output, and training the depth model by using a back propagation algorithm of a neural network; step S2, constructing a transportation demand matrix sequence with rich context information; and step S3, the depth model parameters obtained through training in the step S1 and the depth model are used as a final predictor, a continuous travel demand matrix sequence is input, and an unknown travel demand matrix of the next time period is predicted.
Description
Technical Field
The invention relates to the technical fields of safety monitoring, urban traffic management, deep learning and the like, in particular to a method and a device for predicting inter-area travel demands of a multi-space-time context information fusion mechanism based on deep learning.
Background
The inter-area travel demand prediction is an important task and has important application in the problems of urban traffic intelligent management, traffic resource scheduling in advance and the like. The regional travel demand analysis obtains the quantity of different regional travel demands by analyzing passenger carrying information and GPS position information of vehicles at historical moments, so that the travel demands at the next moments are predicted. And forecasting the inter-area traffic demand, further refining the travel demand of the areas, and forecasting the travel demand from one area to another area. The key to these travel demand forecasting problems is how to adequately capture spatial and temporal correlations, as well as specific global contextual information (e.g., relevance of work and residential areas in some contexts), to obtain accurate forecasts.
In academic research, a scholars puts forward a lot of work for space-time modeling of regional travel demands, but the work for the regional travel demands is relatively less, and the demands of different regional pairs not only have difference in quantity, but also have obvious difference in change rule. The existing research screens the demand of high-frequency region pairs, which is insufficient for the overall research of the trip demand among regions in complex change, ignores the association and global information among different region pairs and is difficult to capture the change trend of data; although the combination of the static global context information is applied to the prediction of the regional travel demand, the static pre-generated global context information is difficult to reflect the sudden change of the global context information, and a piece of dynamically generated global context information is required for prediction.
Disclosure of Invention
In order to overcome the defects in the prior art, the present invention provides a method and an apparatus for predicting inter-area travel demand, so as to extract multiple context information by using a Temporal-Spatial-Temporal Network (CSTN) to fully utilize historical data of inter-area travel demand and other external information, predict travel demand of all area pairs, and improve accuracy of inter-area travel demand prediction.
To achieve the above and other objects, the present invention provides a method for predicting inter-area travel demand, comprising the steps of:
step S1, constructing a depth model with multiple context information extraction, using the traffic travel demand matrix sequence of multiple historical time periods as input and the actual traffic travel demand matrix of the next time period of the corresponding sequence as target output, and training the depth model by using a back propagation algorithm of a neural network;
step S2, constructing a transportation demand matrix sequence with rich context information;
and step S3, the depth model parameters obtained through training in the step S1 and the depth model are used as a final predictor, a continuous travel demand matrix sequence is input, and an unknown travel demand matrix of the next time period is predicted.
Preferably, the step S1 further includes:
s100, constructing a space-time situation demand machine CSTN based on a depth model extracted by multiple context information;
s101, constructing a traffic travel demand matrix sequence with rich context information, and taking traffic travel demand matrixes of a plurality of time periods continuous in historical time and external information of corresponding time periods as an input sequence of the space-time situation demand machine;
step S102, taking the traffic travel demand matrix sequences of a plurality of historical time periods and the external information of the corresponding time periods as input and the actual traffic travel demand matrix of the next time period of the corresponding sequence as target output, training a depth model by using a back propagation algorithm of a neural network, and updating parameters of each layer of the depth model.
Preferably, in step S100, the spatio-temporal context requirement machine CSTN includes a local spatial association sub-network, a time sequence evolution sub-network, and a global spatial cooperation sub-network, and the three sub-networks respectively model spatial, temporal, and global context contexts.
Preferably, step S101 includes:
acquiring vehicle GPS information and passenger carrying state information, and counting the traffic travel demand according to the vehicle state change;
dividing a grid Gird { H, W } of the urban area, and mapping the travel demand to a traffic travel demand matrix OD ∈ R according to a starting point and a destination index of the travel demandN×NAnd in the period, N is H × W, and the traffic travel demand matrix OD obtained by statistics in each time period is marked as Xt；
And taking the traffic travel demand matrix OD of a plurality of time periods which are continuous in time and the external information of the corresponding time period as the input sequence of the depth model.
Preferably, step S102 includes:
step S102a, extracting spatial local context information from the inter-area traffic travel demand matrix in the time slot by using the local spatial correlation sub-network;
step S102b, the time sequence evolution sub-network receives the spatial local context information of the historical moment extracted by the local spatial correlation sub-network, and predicts the spatial local context information of the travel demand at the next moment;
step S102c, the global space collaboration sub-network receives the output of the sequential evolution sub-network TEC as characteristic input, weights and characteristic information of other region input characteristics are added to the characteristics of different regions by extracting global situation characteristics and calculating the global characteristic similarity between the regions as weights, and the final characteristics are used for predicting the travel requirements between the regions;
and step S102d, performing backward propagation according to the prediction matrix and the error of the target output to obtain the network parameter gradient of the depth model, updating, and iterating the process until the depth model can accurately predict the travel demand matrix.
Preferably, in step S102a, the local spatial correlation sub-network learns spatial features of a departure region and an arrival region of travel demand, respectively, to obtain the spatial features of the departure region and the arrival region, then fuses the spatial features of the departure region and the arrival region, and finally fuses the spatial features of the departure region and the arrival region with features of external information.
Preferably, after the external information is subjected to feature expression through three fully-connected neural networks, a scalar of each dimension of the feature vector is copied to form a two-dimensional matrix, the two-dimensional matrix is stacked to form a feature cube, and the feature cube and the airspace feature are fused through a two-dimensional convolution layer.
Preferably, in step S102b, the time-series evolution sub-network takes a fusion feature of spatial features of a plurality of consecutive times and features of external information of corresponding times as an input of the time-series evolution sub-network, and the last time-series hidden layer h of the time-series evolution sub-networkt∈RC×W×HF is obtained by regression of the features through a convolution layerlt∈R(W×H)×W×HInput into the global spatial cooperative subnetwork, the time sequence evolution subnetwork is a convolution long-short term memory networkThe dimension of the hidden state generated at each step is kept consistent with the input dimension.
Preferably, in step S102c, the global spatial collaborative subnetwork is divided into three paths:
a) stretching
Has the dimension of
N＝H×W；
b) Using convolutional layer pairs FltPerforming characteristic compression and reforming to obtain characteristics
The dimension of the stretch s is
N is H multiplied by W and transposed to obtain
Performing dot multiplication on the two matrixes to obtain a correlation matrix S ', performing softmax operation on each row of S' to obtain a final global context correlation matrix S, and calculating FltThe point multiplication of the sum S yields a global context feature FgAnd adjusting FgHas the dimension of
c) Global context feature FgAnd FltAdding to obtain final comprehensive characteristics Fltg。
In order to achieve the above object, the present invention further provides an inter-area travel demand prediction apparatus, including:
the model construction unit is used for constructing a depth model extracted by multiple context information, taking a traffic travel demand matrix sequence of a plurality of historical time periods as input and an actual traffic travel demand matrix of the next time period of the corresponding sequence as target output, and training the depth model by using a back propagation algorithm of a neural network;
the system comprises a traffic travel demand matrix sequence construction unit, a traffic travel demand matrix sequence generation unit and a traffic travel demand matrix sequence generation unit, wherein the traffic travel demand matrix sequence construction unit is used for constructing a traffic travel demand matrix sequence with rich context information;
and the prediction unit is used for inputting a continuous traffic travel demand matrix sequence by taking the depth model parameters obtained by training of the model construction unit and the depth model as a final predictor and predicting an unknown traffic travel demand matrix of the next time period.
Compared with the prior art, the inter-area travel demand prediction method based on deep learning is characterized in that characteristics of an inter-area travel demand matrix sequence at historical time are extracted and fused by effectively combining multiple context information of time space and the whole world, the travel demands from all areas and from area to area at the next time of a research area are accurately predicted, the travel demands among a plurality of all areas at the next time can be predicted at the same time only by the inter-area travel matrix at the historical time, and the accuracy of inter-area travel demand prediction is improved.
Drawings
FIG. 1 is a flowchart illustrating steps of a method for predicting inter-area travel demand according to the present invention;
FIG. 2 is a schematic diagram of a depth model framework in an embodiment of the invention;
FIGS. 3a and 3b are schematic diagrams of city grid division, variation curves of a plurality of OD demands and a matrix of inter-area traffic travel demands, taking a Manhattan island as an example, according to an embodiment of the present invention;
FIG. 4 is a detailed schematic diagram of a depth model network according to an embodiment of the present invention;
FIG. 5 is a diagram of a transpose in a local spatial context extraction subnetwork in an embodiment of the present invention;
FIG. 6 is a flowchart illustrating a process of predicting inter-area travel demand according to an embodiment of the present invention;
FIG. 7 is a system configuration diagram of an inter-area travel demand prediction apparatus;
FIG. 8 is a detailed structural diagram of a model building unit according to an embodiment of the present invention;
FIG. 9 is a diagram illustrating a part of the result of predicting the demand for travel in New York City according to the embodiment of the present invention.
Detailed Description
Other advantages and capabilities of the present invention will be readily apparent to those skilled in the art from the present disclosure by describing the embodiments of the present invention with specific embodiments thereof in conjunction with the accompanying drawings. The invention is capable of other and different embodiments and its several details are capable of modification in various other respects, all without departing from the spirit and scope of the present invention.
Fig. 1 is a flowchart illustrating steps of a method for predicting inter-area travel demand according to the present invention. As shown in fig. 1, the method for predicting inter-area travel demand of the present invention includes the following steps:
step S1, constructing a depth model of multiple context information extraction, namely, a real-time empty-situation demand machine (CSTN), using a traffic demand matrix sequence of multiple historical time periods as an input and an actual traffic demand matrix of a next time period of the corresponding sequence as a target output, and training the depth model by using a back propagation algorithm of a neural Network to obtain a final predictor. Specifically, the space-time situation demand machine receives a traffic travel demand matrix sequence, extracts space-time context and global situation context characteristics, predicts a traffic travel demand matrix of the next time period, performs back propagation according to a prediction matrix and an error of target output to solve a depth model network parameter gradient, updates the depth model network parameter gradient, and iterates the process until the depth model can accurately predict the travel demand matrix.
Specifically, step S1 further includes:
and S100, constructing a space-time situation demand machine based on a depth model.
The space-time situation demand machine CSTN is a deep neural network for the fusion of time-space and global situation context information, and fully captures the context dependence relationship of inter-area travel demands in the space domain and the time domain, and the context information of global situations, such as the demand similarity of adjacent areas in the space, the regularity of the temporal demands and the relevant periodicity of the inter-functional-area travel demands.
In the specific embodiment of the present invention, as shown in fig. 2, the CSTN is composed of three sub-networks, which are respectively a local spatial correlation sub-network, a time sequence evolution sub-network and a global spatial coordination sub-network, the three sub-networks respectively model spatial, temporal and global context contexts, and simultaneously introduce external information into a depth model, and perform temporal context information extraction by fusing the external information with spatial features as time information, as shown in fig. 2, the three sub-networks are: 1) a Local Spatial Context modeling (LSC) for extracting Spatial Local Context information from an inter-area traffic travel demand matrix OD in a time period; 2) the system comprises a time sequence Evolution sub-network (TEC), a time sequence Evolution sub-network and a controller, wherein the TEC receives spatial local Context information of historical moments extracted by the LSC and predicts the spatial local Context information of travel demands at the next moment; 3) and a Global spatial collaborative Context (GCC), wherein the GCC receives the output of the TEC as a characteristic input, adds the weighting and characteristic information of other region input characteristics to the characteristics of different regions by extracting Global Context characteristics and calculating the Global characteristic similarity between the regions as a weight, and predicts the trip requirements between the regions by using the final characteristics.
Step S101, constructing a transportation demand matrix sequence OD with rich context information, and taking the transportation demand matrix OD of a plurality of continuous time periods in historical time and external information (such as weather, holidays and the like) of corresponding time periods as input sequences of a depth model.
Specifically, step S101 includes:
step S101a, obtaining vehicle GPS information and passenger carrying state information, and counting the traffic travel demand according to the vehicle state change, namely, dividing and counting according to the set time granularity to obtain the traffic travel demand in each time period;
step S101b, dividing the city Grid { H, W }, mapping the travel demand to a transportation travel demand matrix OD ∈ R according to the index of the departure point and the destination of the travel demand as shown in FIG. 3aN×NIn the above description, N is H × W, and as shown in fig. 3b, the travel demand matrix OD obtained by statistics in each time period is denoted as XtThe travel demand matrix for the time period t is Xt∈RW×H×W×HWherein W and H are the width and height of the divided region grid,
representing the number of travel demands from region (k, l) to region (m, n) at the t-th history period;
step S101c, the travel demand matrix OD of a plurality of time slots continuous in historical time and the external information M of the corresponding time slott(e.g., weather, holidays, etc.) as an input sequence for the depth model.
Step S102, taking the traffic travel demand matrix sequences of a plurality of historical time periods and the external information of the corresponding time periods as input and the actual traffic travel demand matrix of the next time period of the corresponding sequence as target output, training a depth model by using a back propagation algorithm of a neural network, and updating parameters of each layer of the depth model.
Specifically, step S102 includes:
in step S102a, the Local Spatial Context modeling (LSC) is used to extract Spatial Local Context information for the inter-area traffic demand matrix OD in the time segment. Specifically, the local spatial correlation sub-network LSC learns the spatial characteristics of a departure region and an arrival region of travel demands respectively to obtain the spatial characteristics of the departure region and the arrival region, and in the invention, when the spatial characteristics of the departure region or the arrival region of the travel demands are learned, the spatial characteristics of different departure regions of the same target region or the spatial characteristics of the target region of the same departure region are learned through a convolutional neural network; after extracting the airspace features of the departure region and the arrival region of the travel demand, finally fusing the two features by using the convolutional layer.
As shown in fig. 4, the local spatial correlation sub-network is composed of convolution layers, a transposition operation and a plurality of deformation operations, and the structure thereof is divided into two paths at first, and each path is composed of a plurality of convolution layers; as shown in FIG. 5, the inter-area travel demand matrix XiObtained by a transposition operation
Inputting two sub-networks of the local space correlation sub-networks, and respectively performing local space correlation characteristic extraction on different dimensions (departure area and arrival area) of the input traffic travel demand matrix; and stacking the two obtained features, and fusing the features by using a convolution layer to obtain the final spatial local context information expression.
Specifically, step S102a includes:
in step S102a1, spatial local context information of the departure area is extracted.
First, adjust Xi∈RW×H×W×HDimension of (2), Xi∈RN×W×HAnd N is H × W, then
Representing the travel demand from the region (k, l) to the region m in the ith time period; then changing XiSuch that X isi∈RN×W×HK consecutive (see fig. 4) convolutional layers with kernel size 3 × 3 are input, preferably K equals 3, to extract different features for the trip needs to different regions m, respectively. The obtained spatial local context information of the departure area is recorded as Fi o。
In step S102a2, the spatial local context information of the arrival area is extracted.
First, adjust Xi∈RW×H×W×HDimension of (2), Xi∈RN×W×HN — H × W, obtained by a transpose operation as shown in fig. 5
Then
Representing the travel demand from the region k to the region (m, n) in the ith time period; then change
Is such that
K consecutive (see fig. 4) convolutional layers with kernel size 3 × 3 are input, preferably K is 3, to extract different features for the travel needs to different regions K, respectively. Remembered spatial local context information F of the arrival regioni d。
Step S102a3, the spatial context information of the departure and arrival areas is fused to obtain the final local spatial context feature. I.e. finally, Fi oAnd Fi dStacking the layers, and passing through another convolution layer to obtain the final local spatial context characteristic Fi l。
The influence on the inter-area travel demand on the weather and the time information at the historical time is not negligible. Therefore, after the spatial context information of the departure and arrival regions is fused to obtain the final local spatial context feature, the spatial context feature is further fused with the feature of the external information, that is, the external information MiAfter feature expression is obtained through three fully-connected neural networks, the scalar quantity of each dimension of the feature vector is copied to form a two-dimensional matrix, and the two-dimensional matrix is stacked to form a feature cube Fi mAnd space domain feature Fi lFusing by two-dimensional convolution layers to obtain local spatial context characteristics
Step S102b, the Temporal Evolution Context Modeling (TEC) receives the spatial local Context information of the historical time extracted by the local spatial correlation sub-network LSC, and predicts the spatial local Context information of the travel demand at the next time. Preferably, a fusion feature of spatial domain features of a plurality of continuous time instants and features of external information of corresponding time instants is taken as an input of the time sequence evolution sub-network TEC.
In the present invention, the travel demand matrix X of historical time periods 0,1, …, t0,X1,...,XtExtracting local spatial context information of local spatial correlation sub-network to obtain characteristic sequence
As input to the timing evolution sub-network TEC. The sequential evolution sub-network TEC is composed of a convolution long-short term memory network, the convolution long-short term memory network controls the dynamic modeling of each sequential state on the characteristics of the previous time through 3 gates, the sequential evolution context information is extracted, and the input characteristic diagram of the ith time period is recorded as Fi∈RC×W×HWherein W, H, and C respectively represent the width, height, and number of channels of the feature map (it is stated that C of different subscripts in the following is only a channel representing different tensors). Hidden state HiCan be expressed as:
Hi＝ConvLSTM(Hi-1,Ci-1,Fi)
wherein C isi-1Is the memory state of ConvLSTM in the (i-1) th time period. Hidden state HiA dynamic modeling of the characteristics of the previous time instants is expressed.
Compared with the conventional long and short term memory network, the connection between the input and each department in the convolutional long and short term memory network ConvLSTM is a convolution operation, and the dependence between the states is also realized by a convolution operation, and the realization principle can be expressed by the following formula.
Wherein x represents the operation of convolution,
representing multiplication of corresponding elements of the matrix, it，ft，Ct，ot，HtRespectively representing an input gate, a forgetting gate, a cell state, an output gate and a hidden state, and the dimensionalities of the input gate, the forgetting gate, the cell state, the output gate and the hidden state are three-dimensional, XtFor the t-th input of the input sequence, W and b in each gate correspond to input XtHidden state HtAnd cell state CtThe weight parameter of the convolution or multiplication between and the bias.
The influence on the inter-area travel demand is not negligible corresponding to the weather and the time information of the historical time, and the inter-area travel demand of the spatial local context information extraction model in the historical time period 0,1, … and t is characterized
Using external information M corresponding to time of day0,M1,...,MtTo improve the predictive power of the model. As shown in FIG. 4, the output characteristic diagram of the local spatial correlation sub-network at the ith time point is
And the external information inputted is MtA) external information MtInputting a 3-layer full-connection layer for feature extraction to obtain a feature vector
b) Expanding each dimension of replication of feature vectors to Fi lThe size of the feature map of (2) is obtained as Fi m(ii) a c) F is to bei lAnd Fi mAre stacked into
Then, carrying out feature fusion through another convolution layer to obtain comprehensive features F of the time period ii lmThen the aforementioned ConvLSTM input becomes
After the characteristics of the historical time are sequentially input into ConvLSTM, the hidden state H of the last time is takentThe feature is enhanced by another two-dimensional convolution, the output feature is FltAnd (3) performing feature expression of the context information in the time domain as a feature sequence.
Step S102c, the global space collaboration sub-network receives the output of the sequential evolution sub-network TEC as a feature input, and adds the weighting and feature information of other region input features to the features of different regions by extracting global situation features and calculating the global feature similarity between the regions as a weight, and uses the final features to predict the travel requirements between the regions.
Specifically, the context features of the two previous sub-networks are extracted and the obtained features are recorded
It is used as the input of the global spatial collaboration sub-network. As shown in fig. 4, the global spatial collaboration sub-network is composed of a plurality of two-dimensional convolutional layers. The global spatial collaboration sub-network is divided into three paths,
a) stretching
Has the dimension of
N＝H×W；
b) Using convolutional layer pairs FltPerforming characteristic compression and reforming to obtain characteristics
Stretching FsHas the dimension of
N is H multiplied by W and transposed to obtain
The two are subjected to matrix dot multiplication to obtain a correlation matrix S ', each row of S' is subjected to softmax operation,
obtaining a final global context correlation matrix S, and calculating FltThe point multiplication of the sum S yields a global context feature FgAnd adjusting FgHas the dimension of
c) Global context feature FgAnd FltAdding to obtain final comprehensive characteristics Fltg。
General characteristics FltgInputting a1 x 1 convolutional layer regression to obtain the final characteristic u, and the inter-region travel demand of the next time period
tanh guarantees a range of output (-1, 1). The predicted values are then calculated according to the normalization method used.
And step S102d, performing backward propagation according to the prediction matrix and the error of the target output to obtain the network parameter gradient of the depth model, updating, and iterating the process until the depth model can accurately predict the travel demand matrix.
And step S2, constructing a transportation demand matrix sequence OD with rich context information.
Specifically, similar to step 101, this step includes:
step S200, acquiring vehicle GPS information and passenger carrying state information, and counting traffic travel demands according to vehicle state changes, namely dividing and counting according to set time granularity to obtain the traffic travel demands in each time period;
step S201, dividing a city Grid { H, W }, mapping a travel demand to a traffic travel demand matrix OD ∈ R according to a departure point and a destination index of the travel demandN×NIn the time period, N is H × W, as shown in fig. 3b, the travel demand matrix OD obtained by statistics in each time period is X, and the travel demand matrix for the time period i is Xi∈RW×H×W×HWherein W and H are the width and height of the divided region grid,
represents the number of travel demands from the area (k, l) to the area (m, n) in the i-th history period.
And step S3, the trained depth model parameters and the trained depth model are used as a final predictor, and a continuous travel demand matrix sequence is input to predict the unknown travel demand matrix of the next time period.
Fig. 6 is a flowchart illustrating an inter-area travel demand prediction process according to an embodiment of the present invention. The present invention is further explained with reference to fig. 6, and as shown in fig. 6, the inter-area travel demand prediction process is as follows:
dividing a city Grid { H, W }, collecting vehicle GPS information and passenger carrying state information, counting traffic travel demands according to vehicle state change, dividing according to selected time granularity to obtain the traffic travel demands in different time periods, and mapping according to departure region and arrival region indexes to obtain a traffic travel demand matrix OD belonging to R in different time periodsHW×HW；
Step two, constructing a depth model for extracting multiple Context information, namely a space-time situation demand machine (CSTN) which comprises three sub-networks, a) a Local space associated sub-Network (LSC) for extracting space-domain Context information, wherein the LSC is composed of a convolutional layer, a transposition and a deformation operation; b) a Temporal Evolution Context Modeling (TEC) used for extracting time domain Context information and composed of a two-dimensional convolution long and short memory neural network; c) a Global spatial coordination Context (GCC) for Global Context information learning, which is composed of two-dimensional convolutional layers. And (3) sequentially connecting the three sub-networks, wherein the output of the sub-networks is used as the input of the next sub-network, the LSC receives the traffic travel demand matrix sequence of the plurality of historical time periods obtained in the first step as the input, and the GCC output is the prediction of the travel demand OD matrix of the next time period. In addition, a small network is used for extracting the characteristics of external information, fusing the external information with a local space correlation sub-network and inputting the external information into a time sequence evolution sub-network, wherein the time sequence evolution sub-network consists of a full connection layer and a two-dimensional convolution layer;
thirdly, taking a traffic travel demand matrix sequence of a plurality of historical time periods as input and an actual traffic travel demand matrix of a corresponding sequence in a next time period as target output, training a depth model by using a back propagation algorithm of a neural network, receiving the traffic travel demand matrix sequence by a space-time situation demand machine, extracting space-time context and global situation context characteristics, predicting the traffic travel demand matrix of the next time period, performing back propagation according to errors of the prediction matrix and the target output to obtain a network parameter gradient of the depth model, updating, and iterating the process until the depth model can accurately predict the travel demand matrix;
and step four, the depth model parameters and the depth model obtained in the step three are used as a final predictor, and the unknown traffic travel demand matrix of the next time period can be accurately predicted for the input continuous traffic travel demand matrix sequence.
Fig. 7 is a schematic structural diagram of an inter-area travel demand prediction apparatus according to the present invention, and as shown in fig. 7, the inter-area travel demand prediction apparatus according to the present invention includes:
the model building unit 701 is configured to build a depth model for extracting multiple context information, namely, a continuous Spatial-Temporal Network (CSTN), train the depth model by using a back propagation algorithm of a neural Network, and obtain a final predictor, by using a traffic demand matrix sequence of a plurality of historical time periods as an input and an actual traffic demand matrix of a next time period of a corresponding sequence as a target output. Specifically, the space-time situation demand machine receives a traffic travel demand matrix sequence, extracts space-time context and global situation context characteristics, predicts a traffic travel demand matrix of the next time period, performs back propagation according to a prediction matrix and an error of target output to solve a depth model network parameter gradient, updates the depth model network parameter gradient, and iterates the process until the depth model can accurately predict the travel demand matrix.
Specifically, as shown in fig. 8, the model building unit 701 further includes:
the CSTN constructing unit 7010 is configured to construct a depth model-based spatio-temporal context requirement machine.
The space-time situation demand machine CSTN is a deep neural network for fusion of time-space and global situation context information, and fully captures the context dependence relationship of inter-area travel demands in space domain and time domain, and the context information of global situation, such as the demand similarity of adjacent areas in space, the regularity of temporal demands and the relevant periodicity of the travel demands in functional areas.
In the specific embodiment of the invention, the CSTN is composed of three sub-networks, namely a local space association sub-network, a time sequence evolution sub-network and a global space cooperation sub-network, wherein the three sub-networks respectively model the space domain, the time domain and the global context, and simultaneously introduce external information into the depth model and fuse the external information with the spatial characteristics as time information to extract the time domain context information.
The travel demand matrix sequence obtaining unit 7011 is configured to construct a travel demand matrix sequence OD with rich context information, and use the travel demand matrix OD of multiple time periods continuous in historical time and external information (such as weather, holidays, and the like) of corresponding time periods as input sequences of the depth model.
The traffic travel demand matrix sequence acquisition unit is specifically configured to:
acquiring vehicle GPS information and passenger carrying state information, and counting the traffic travel demands according to vehicle state changes, namely dividing and counting according to set time granularity to obtain the traffic travel demands in each time period;
dividing a city Grid (H, W), and mapping the travel demand to a traffic travel demand matrix OD ∈ R according to a departure point and a destination index of the travel demandN×NAnd in the time period, N is H × W, counting the travel demand matrix OD obtained in each time period, and recording the travel demand matrix for the time period t as Xt∈RW×H×W×HWherein W and H are the width and height of the divided region grid,
representing the number of travel demands from region (k, l) to region (m, n) at the t-th history period;
the input unit is used for taking the travel demand matrix OD of a plurality of time periods continuous in historical time and external information (such as weather, holidays and the like) of corresponding time periods as an input sequence of the depth model.
The model training unit 7012 is configured to train the depth model by using the traffic travel demand matrix sequence of a plurality of historical time periods and the external information of the corresponding time period as inputs, and the actual traffic travel demand matrix of the next time period of the corresponding sequence as a target output, and update parameters of each layer of the depth model by using a back propagation algorithm of the neural network.
Model training unit 7012 is specifically configured to:
1. local Spatial Context modeling (LSC)
The Local Spatial Context modeling (LSC) extracts Spatial Local Context information for the inter-area traffic travel demand matrix OD in the time segment. Specifically, the local spatial correlation sub-network LSC learns the spatial characteristics of a departure region and an arrival region of travel demands respectively to obtain the spatial characteristics of the departure region and the arrival region, and in the invention, when the spatial characteristics of the departure region or the arrival region of the travel demands are learned, the spatial characteristics of different departure regions of the same target region or the spatial characteristics of the target region of the same departure region are learned through a convolutional neural network; after spatial domain features of a departure region and an arrival region of a travel demand are extracted, the two features are fused by using a two-dimensional convolution layer.
The local spatial correlation sub-network consists of convolution layers, a transposition operation and a plurality of deformation operations, the structure of the local spatial correlation sub-network is divided into two paths at first, each path consists of a plurality of convolution layers, and preferably, K is 3; respectively inputting the input traffic travel demand matrix and the transpose thereof into the two sub-networks, and respectively extracting local spatial correlation characteristics of the departure area and the arrival area; and then stacking the two obtained features, and fusing the features by using a two-dimensional convolution layer to obtain the final spatial local context information expression.
1.1, extracting the spatial local context information of the starting area.
First, adjust Xi∈RW×H×W×HDimension of (2), Xi∈RN×W×HAnd N is H × W, then
Representing the travel demand from the region (k, l) to the region m in the ith time period; then changing XiSuch that X isi∈RN×W×HK consecutive convolutional layers with a kernel size of 3 × 3 are input, connected across layers (see fig. 4), preferably K ═ 3, to extract different features for the travel needs to different regions m, respectively. The obtained spatial local context information of the departure area is recorded as Fi o
1.2, extracting the spatial local context information of the arrival area.
Adjusting Xi∈RW×H×W×HDimension of (2), Xi∈RN×W×HN ═ hxw, by following the scheme in fig. 5The transpose operation of the display is obtained
Then
Representing the travel demand from the region k to the region (m, n) in the ith time period; then change
Is such that
K consecutive convolutional layers with kernel size 3 × 3 (see fig. 4) connected across the layers are input, preferably K3, to extract different features for the travel needs to different regions K, respectively. Remembered spatial local context information F of the arrival regioni d
And 1.3, fusing the spatial context information of the departure region and the arrival region to obtain the final local spatial context characteristics. I.e. finally, Fi oAnd Fi dStacking the layers, and passing through another convolution layer to obtain the final local spatial context characteristic Fi l。
The influence on the inter-area travel demand on the weather and the time information at the historical time is not negligible. Therefore, after the spatial context information of the departure and arrival regions is fused to obtain the final local spatial context feature, the spatial context feature is further fused with the feature of the external information, that is, the external information MiAfter feature expression is obtained through three fully-connected neural networks, the scalar quantity of each dimension of the feature vector is copied to form a two-dimensional matrix, and the two-dimensional matrix is stacked to form a feature cube Fi mAnd space domain feature Fi lFusing by two-dimensional convolution layers to obtain local spatial context characteristics
2. Time Evolution sub-network (TEC)
The Temporal Evolution Context Modeling (TEC) receives spatial local Context information of a historical moment extracted by the local spatial correlation sub-network LSC, and predicts travel demand spatial local Context information of a next moment. Preferably, a fusion feature of spatial domain features of a plurality of continuous time instants and features of external information of corresponding time instants is taken as an input of the time sequence evolution sub-network TEC.
In the present invention, the travel demand matrix X of historical time periods 0,1, …, t0,X1,...,XtExtracting local spatial context information of local spatial correlation sub-network to obtain characteristic sequence
As input to the timing evolution sub-network TEC. The sequential evolution sub-network TEC is composed of a convolution long-term and short-term memory network, the convolution long-term and short-term memory network controls the dynamic modeling of each sequential state on the characteristics of the previous moment through 3 gates, the sequential evolution context information is extracted, and the input characteristic diagram of the ith time period is assumed to be Fi∈RC×W×HWherein W, H and C respectively represent the width, height and channel number of the characteristic diagram. Hidden state HiCan be expressed as:
Hi＝ConvLSTM(Hi-1,Ci-1,Fi)
wherein C isi-1Is the memory state of ConvLSTM in the (i-1) th time period. Hidden state HiThe dynamic modeling of the characteristics of the previous moments is expressed, taking the hidden state H of the last momenttAnd (3) performing feature expression of the context information in the time domain as a feature sequence.
Compared with the conventional long and short term memory network, the connection between the input and each department in the convolutional long and short term memory network ConvLSTM is a convolution operation, and the dependence between the states is also realized by a convolution operation, and the realization principle can be expressed by the following formula.
Wherein x represents the operation of convolution,
representing multiplication of corresponding elements of the matrix, it，ft，Ct，ot，HtRespectively representing an input gate, a forgetting gate, a cell state, an output gate and a hidden state, and the dimensionalities of the input gate, the forgetting gate, the cell state, the output gate and the hidden state are three-dimensional, XtFor the t-th input of the input sequence, W and b in each gate correspond to input XtHidden state HtAnd cell state CtThe weight parameter of the convolution or multiplication between and the bias.
The influence on the inter-area travel demand is not negligible corresponding to the weather and the time information of the historical time, and the inter-area travel demand of the spatial local context information extraction model in the historical time period 0,1, … and t is characterized
Using external information M corresponding to time of day0,M1,...,MtTo improve the predictive power of the model. As shown in FIG. 4, assume that the output characteristics of the local spatially correlated sub-network at the ith time point are plotted as
And the external information inputted is MtA) external information MtInputting a 3-layer full-connection layer for feature extraction to obtain a feature vector
b) Expanding each dimension of replication of feature vectors to Fi lThe size of the feature map of (2) is obtained as Fi m(ii) a c) F is to bei lAnd Fi mAre stacked into
Then, carrying out feature fusion through another convolution layer to obtain comprehensive features F of the time period ii lmThen the aforementioned ConvLSTM input becomes
After the characteristics of the historical time are sequentially input into ConvLSTM, the hidden state H of the last time is takentThe feature is enhanced by another two-dimensional convolution, the output feature is FltAnd (3) performing feature expression of the context information in the time domain as a feature sequence.
3. Global spatial collaboration sub-network
The global space cooperation sub-network receives the output of the sequential evolution sub-network TEC as characteristic input, global situation characteristics are extracted, inter-area global characteristic similarity is calculated as weight, weighting and characteristic information of other area input characteristics are added to the characteristics of different areas, and final characteristics are used for predicting inter-area travel requirements.
Specifically, through the context feature extraction of the first two sub-networks, the obtained feature is recorded as
It is used as the input of the global spatial collaboration sub-network. As shown in FIG. 4, a global spatial collaboration sub-network is formed from a plurality of two-dimensional volumesAnd (4) laminating. The global spatial collaboration sub-network is divided into three paths,
a) stretching
Has the dimension of
N＝H×W；
b) Using convolutional layer pairs FltPerforming characteristic compression and reforming to obtain characteristics
Stretching FsHas the dimension of
N is H multiplied by W and transposed to obtain
The two are subjected to matrix dot multiplication to obtain a correlation matrix S ', each row of S' is subjected to softmax operation,
obtaining a final global context correlation matrix S, and calculating FltPoint multiplication of the sum S yields a global context feature
And adjust FgHas the dimension of
c) Global context feature FgAnd FltAdding to obtain final comprehensive characteristics Fltg。
General characteristics FltgInputting a1 x 1 convolutional layer regression to obtain the final characteristic u, and the inter-region travel demand of the next time period
tanh ensures output in the range of (-1),1). The predicted values are then calculated according to the normalization method used.
4. And (4) performing back propagation according to the prediction matrix and the error of target output to obtain the network parameter gradient of the depth model, updating, and iterating the process until the depth model can accurately predict the travel demand matrix.
A transit trip demand matrix sequence construction unit 702, configured to construct a transit trip demand matrix sequence OD with rich context information.
Specifically, similar to the transit trip demand matrix sequence obtaining unit, the transit trip demand matrix sequence constructing unit 702 is specifically configured to:
acquiring vehicle GPS information and passenger carrying state information, and counting the traffic travel demands according to vehicle state changes, namely dividing and counting according to set time granularity to obtain the traffic travel demands in each time period;
dividing a city Grid (H, W), and mapping the travel demand to a traffic travel demand matrix OD ∈ R according to a departure point and a destination index of the travel demandN×NAnd in the time period, N is H × W, counting the travel demand matrix OD obtained in each time period, and recording the travel demand matrix for the time period i as Xt∈RW×H×W×HWherein W and H are the width and height of the divided region grid,
representing the number of travel demands from the region (k, l) to the region (m, n) at the i-th history period;
the predicting unit 703 is configured to input a continuous travel demand matrix sequence by using the trained depth model parameters and the trained depth model as a final predictor, so as to predict an unknown travel demand matrix of the next time period.
Therefore, in the invention, the evolution analysis of the inter-area travel demand is focused firstly, and the travel demand matrix among areas in the next time period is deduced from the historical inter-area travel demand matrix without observing the travel demand state of the current time period and requiring the required data timeliness requirement; secondly, the method based on the depth model can utilize a large amount of traffic data to learn, does not need complicated characteristic engineering of manual design, and can synergistically predict the OD demand matrix among all the area pairs rather than the OD demand of a few hot area pairs; meanwhile, the spatial domain time domain and global context information are dynamically extracted by a multiple spatial and temporal context information fusion mechanism, so that model interpretability is improved, and prediction accuracy is also improved. Fig. 9 shows a part of the result of predicting the transportation demand in new york city by applying the present invention, and it can be seen that the predicted travel demand is substantially consistent with the actual transportation demand.
The foregoing embodiments are merely illustrative of the principles and utilities of the present invention and are not intended to limit the invention. Modifications and variations can be made to the above-described embodiments by those skilled in the art without departing from the spirit and scope of the present invention. Therefore, the scope of the invention should be determined from the following claims.
Claims (5)
1. A method for predicting inter-area travel demand comprises the following steps:
step S1, constructing a depth model with multiple context information extraction, using the traffic travel demand matrix sequence of multiple historical time periods as input and the actual traffic travel demand matrix of the next time period of the corresponding sequence as target output, and training the depth model by using a back propagation algorithm of a neural network;
step S2, constructing a transportation demand matrix sequence with rich context information;
step S3, the depth model parameters obtained through training in the step S1 and the depth model are used as a final predictor, a continuous travel demand matrix sequence is input, and an unknown travel demand matrix of the next time period is predicted;
step S1 further includes:
s100, constructing a space-time situation demand machine CSTN based on a depth model extracted by multiple context information;
s101, constructing a traffic travel demand matrix sequence with rich context information, and taking traffic travel demand matrixes of a plurality of time periods continuous in historical time and external information of corresponding time periods as an input sequence of the space-time situation demand machine;
step S102, taking a traffic travel demand matrix sequence of a plurality of historical time periods and external information of a corresponding time period as input and an actual traffic travel demand matrix of a next time period of the corresponding sequence as target output, training a depth model by using a back propagation algorithm of a neural network, and updating parameters of each layer of the depth model;
in step S100, the spatio-temporal context demand machine CSTN includes a local spatial correlation sub-network, a time sequence evolution sub-network, and a global spatial coordination sub-network, and the three sub-networks respectively model a spatial domain, a time domain, and a global context;
step S102 includes:
step S102a, extracting spatial local context information from the inter-area traffic travel demand matrix in the time slot by using the local spatial correlation sub-network;
step S102b, the time sequence evolution sub-network receives the spatial local context information of the historical moment extracted by the local spatial correlation sub-network, and predicts the spatial local context information of the travel demand at the next moment;
step S102c, the global space collaboration sub-network receives the output of the time sequence evolution sub-network as feature input, and adds the weighting and feature information of other region input features to the features of different regions by extracting global situation features and calculating the global feature similarity between the regions as weight, and uses the final features to predict the travel demand between the regions;
step S102d, performing backward propagation according to the prediction matrix and the error of target output to obtain the network parameter gradient of the depth model, updating, and iterating the process until the depth model can accurately predict the travel demand matrix;
in step S102a, the local spatial correlation sub-network learns the spatial features of the departure area and the arrival area of the travel demand, respectively, to obtain the spatial features of the departure area and the arrival area, then fuses the spatial features of the departure area and the arrival area, and finally fuses the spatial features of the departure area and the arrival area with the features of the external information.
2. The method for predicting demand for interregional travel according to claim 1, wherein the step S101 includes:
acquiring vehicle GPS information and passenger carrying state information, and counting the traffic travel demand according to the vehicle state change;
dividing a grid Gird { H, W } of the urban area, and mapping the travel demand to a traffic travel demand matrix OD ∈ R according to a starting point and a destination index of the travel demandN×NAnd in the period, N is H × W, and the traffic travel demand matrix OD obtained by statistics in each time period is marked as XtWherein W and H are the width and height of the divided region grid;
and taking the traffic travel demand matrix OD of a plurality of time periods which are continuous in time and the external information of the corresponding time period as the input sequence of the depth model.
3. The method of predicting demand for interregional travel according to claim 1, wherein: after the external information is subjected to feature expression through three fully-connected neural networks, the scalar quantity of each dimension of the feature vector is copied to form a two-dimensional matrix, the two-dimensional matrix is stacked to form a feature cube, and the feature cube is fused with the airspace features through a two-dimensional convolution layer.
4. The method of predicting demand for interregional travel according to claim 1, wherein: in step S102b, the time-series evolution sub-network takes the fusion feature of the spatial domain features of a plurality of consecutive times and the feature of the external information of the corresponding time as the input of the time-series evolution sub-network, and the hidden layer H of the last time series of the time-series evolution sub-networkt∈RC×W×HF is obtained by regression of the features through a convolution layerlt∈R(W×H)×W×HInputting the data into the global space cooperative sub-network, wherein the time sequence evolution sub-network is a convolution long-term and short-term memory network, and the generated concealment of each stepThe dimension of the state remains consistent with the input dimension.
5. The method according to claim 4, wherein the inter-area travel demand prediction method comprises: in step S102c, the global spatial collaboration sub-network is divided into three paths:
a) stretching
Has the dimension of
N＝H×W；
b) Using convolutional layer pairs FltPerforming characteristic compression and reforming to obtain characteristics
The dimension of the stretch s is
Is transposed to obtain
Performing dot multiplication on the two matrixes to obtain a correlation matrix S ', performing softmax operation on each row of S' to obtain a final global context correlation matrix S, and calculating FltThe point multiplication of the sum S yields a global context feature FgAnd adjusting FgHas the dimension of
c) Global context feature FgAnd FltAdding to obtain final comprehensive characteristics Fltg。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810864027.3A CN109190795B (en) | 2018-08-01 | 2018-08-01 | Inter-area travel demand prediction method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810864027.3A CN109190795B (en) | 2018-08-01 | 2018-08-01 | Inter-area travel demand prediction method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109190795A CN109190795A (en) | 2019-01-11 |
| CN109190795B true CN109190795B (en) | 2022-02-18 |
Family
ID=64920250
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810864027.3A Active CN109190795B (en) | 2018-08-01 | 2018-08-01 | Inter-area travel demand prediction method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109190795B (en) |
Families Citing this family (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110060471B (en) * | 2019-04-01 | 2020-09-08 | 长安大学 | Vehicle OD flow prediction model construction method and vehicle OD flow prediction method |
| CN111784018A (en) * | 2019-04-03 | 2020-10-16 | 北京嘀嘀无限科技发展有限公司 | Resource scheduling method and device, electronic equipment and storage medium |
| CN109978279B (en) * | 2019-04-10 | 2023-05-02 | 青岛农业大学 | Ocean surface temperature area prediction method |
| CN110119838A (en) * | 2019-04-17 | 2019-08-13 | 成都信息工程大学 | A kind of shared bicycle demand forecast system, method and device |
| CN110210664B (en) * | 2019-05-29 | 2020-07-24 | 东南大学 | Deep learning method for short-term prediction of using behaviors of multiple individual vehicles |
| CN110298550A (en) * | 2019-05-30 | 2019-10-01 | 特斯联（北京）科技有限公司 | Community resident's service dispatch system and method based on grid cell |
| CN110322064B (en) * | 2019-06-27 | 2022-04-15 | 清华大学 | Urban trip demand prediction method |
| CN111862585B (en) * | 2019-07-23 | 2021-11-02 | 北京嘀嘀无限科技发展有限公司 | System and method for traffic prediction |
| CN111160747B (en) * | 2019-12-23 | 2022-07-22 | 北京百度网讯科技有限公司 | Scheduling method and device for unmanned robot taxi and electronic equipment |
| CN111753910A (en) * | 2020-06-27 | 2020-10-09 | 长安大学 | Method and device for predicting drip order demand based on LSTM |
| CN112102239A (en) * | 2020-08-10 | 2020-12-18 | 北京工业大学 | Image processing method and system for full-layer brain CT image |
| CN111815075B (en) * | 2020-08-27 | 2021-01-15 | 北京航空航天大学 | Prediction method for transportation travel demand under major public health incident |
| CN112613630B (en) * | 2021-01-05 | 2022-09-20 | 大连理工大学 | Short-term traffic demand prediction method integrating multi-scale space-time statistical information |
| CN113128668B (en) * | 2021-04-07 | 2023-07-25 | 中山大学 | Link scheduling method considering high throughput and fairness in data center network |
| CN113837383B (en) * | 2021-10-18 | 2023-06-23 | 中国联合网络通信集团有限公司 | Model training method and device, electronic equipment and storage medium |
| CN114327840B (en) * | 2022-03-11 | 2022-06-10 | 北京骑胜科技有限公司 | Data processing method and data processing device |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105389584A (en) * | 2015-10-13 | 2016-03-09 | 西北工业大学 | Streetscape semantic annotation method based on convolutional neural network and semantic transfer conjunctive model |
| CN105701571A (en) * | 2016-01-13 | 2016-06-22 | 南京邮电大学 | Short-term traffic flow prediction method based on nerve network combination model |
| CN106297297A (en) * | 2016-11-03 | 2017-01-04 | 成都通甲优博科技有限责任公司 | Traffic jam judging method based on degree of depth study |
| CN106934497A (en) * | 2017-03-08 | 2017-07-07 | 青岛卓迅电子科技有限公司 | Wisdom cell power consumption real-time predicting method and device based on deep learning |
| CN106991510A (en) * | 2017-05-31 | 2017-07-28 | 福建江夏学院 | A kind of method based on the traffic accident of spatial-temporal distribution characteristic predicted city |
| CN107103758A (en) * | 2017-06-08 | 2017-08-29 | 厦门大学 | A kind of city area-traffic method for predicting based on deep learning |
| CN107180530A (en) * | 2017-05-22 | 2017-09-19 | 北京航空航天大学 | A kind of road network trend prediction method based on depth space-time convolution loop network |
-
2018
- 2018-08-01 CN CN201810864027.3A patent/CN109190795B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105389584A (en) * | 2015-10-13 | 2016-03-09 | 西北工业大学 | Streetscape semantic annotation method based on convolutional neural network and semantic transfer conjunctive model |
| CN105701571A (en) * | 2016-01-13 | 2016-06-22 | 南京邮电大学 | Short-term traffic flow prediction method based on nerve network combination model |
| CN106297297A (en) * | 2016-11-03 | 2017-01-04 | 成都通甲优博科技有限责任公司 | Traffic jam judging method based on degree of depth study |
| CN106934497A (en) * | 2017-03-08 | 2017-07-07 | 青岛卓迅电子科技有限公司 | Wisdom cell power consumption real-time predicting method and device based on deep learning |
| CN107180530A (en) * | 2017-05-22 | 2017-09-19 | 北京航空航天大学 | A kind of road network trend prediction method based on depth space-time convolution loop network |
| CN106991510A (en) * | 2017-05-31 | 2017-07-28 | 福建江夏学院 | A kind of method based on the traffic accident of spatial-temporal distribution characteristic predicted city |
| CN107103758A (en) * | 2017-06-08 | 2017-08-29 | 厦门大学 | A kind of city area-traffic method for predicting based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109190795A (en) | 2019-01-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109190795B (en) | Inter-area travel demand prediction method and device | |
| Chen et al. | A novel fuzzy deep-learning approach to traffic flow prediction with uncertain spatial–temporal data features | |
| Bi et al. | Daily tourism volume forecasting for tourist attractions | |
| US11270579B2 (en) | Transportation network speed foreeasting method using deep capsule networks with nested LSTM models | |
| CN111860951B (en) | Rail transit passenger flow prediction method based on dynamic hypergraph convolutional network | |
| Ye et al. | Meta graph transformer: A novel framework for spatial–temporal traffic prediction | |
| CN112508173A (en) | Traffic space-time sequence multi-step prediction method, system and storage medium | |
| Zhang et al. | A Traffic Prediction Method of Bicycle-sharing based on Long and Short term Memory Network. | |
| CN113762595B (en) | Traffic time prediction model training method, traffic time prediction method and equipment | |
| CN115204478A (en) | Public traffic flow prediction method combining urban interest points and space-time causal relationship | |
| CN114692984B (en) | Traffic prediction method based on multi-step coupling graph convolution network | |
| CN114495500B (en) | Traffic prediction method based on dual dynamic space-time diagram convolution | |
| CN115376317B (en) | Traffic flow prediction method based on dynamic graph convolution and time sequence convolution network | |
| Huang et al. | Identification and prediction of urban traffic congestion via cyber-physical link optimization | |
| Zhonghe et al. | Consensus based approach to the signal control of urban traffic networks | |
| Jin et al. | An agent-based traffic recommendation system: Revisiting and revising urban traffic management strategies | |
| Xia et al. | A parallel NAW-DBLSTM algorithm on Spark for traffic flow forecasting | |
| Xiong et al. | DCAST: a spatiotemporal model with DenseNet and GRU based on attention mechanism | |
| CN117195976A (en) | Traffic flow prediction method and system based on layered attention | |
| CN115762147B (en) | Traffic flow prediction method based on self-adaptive graph meaning neural network | |
| CN115080795A (en) | Multi-charging-station cooperative load prediction method and device | |
| CN114997464A (en) | Popularity prediction method based on graph time sequence information learning | |
| CN114385910A (en) | Knowledge tracking based online learning content recommendation method and system | |
| Chen et al. | Multitime resolution hierarchical attention-based recurrent highway networks for taxi demand prediction | |
| CN112508170A (en) | Multi-correlation time sequence prediction system and method based on generation countermeasure network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |