CN110310479A - A kind of Forecast of Urban Traffic Flow forecasting system and method - Google Patents
A kind of Forecast of Urban Traffic Flow forecasting system and method Download PDFInfo
- Publication number
- CN110310479A CN110310479A CN201910534104.3A CN201910534104A CN110310479A CN 110310479 A CN110310479 A CN 110310479A CN 201910534104 A CN201910534104 A CN 201910534104A CN 110310479 A CN110310479 A CN 110310479A
- Authority
- CN
- China
- Prior art keywords
- model
- flow
- period
- unit time
- output stream
- 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
Classifications
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0125—Traffic data processing
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0125—Traffic data processing
- G08G1/0129—Traffic data processing for creating historical data or processing based on historical data
Landscapes
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention discloses a kind of Forecast of Urban Traffic Flow forecasting system and method, which models different time dimension by using three models, and it is seasonal to solve the problems, such as that the conventional time series model needs magnitude of traffic flow is presented.The present invention optimizes residual unit on the basis of ResNet model simultaneously, such as BN layers of addition, and the module unit design philosophy for using for reference MobileNetV2 optimizes its mode for extracting feature, can be improved the precision of prediction model.Finally using RMSE as the evaluation criteria of model, the data instance announced with certain Chinese city, it is more excellent that the forecasting system compares the performance of ARIMA, SARIMA, VAR, RNN, LSTM and GRU effect, so as to extract the Forecast of Urban Traffic Flow actual time well, the feature in space solves the problems, such as that time-space data analysis method existing for existing model is bad and itself existing defects.
Description
Technical field
The present invention relates to traffic management technology fields, more particularly to a kind of Forecast of Urban Traffic Flow forecasting system and method.
Background technique
Recently as science and technology, expanding economy, urban population density is increasing, possesses the number of automobile also year by year
Increase, but people are unable to get the raising of trip demand and timely meet.That is, the speed of transport development is caught up with
The not upper increased degree of traffic congestion.The requirement that people experience traffic is also higher and higher, no matter the congestion level of trip, also
Be traffic safety be all traffic experience important indicator, under the premise of can not improve immediately traffic hardware facility, rationally
Trip strategy and intelligentized traffic administration will be expected to improve traffic quality.In recent years, not due to computer performance
Disconnected to be promoted, Successful utilization and is proved to be the artificial intelligence technology of forefront into many applications to deep learning.It will
These Technology applications to space-time data to a series of space-time using extremely important, will be expected to including urban transportation, city peace
Entirely, urban planning etc. brings breakthrough progress, to carry out the traffic trip environment an of blocking-resistant safety to urban belt.
Predicting traffic flow amount problem can regard a spatio-temporal prediction problem as.Traffic flow forecasting initial stage can be divided into the recent period
And long-range forecasting.Last-period forecast can use mathematical statistics method, data processing be carried out to the magnitude of traffic flow investigated in the past, according to it
The rule of development and trend calculate and obtain prediction result.Long-range forecasting can use traffic prediction model, and to future city, resident goes out
Comprehensive analysis and research are made in row activity, are planned according to land use situation and Traffic Net, by the distribution of flow of the people, hand over
The division of logical mode and etc. predict traffic forecast.The deficiency of traditional and recent prediction technique is as follows:
(1) traditional time series predicting model
Common model has history averaging model, sliding autoregression averaging model (ARIMA), seasonal difference autoregression sliding
Dynamic averaging model (SARIMA), venture worth model (VAR), artificial nerve network model (ANN) etc..History averaging model can not
Dynamic event (such as traffic accident) is responded；ARIMA is needed to be treated differently before analysis, be not suitable for having missing values when
Between sequence；SARIMA is disadvantageous in that it is very time-consuming；VAR can ignore the relationship between predicted value and residual error；The line of ANN
Property modeling ability is not good enough.
(2) recent deep neural network model
Application of the recent deep neural network model on time, space achieves more achievement and good hair
Exhibition, however, convolutional neural networks (CNN) and Recognition with Recurrent Neural Network (RNN) all can only capture space or time single dependences
Property.Even if recent shot and long term memory network (LSTM) can simultaneously learning time and spatial dependence, but its can not model it is very deep
Network and very long dependence.
The reason of prior art Shortcomings is: traffic data is space-time data, complex, the factor that can be influenced compared with
It is more, and there is biggish uncontrollabilities for actual traffic conditions.Furthermore prediction model itself existing defects, the side solved the problems, such as
To single, and there are contradictions for complicated traffic data, therefore existing model is only reducing error as much as possible, and can not be good
The characteristic rule of actual traffic is extracted well.
Due to the specific regional subjective factor of crowd and the behavioural characteristic of trip.Meanwhile the space of the magnitude of traffic flow in city
Characteristic rule is sufficiently complex, and usually in non-linear, factors of the corresponding time series feature generally also because of many complexity
Be it is non-linear and uncontrollable, it is inadequate usually to there is precision in traditional and existing model, therefore, it is necessary to traffic space-time data
Complexity, it is non-linear that the physical meaning in conjunction with traffic is gone using new method, seek its characteristic rule and analyzed and studied.
Summary of the invention
The invention mainly solves the technical problem of providing a kind of Forecast of Urban Traffic Flow forecasting system and methods, can be fine
Ground extracts the Forecast of Urban Traffic Flow actual time, and the feature in space solves time-space data analysis method existing for existing model not
The problem of good and itself existing defects.
In order to solve the above technical problems, one technical scheme adopted by the invention is that: it is pre- to provide a kind of Forecast of Urban Traffic Flow
Examining system, comprising: data collection module, before according to cellphone GPS signal and/or vehicle GPS signal collection current date
The tracing point of urban transportation in predetermined amount of time counts the inlet flow and output stream of each unit time period in each net region, and
Inlet flow and output stream are integrated into a tensor, wherein inlet flow is not in net region but to tie when starting the unit period
The sum of all tracing points when beam in net region, output stream be the unit period when starting in net region but at the end of not
The sum of all tracing points in net region, urban area are divided into multiple net regions, several unit time periods are one
It；Data preprocessing module for the characteristic dimension of inlet flow and output stream to be zoomed to the range of prediction model requirement, and is selected
A date is selected as target date, screens the tensors of preceding 3 unit time periods of target date as recent input data, previous
It tensor as periodical input data, the tensor of the last week as trend input data, using the tensor of target date as surveying
Try data；Model training module obtains recent output stream for extracting feature to recent input data by recent model
Vector extracts feature to periodical input data by periodic model and obtains period output flow vector, passes through trend model pair
Trend input data extracts feature and obtains trend output flow vector, by recent output stream vector, period output stream vector sum
Trend output flow vector integrate the simulating traffic figure of target date, by the feature ruler of the simulating traffic figure of target date
Degree zooms to the range of prediction model requirement, reversely adjusts institute using the simulating traffic figure of target date and the error of test data
State prediction model, wherein the recent model, periodic model, trend model are all made of identical depth residual error neural network frame
Structure；Model evaluation module, for assessing the prediction model using root-mean-square error test；Traffic forecast module, for utilizing
The prediction model exports the predicted flow rate figure of the following predetermined time, predicted flow rate figure is zoomed to initial value, and use thermodynamic chart
Visable representation is done to the predicted flow rate figure.
Preferably, the data preprocessing module be also used to search missing tracing point to completion unit time period, and judge
Unit time period to the same time of predetermined number of days before completion unit time period whether there is tracing point, when there are tracing point,
Completion unit time period is treated by SES function according to the inlet flow of the unit time period of the same time of the continuous predetermined number of days
Inlet flow is filled, and passes through SES function pair according to the output stream of the unit time period of the same time of the continuous predetermined number of days
Output stream to completion unit time period is filled, and when tracing point is not present, will be to day where completion unit time period
Whole track point deletions.
Preferably, the SES function representation is as follows:
Ft+1=α xt+(1-α)Ft
Wherein, t is time current period, and α is coefficent of exponential smoothing, xtFor the actual observation value of t phase, FtFor the prediction of t phase
Value, Ft+1For the predicted value of t+1 phase.
Preferably, the inlet flow indicates are as follows:
The output stream indicates are as follows:
Wherein, Tt is the track of line at the beginning and end of unit time period, and i indicates the length of net region, and j indicates grid regions
The width in domain, P indicate time interval tthTrack set, gk-1Indicate tracing point when unit time period starts, gkIndicate unit time period
At the end of tracing point, K indicate tracing point number；
The tensor representation are as follows:
X∈R2×I×J。
Preferably, the feature scaling algorithm that the data preprocessing module uses is as follows:
X=x*2-1
Wherein, xmaxIndicate the maximum value in all data, xminIndicate the minimum value in all data；
Initial value retrieving algorithm corresponding with feature scaling algorithm is as follows:
X=x* (xmax-xmin)+xmin。
Preferably, the depth residual error neural network framework successively includes 3*3 convolutional layer, 12 residual units, 3*3 convolution
Layer, BN layers, relu activation primitive, 3*3 convolutional layer, the merge of residual error, activation primitive and convolutional layer.
Preferably, the prediction model is trained using dynamic sliding time window.
Preferably, the root-mean-square error indicates are as follows:
Wherein, n indicates the quantity of test data, wherein xiWithRespectively indicate the true value and corresponding emulation of test data
Value.
In order to solve the above technical problems, another technical solution used in the present invention is: providing a kind of Forecast of Urban Traffic Flow
Prediction technique, comprising: S1: according in predetermined amount of time before cellphone GPS signal and/or vehicle GPS signal collection current date
The tracing point of urban transportation, counts the inlet flow and output stream of each unit time period in each net region, and by inlet flow and defeated
Stream is integrated into a tensor out, wherein inlet flow be the unit period when starting not in net region but at the end of in grid regions
The sum of all tracing points in domain, output stream be the unit period when starting in net region but at the end of not in net region
The sum of all tracing points, urban area is divided into multiple net regions, several unit time periods are one day；S2: it will input
Stream and the characteristic dimension of output stream zoom to the range of prediction model requirement, and select a date as target date, screening
The tensor of preceding 3 unit time periods of target date is as recent input data, the tensor of the previous day as periodical input data, preceding
One week tensor is as trend input data, using the tensor of target date as test data；S3: by recent model to recent
Input data extracts feature and obtains exporting flow vector in the recent period, extracts feature to periodical input data by periodic model
Period output flow vector is obtained, feature is extracted to trend input data by trend model and obtains trend output flow vector,
Recent output stream vector, period output stream vector sum trend output flow vector integrate the simulating traffic of target date
The characteristic dimension of the simulating traffic figure of target date, is zoomed to the range of prediction model requirement by figure, utilizes imitating for target date
The error of true flow diagram and test data reversely adjusts the prediction model, wherein the recent model, periodic model, trend
Model is all made of identical depth residual error neural network framework；S4: the prediction model is assessed using root-mean-square error test；
S5: the predicted flow rate figure of the following predetermined time is exported using the prediction model, predicted flow rate figure is zoomed into initial value, will be predicted
Flow diagram zooms to initial value, and does visable representation to the predicted flow rate figure using thermodynamic chart.
It is in contrast to the prior art, the beneficial effects of the present invention are:
Fitting space-time can not be removed using deep layer network in the case where guaranteeing precision by solving tradition and existing prediction model
Data, the problem of finding its spatial homing.Different time dimension is modeled by using three models, when solving tradition
Between series model require the magnitude of traffic flow that seasonal problem is presented.The present invention optimizes residual on the basis of ResNet model simultaneously
Poor unit, such as BN layers of addition, the module unit design philosophy for using for reference MobileNetV2 optimize its mode for extracting feature, Neng Gouti
The precision of high prediction model.And during training pattern, the training method of dynamic sliding time window is taken, not only
Trained speed can be greatly improved, the training time of model is reduced, and improves the precision of prediction of model, Ke Yigao
Obtain a good prediction model to effect.Finally announced using RMSE as the evaluation criteria of model with certain Chinese big city
Data instance, the forecasting system compare ARIMA, SARIMA, VAR, RNN, LSTM and the performance of gating cycle unit (GRU) effect more
It is excellent.
Detailed description of the invention
Fig. 1 is the configuration diagram of the Forecast of Urban Traffic Flow forecasting system of the embodiment of the present invention.
Fig. 2 is the data definition figure of Forecast of Urban Traffic Flow forecasting system.
Fig. 3 is the magnitude of traffic flow exemplary diagram in certain city.
Fig. 4 is the structure chart for the prediction model that the present invention uses.
Fig. 5 is that the Forecast of Urban Traffic Flow forecasting system of the embodiment of the present invention carries out the effect picture of short-term forecast.
Fig. 6 is that the Forecast of Urban Traffic Flow forecasting system of the embodiment of the present invention carries out the effect picture of long-term forecast.
Fig. 7 is the flow diagram of the Forecast of Urban Traffic Flow prediction technique of the embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that the described embodiments are merely a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
It is the configuration diagram of the Forecast of Urban Traffic Flow forecasting system of the embodiment of the present invention referring to Fig. 1.The city of the present embodiment
City's traffic flow forecasting system includes data collection module 10, data preprocessing module 20, model training module 30, model evaluation
Module 40 and traffic forecast module 50.
Data collection module 10 is used for according to predetermined before cellphone GPS signal and/or vehicle GPS signal collection current date
The tracing point of urban transportation in period counts the inlet flow and output stream of each unit time period in each net region, and will be defeated
Become a mandarin and be integrated into a tensor with output stream, wherein inlet flow be the unit period when starting not in net region but at the end of
The sum of all tracing points in net region, output stream be the unit period when starting in net region but at the end of not in net
The sum of all tracing points in lattice region, urban area are divided into multiple net regions, several unit time periods are one day.This
The urban area that invention divides is center city, because of the lesser region of the too remote or density of population, practical application is not present
Value.
Value in each net region represents the size of the pixel of flow diagram, while also representing traffic in the net region
The inlet flow of flow and the size of output stream.Referring to Fig.2, being the data definition figure of Forecast of Urban Traffic Flow forecasting system.Such as Fig. 2
It (a) is inlet flow and output stream shown in.Inlet flow is the traffic total flow for entering a region in given time section.It is defeated
Stream is the traffic total flow flowed out in a period of time from a region out.Inlet flow and output stream can pass through the number of disengaging pedestrian
Mesh, the number of automobile, the number of bus estimate.An example is given in Fig. 2 (b), we can use mobile phone
Signal estimates the number of pedestrian, shows that the inlet flow of people and output stream are 3 and 1 respectively in the region r2.Similar, use vehicle
GPS track, the inlet flow of vehicle and output stream are respectively 0 and 3 in the region r2.Therefore, inlet flow in total and output flow point
It is not 3 and 4.
It is the magnitude of traffic flow exemplary diagram in certain city refering to Fig. 3.In the present embodiment, the urban area in certain city is divided into
32 × 32 net region, the statistics of data collection module 10 obtain inlet flow and the output of each net region unit time period
Stream.Inlet flow indicates are as follows:
Output stream indicates are as follows:
Wherein, Tt is the track of line at the beginning and end of unit time period, and i indicates the length of net region, and j indicates grid regions
The width in domain, P indicate time interval tthTrack set, gk-1Indicate tracing point when unit time period starts, gkIndicate unit time period
At the end of tracing point, K indicate tracing point number；
Tensor representation are as follows:
X∈R2×I×J
That is, the inlet flow and output stream of each unit time period in each net region (for example, 30 minutes) are counted,
The matrix of an i × j is obtained after having counted, is a twin-channel flow diagram by the matrix conversion of i × j, be denoted as 2 × i ×
J respectively indicates the inlet flow and output stream of unit time period wherein 2 be port number i.e. binary channels, which is by flow number
According to the data format for being converted into tensor.
Data preprocessing module 20 is used to zoom to the characteristic dimension of inlet flow and output stream the model of prediction model requirement
It encloses, and selects a date as target date, screen the tensor of preceding 3 unit time periods of target date as input number in the recent period
According to the tensor of, the previous day as periodical input data, the tensor of the last week as trend input data, by the tensor of target date
As test data.
Wherein, why feature scaling is carried out to inlet flow, the characteristic dimension of output stream, is because in face of multidimensional characteristic
When problem, needing to guarantee these features all has similar scale, so as to help gradient descent algorithm quickly to receive
It holds back, to improve efficiency of algorithm.In the present embodiment, the range that prediction model requires is between -1 to 1.In the present embodiment,
The feature scaling algorithm that data preprocessing module 20 uses is as follows:
X=x*2-1
Wherein, xmaxIndicate the maximum value in all data, xminIndicate the minimum value in all data；
Initial value retrieving algorithm corresponding with feature scaling algorithm is as follows:
X=x* (xmax-xmin)+xmin
That is, can use feature scaling algorithm when carrying out feature scaling and zoom in and out, needing data also
When original is at initial value, then it can use initial value retrieving algorithm and restored.
Data before input data, periodical input data, trend input data are all target dates in the recent period, and all to open
The form of spirogram stores, and for example, test data is the data of target date 2013071201, and storage format is [2] [32]
[32], then the data that recent input data is 2013071148,2013071147,2013071146, storage format are [6] [32]
[32]；The data that periodical input data are 2013071101 to 2013071148, storage format are [2] [32] [32]；Trend is defeated
Enter the data that data are 2013070501 to 2013071148, storage format is [2] [32] [32].
Model training module 30 is exported in the recent period for extracting feature to recent input data by recent model
Flow vector extracts feature to periodical input data by periodic model and obtains period output flow vector, passes through trend model
Feature is extracted to trend input data and obtains trend output flow vector, recent output stream vector, period are exported into flow vector
The simulating traffic figure for integrate with trend output flow vector target date, by the feature of the simulating traffic figure of target date
The range that scaling is required to prediction model is reversely adjusted using the simulating traffic figure of target date and the error of test data
Prediction model, wherein recent model, periodic model, trend model are all made of identical depth residual error neural network framework.
In the present embodiment, depth residual error neural network framework successively includes 3*3 convolutional layer, 12 residual units, 3*3 volumes
Lamination, BN (Batch Normalization) layer, relu activation primitive, 3*3 convolutional layer, the merge of residual error, activation primitive and
Convolutional layer.Recent input data, periodical input data, trend input data are after three models, finally in three models
Merge is linked to full articulamentum later, then by tanh activation primitive, the characteristic dimension of the simulating traffic figure of target date is contracted
Be put into it is consistent with the characteristic dimension of data prediction scaling between [- 1,1], thus obtain meeting prediction data form as a result,
To be compared with actual result, reversed adjusting prediction model parameters are trained.
It is the structure chart for the prediction model that the present invention uses refering to Fig. 4.Three models rely on different time and carry out feature
It extracts, is finally integrated, the range then required by tanh Function Mapping to prediction model, to obtain simulating traffic
Figure.
The present invention proposes addition BN on the basis of the block of ResNet and meets the residual unit of data demand, then
The residual unit that ResNet model is optimized in conjunction with the module unit of Mobilenet-V2 model has borrowed lightweight network
The thought " expansion → convolution proposes feature → compression " of MobileNet-V2 model module unit, by the convolution unit of original residual error network
Convolution mode 2 → 4 → 64 be changed to 2 → 256 → 64 so that model accuracy promoted.The recent convolution principle of this paper formalizes table
Up to for (period and trend model are similar):
Wherein what X was referred to is the data and target data of recent model, and wherein * represents convolution, and f represents activation primitive, Wc (1)And bc (1)Represent the parameter for needing to learn.
ResNet model form of the invention is expressed as (period and trend model are same):
WhereinRepresent all learning parameters that L layers of residual unit are included.
So as under the premise of depth is increased, guarantee that accuracy rate will not decline.
In the present embodiment, prediction model is trained using dynamic sliding time window.Specifically, first by data
It arranges sequentially in time, then defines the time slide window of dynamic size, to the Time of Day addition time in data
Window size dynamic setting is 21 days to 35 days (3 weeks to 5 weeks), every time when screening window size, successively by correlation calculations
Compare the correlation between the corresponding time data of each number of days, and sum to it, be finally averaging, then makes even homogeneous
The maximum window size of closing property, is then trained.Then 1 week data are as test set backward for window, then at the beginning of the data
Beginning position starts to slide backward, and slides backward 1 week time every time, window size dynamic select is then done, finally until training
Data deficiencies terminates, and completes the training of model.Model training speed greatly is optimized, specific algorithm process can be the time
The calculating of sliding window optimization algorithm or the dynamic sliding time window degree of correlation.
The encoding examples of time slide window optimization algorithm are as follows:
INPUT:
The trainData: // preprocessed good data for training pattern
OUTPUT:
{ model } // trained model, later period can call directly its method, preservation model parameter
1 index=0 // start to index
2 while (index+size*48 < len (trainData)): // terminate training condition
3 size=maxCorrelation (data, index) // dynamic window size calculates (relatedness computation)
The training number of 4 trainData_window=trainData [index:index+size*48] // mono- window
According to
5 testData_window=trainData [index+size*48:index+ (size+7) * 48] // test number
According to
6 model.load (the before_model) // trained model of the upper window of load
7 model.fit (trainData_window, testData_window) // iteration more new model
9 index=index+7*48 // wherein 48 are the time slot quantity of each day
The encoding examples of the calculating of the dynamic sliding time window degree of correlation are as follows:
INPUT:
X2: // corresponding time series, it is integer that the time is processed, such as 2015050101,2015050102 ...
Y2: // corresponding serial number according to this, such as 1,2 ...
1 XMean=np.mean (X2)；YMean=np.mean (Y2) // formula calculates # mean value
2 XSD=np.std (X2)；YSD=np.std (Y2) // standard deviation
3 ZX=(X2-XMean)/XSD //z-score
4 ABS_ZX=(ZX)；ZY=(Y2-YMean)/YSD；ABS_ZY=(ZY) # related coefficient
// average degree of correlation, compares for the later period
5r=(np.sum (ABS_ZX)+np.sum (ABS_ZY))/(len (Y2)+len (X2) * 5)
Model evaluation module 40 is used to test assessment prediction model using root-mean-square error.
In the present embodiment, root-mean-square error is indicated as assessment models are as follows:
Wherein, n indicates the quantity of test data, wherein xiWithRespectively indicate the true value and corresponding emulation of test data
Value.In specific application, it can be estimated that repeatedly, be finally averaged.
For example, model evaluation module 40 selects 4 years data, has respectively corresponded 1-October 29 in July in 2013
Day, 1-June 27 March in 2014,1-June 27 March in 2015 and on 2 10th, 1 on the 1st November in 2015
Data as training data, on April 10th, 2 months in 2016 data as test data, the format of data is
(sample size, image channel, image length, picture traverse), unit time period are 30 minutes, input 6 × M × N, 2 × M × N, 2 × M
× N exports as 2 × M × N, and using RMSE assessment prediction model, the flow maximum in each area is 1293, the RMSE of prediction
It is 0.0298, being folded to true value is 17.30, and root-mean-square error RMSE is smaller, indicates that the precision of prediction model is higher.
Using RMSE as evaluation criteria, the comparison of algorithms of different model traffic flow forecasting is as shown in table 1:
The different model result contrast tables of table 1
Traffic forecast module 50 is used to export the predicted flow rate figure of the following predetermined time using prediction model, by predicted flow rate
Figure zooms to initial value, and does visable representation to predicted flow rate figure using thermodynamic chart.Wherein, predicted flow rate figure is restored by initial value
Algorithm obtains initial value, thus the result predicted.
It is that the Forecast of Urban Traffic Flow forecasting system of the embodiment of the present invention carries out the effect picture of short-term forecast refering to Fig. 5.This reality
It applies example to show prediction effect by the thermodynamic chart of echart, prediction result is the traffic of the April in 2016 of 00:30 on the 10th
The inlet flow and output stream situation of flow, there is specific value in each region.
It is that the Forecast of Urban Traffic Flow forecasting system of the embodiment of the present invention carries out the effect picture of long-term forecast refering to Fig. 6.This reality
It applies example to show prediction effect by the thermodynamic chart of echart, prediction result is on the April 10th, 2016 of 00:30 and 2016
On April 10,01:00 the magnitude of traffic flow inlet flow and output stream situation, there is specific value in each region.Its long-term forecast
Principle is first to carry out short-term forecast, is then put into prediction model for short-term forecast result as input data, is carried out again short-term
Prediction, thus achieve the effect that long-term forecast,
In the present embodiment, data preprocessing module 20 be also used to search missing tracing point to completion unit time period, and
Unit time period to the same time of predetermined number of days before completion unit time period is judged with the presence or absence of tracing point, there are tracing points
When, completion unit time period is treated by SES function according to the inlet flow of the unit time period of the same time of continuous predetermined number of days
Inlet flow is filled, and treats benefit by SES function according to the output stream of the unit time period of the same time of continuous predetermined number of days
The output stream of full unit time period is filled, and when tracing point is not present, by the whole to day where completion unit time period
Track point deletion.Predetermined number of days is, for example, 5-14 days, that is to say, that if first 5-14 days to the date where completion unit time period
Each unit time period there is tracing point, then just to this wait for completion unit time period inlet flow and output stream carry out completion,
Otherwise the tracing point of all unit time periods for waiting for the completion unit time period place date is rejected.
Further, SES function representation is as follows:
Ft+1=α xt+(1-α)Ft
Wherein, t is time current period, and α is coefficent of exponential smoothing, xtFor the actual observation value of t phase, FtFor the prediction of t phase
Value, Ft+1For the predicted value of t+1 phase.
It is the flow diagram of the Forecast of Urban Traffic Flow prediction technique of the embodiment of the present invention refering to Fig. 7.The city of the present embodiment
City's traffic flow forecasting method the following steps are included:
S1: it is handed over according to city in predetermined amount of time before cellphone GPS signal and/or vehicle GPS signal collection current date
Logical tracing point counts the inlet flow and output stream of each unit time period in each net region, and inlet flow and output stream is whole
Be combined into a tensor, wherein inlet flow be the unit period when starting not in net region but at the end of in net region
The sum of all tracing points, output stream be the unit period when starting in net region but at the end of it is all not in net region
The sum of tracing point, urban area are divided into multiple net regions, several unit time periods are one day；
S2: the characteristic dimension of inlet flow and output stream is zoomed to the range of prediction model requirement, and selects a date
As target date, the tensor for screening preceding 3 unit time periods of target date is made as the tensor of recent input data, the previous day
It is periodical input data, the tensor of the last week as trend input data, using the tensor of target date as test data；
S3: feature is extracted to recent input data by recent model and obtains exporting flow vector in the recent period, passes through the period
Model to periodical input data extract feature obtain the period output flow vector, by trend model to trend input data into
Row extracts feature and obtains trend output flow vector, and recent output stream vector, period output stream vector sum trend are exported flow vector
The characteristic dimension of the simulating traffic figure of target date is zoomed to prediction mould by the simulating traffic figure for integrate target date
The range that type requires, reversely adjusts prediction model using the simulating traffic figure of target date and the error of test data, wherein close
Phase model, periodic model, trend model are all made of identical depth residual error neural network framework；
S4: the prediction model is assessed using root-mean-square error test；
S5: being exported the predicted flow rate figure of the following predetermined time using prediction model, predicted flow rate figure zoomed to initial value, and
Visable representation is done to the predicted flow rate figure using thermodynamic chart.
The method of the present embodiment has technical characteristic identical with the system of previous embodiment, and details are not described herein.
The above description is only an embodiment of the present invention, is not intended to limit the scope of the invention, all to utilize this hair
Equivalent structure or equivalent flow shift made by bright specification and accompanying drawing content is applied directly or indirectly in other relevant skills
Art field, is included within the scope of the present invention.
Claims (9)
1. a kind of Forecast of Urban Traffic Flow forecasting system characterized by comprising
Data collection module, for according to the predetermined time before cellphone GPS signal and/or vehicle GPS signal collection current date
The tracing point of urban transportation, counts the inlet flow and output stream of each unit time period in each net region in section, and by inlet flow
Be integrated into a tensor with output stream, wherein inlet flow be the unit period when starting not in net region but at the end of in net
The sum of all tracing points in lattice region, output stream be the unit period when starting in net region but at the end of not in grid regions
The sum of all tracing points in domain, urban area are divided into multiple net regions, several unit time periods are one day；
Data preprocessing module, for the characteristic dimension of inlet flow and output stream to be zoomed to the range of prediction model requirement, and
A date is selected as target date, screens the tensors of preceding 3 unit time periods of target date as recent input data, preceding
One day tensor as periodical input data, the tensor of the last week as trend input data, using the tensor of target date as
Test data；
Model training module obtains exporting flow direction in the recent period for extracting feature to recent input data by recent model
Periodical input data are extracted feature by periodic model and obtain the period and export flow vector by amount, by trend model to becoming
Gesture input data extracts feature and obtains trend output flow vector, and recent output stream vector, period output stream vector sum are become
Gesture output flow vector integrate the simulating traffic figure of target date, by the characteristic dimension of the simulating traffic figure of target date
The range for zooming to prediction model requirement, reversely adjusted using the simulating traffic figure of target date and the error of test data described in
Prediction model, wherein the recent model, periodic model, trend model are all made of identical depth residual error neural network framework；
Model evaluation module, for assessing the prediction model using root-mean-square error test；
Traffic forecast module, for exporting the predicted flow rate figure of the following predetermined time using the prediction model, by predicted flow rate
Figure zooms to initial value, and does visable representation to the predicted flow rate figure using thermodynamic chart.
2. Forecast of Urban Traffic Flow forecasting system according to claim 1, which is characterized in that the data preprocessing module is also
For search missing tracing point to completion unit time period, and judge the same time to predetermined number of days before completion unit time period
Unit time period whether there is tracing point, when there are tracing point, according to the unit of the same time of the continuous predetermined number of days
The inlet flow of period is filled by the inlet flow that SES function treats completion unit time period, according to the continuous predetermined number of days
The output stream of unit time period of same time be filled by the output stream that SES function treats completion unit time period, and
When tracing point is not present, by whole track point deletions to day where completion unit time period.
3. Forecast of Urban Traffic Flow forecasting system according to claim 2, which is characterized in that the SES function representation is as follows:
Ft+1=α xt+(1-α)Ft
Wherein, t is time current period, and α is coefficent of exponential smoothing, xtFor the actual observation value of t phase, FtFor the predicted value of t phase,
Ft+1For the predicted value of t+1 phase.
4. Forecast of Urban Traffic Flow forecasting system according to claim 3, which is characterized in that the inlet flow indicates are as follows:
The output stream indicates are as follows:
Wherein, Tt is the track of line at the beginning and end of unit time period, and i indicates the length of net region, and j indicates net region
Width, P indicate time interval tthTrack set, gk-1Indicate tracing point when unit time period starts, gkIndicate that unit time period terminates
When tracing point, K indicate tracing point number；
The tensor representation are as follows:
X∈R2×I×J。
5. Forecast of Urban Traffic Flow forecasting system according to claim 4, which is characterized in that the data preprocessing module is adopted
It is as follows that feature scales algorithm:
X=x*2-1
Wherein, xmaxIndicate the maximum value in all data, xminIndicate the minimum value in all data；
Initial value retrieving algorithm corresponding with feature scaling algorithm is as follows:
X=x* (xmax-xmin)+xmin。
6. Forecast of Urban Traffic Flow forecasting system according to claim 5, which is characterized in that the depth residual error neural network
Framework successively includes 3*3 convolutional layer, 12 residual units, 3*3 convolutional layer, BN layers, relu activation primitive, 3*3 convolutional layer, residual error
Merge, activation primitive and convolutional layer.
7. Forecast of Urban Traffic Flow forecasting system according to claim 6, which is characterized in that the prediction model is using dynamic
Sliding time window is trained.
8. Forecast of Urban Traffic Flow forecasting system according to claim 7, which is characterized in that the root-mean-square error indicates
Are as follows:
Wherein, n indicates the quantity of test data, wherein xiWithRespectively indicate test data true value and corresponding simulation value.
9. a kind of Forecast of Urban Traffic Flow prediction technique characterized by comprising
S1: according to urban transportation in predetermined amount of time before cellphone GPS signal and/or vehicle GPS signal collection current date
Tracing point, counts the inlet flow and output stream of each unit time period in each net region, and inlet flow and output stream are integrated into
One tensor, wherein inlet flow be the unit period when starting not in net region but at the end of it is all in net region
The sum of tracing point, output stream be the unit period when starting in net region but at the end of all tracks not in net region
The sum of point, urban area is divided into multiple net regions, several unit time periods are one day；
The characteristic dimension of inlet flow and output stream: being zoomed to the range of prediction model requirement by S2, and select a date as
Target date, the tensor for screening preceding 3 unit time periods of target date are used as week as the tensor of recent input data, the previous day
Phase input data, the tensor of the last week are as trend input data, using the tensor of target date as test data；
S3: feature is extracted to recent input data by recent model and obtains exporting flow vector in the recent period, passes through periodic model
Feature is extracted to periodical input data and obtains period output flow vector, trend input data is mentioned by trend model
It takes feature to obtain trend output flow vector, recent output stream vector, period output stream vector sum trend output flow vector is carried out
The characteristic dimension of the simulating traffic figure of target date is zoomed to prediction model and wanted by the simulating traffic figure for integrating target date
The range asked reversely adjusts the prediction model using the simulating traffic figure of target date and the error of test data, wherein institute
It states recent model, periodic model, trend model and is all made of identical depth residual error neural network framework；
S4: the prediction model is assessed using root-mean-square error test；
S5: being exported the predicted flow rate figure of the following predetermined time using the prediction model, predicted flow rate figure zoomed to initial value, and
Visable representation is done to the predicted flow rate figure using thermodynamic chart.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910534104.3A CN110310479A (en) | 2019-06-20 | 2019-06-20 | A kind of Forecast of Urban Traffic Flow forecasting system and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910534104.3A CN110310479A (en) | 2019-06-20 | 2019-06-20 | A kind of Forecast of Urban Traffic Flow forecasting system and method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110310479A true CN110310479A (en) | 2019-10-08 |
Family
ID=68076149
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910534104.3A Pending CN110310479A (en) | 2019-06-20 | 2019-06-20 | A kind of Forecast of Urban Traffic Flow forecasting system and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110310479A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111127879A (en) * | 2019-12-09 | 2020-05-08 | 湖南大学 | Urban traffic flow prediction method based on generative countermeasure network |
| CN111639791A (en) * | 2020-05-11 | 2020-09-08 | 同济大学 | Traffic flow prediction method, system, storage medium and terminal |
| CN111709553A (en) * | 2020-05-18 | 2020-09-25 | 杭州电子科技大学 | Subway flow prediction method based on tensor GRU neural network |
| CN112833884A (en) * | 2020-12-31 | 2021-05-25 | 郑州轻工业大学 | Navigation method and navigation system based on big data |
| CN113192315A (en) * | 2020-01-14 | 2021-07-30 | 香港理工大学深圳研究院 | Traffic flow distribution prediction method, prediction device and terminal equipment |
| CN116029463A (en) * | 2023-03-28 | 2023-04-28 | 松立控股集团股份有限公司 | Traffic flow prediction and parking lot location method based on dynamic space-time correlation learning |
| CN116363878A (en) * | 2023-05-26 | 2023-06-30 | 云南大学 | Traffic flow prediction system and method based on continuous dynamic ordinary differential equation |
| CN117474365A (en) * | 2023-12-27 | 2024-01-30 | 西安衍舆航天科技有限公司 | Intelligent police service method and system based on artificial intelligence technology |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170061020A1 (en) * | 2006-09-19 | 2017-03-02 | Intuitive Control Systems, Llc | Collection, Monitoring, Analyzing and Reporting Decay Rate of Traffic Speed Data Via Vehicle Sensor Devices Placed at Multiple Remote Locations |
| CN107230351A (en) * | 2017-07-18 | 2017-10-03 | 福州大学 | A kind of Short-time Traffic Flow Forecasting Methods based on deep learning |
| CN109448361A (en) * | 2018-09-18 | 2019-03-08 | 云南大学 | Resident's traffic trip volume forecasting system and its prediction technique |
-
2019
- 2019-06-20 CN CN201910534104.3A patent/CN110310479A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170061020A1 (en) * | 2006-09-19 | 2017-03-02 | Intuitive Control Systems, Llc | Collection, Monitoring, Analyzing and Reporting Decay Rate of Traffic Speed Data Via Vehicle Sensor Devices Placed at Multiple Remote Locations |
| CN107230351A (en) * | 2017-07-18 | 2017-10-03 | 福州大学 | A kind of Short-time Traffic Flow Forecasting Methods based on deep learning |
| CN109448361A (en) * | 2018-09-18 | 2019-03-08 | 云南大学 | Resident's traffic trip volume forecasting system and its prediction technique |
Non-Patent Citations (1)
| Title |
|---|
| YUANFANG CHEN 等: "Poster: DeepTFP: Mobile Time Series Data Analytics based Traffic Flow Prediction", 《MOBILE COMPUTING AND NETWORKING》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111127879A (en) * | 2019-12-09 | 2020-05-08 | 湖南大学 | Urban traffic flow prediction method based on generative countermeasure network |
| CN111127879B (en) * | 2019-12-09 | 2021-09-07 | 湖南大学 | Urban traffic flow prediction method based on generative countermeasure network |
| CN113192315B (en) * | 2020-01-14 | 2022-08-23 | 香港理工大学深圳研究院 | Traffic flow distribution prediction method, prediction device and terminal equipment |
| CN113192315A (en) * | 2020-01-14 | 2021-07-30 | 香港理工大学深圳研究院 | Traffic flow distribution prediction method, prediction device and terminal equipment |
| CN111639791A (en) * | 2020-05-11 | 2020-09-08 | 同济大学 | Traffic flow prediction method, system, storage medium and terminal |
| CN111709553A (en) * | 2020-05-18 | 2020-09-25 | 杭州电子科技大学 | Subway flow prediction method based on tensor GRU neural network |
| CN111709553B (en) * | 2020-05-18 | 2023-05-23 | 杭州电子科技大学 | Subway flow prediction method based on tensor GRU neural network |
| CN112833884A (en) * | 2020-12-31 | 2021-05-25 | 郑州轻工业大学 | Navigation method and navigation system based on big data |
| CN112833884B (en) * | 2020-12-31 | 2021-09-28 | 郑州轻工业大学 | Navigation method and navigation system based on big data |
| CN116029463A (en) * | 2023-03-28 | 2023-04-28 | 松立控股集团股份有限公司 | Traffic flow prediction and parking lot location method based on dynamic space-time correlation learning |
| CN116363878A (en) * | 2023-05-26 | 2023-06-30 | 云南大学 | Traffic flow prediction system and method based on continuous dynamic ordinary differential equation |
| CN116363878B (en) * | 2023-05-26 | 2023-08-11 | 云南大学 | Traffic flow prediction system and method based on continuous dynamic ordinary differential equation |
| CN117474365A (en) * | 2023-12-27 | 2024-01-30 | 西安衍舆航天科技有限公司 | Intelligent police service method and system based on artificial intelligence technology |
| CN117474365B (en) * | 2023-12-27 | 2024-03-08 | 西安衍舆航天科技有限公司 | Intelligent police service method and system based on artificial intelligence technology |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110310479A (en) | A kind of Forecast of Urban Traffic Flow forecasting system and method | |
| CN110070713B (en) | Traffic flow prediction method based on bidirectional nested LSTM neural network | |
| CN111210633B (en) | Short-term traffic flow prediction method based on deep learning | |
| CN109508360B (en) | Geographical multivariate stream data space-time autocorrelation analysis method based on cellular automaton | |
| CN108197739B (en) | Urban rail transit passenger flow prediction method | |
| CN110164128A (en) | A kind of City-level intelligent transportation analogue system | |
| CN105513359B (en) | A kind of urban expressway traffic method for estimating state based on smart mobile phone movement detection | |
| CN111860951A (en) | Rail transit passenger flow prediction method based on dynamic hypergraph convolutional network | |
| CN110570651A (en) | Road network traffic situation prediction method and system based on deep learning | |
| CN109034449A (en) | Short-term bus passenger flow prediction technique based on deep learning and passenger behavior mode | |
| CN110458048A (en) | Take population distribution Spatio-temporal Evolution and the cognition of town pattern feature into account | |
| CN109902880A (en) | A kind of city stream of people's prediction technique generating confrontation network based on Seq2Seq | |
| CN110390349A (en) | Bus passenger flow volume based on XGBoost model predicts modeling method | |
| CN110222873A (en) | A kind of subway station passenger flow forecast method based on big data | |
| CN113672865A (en) | Traffic flow prediction method based on depth map Gaussian process | |
| CN110503104B (en) | Short-time remaining parking space quantity prediction method based on convolutional neural network | |
| CN112466117A (en) | Road network short-term traffic flow prediction method based on deep space-time residual error network | |
| CN111160622A (en) | Scenic spot passenger flow prediction method and device based on hybrid neural network model | |
| CN107704970A (en) | A kind of Demand-side load forecasting method based on Spark | |
| CN106815563A (en) | A kind of crowd's quantitative forecasting technique based on human body apparent structure | |
| CN118297775B (en) | Urban planning management and control system based on digital twin technology | |
| CN114202253A (en) | Charging station load adjustable potential evaluation method and system, storage medium and server | |
| CN107316096A (en) | A kind of track traffic one-ticket pass passenger amount of entering the station Forecasting Methodology | |
| CN113537626A (en) | Neural network combined time sequence prediction method for aggregating information difference | |
| CN112419711A (en) | Closed parking lot parking demand prediction method based on improved GMDH algorithm |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191008 |