CN109711640A - A kind of Short-time Traffic Flow Forecasting Methods based on fuzzy C-mean algorithm magnitude of traffic flow cluster and error feedback convolutional neural networks - Google Patents
A kind of Short-time Traffic Flow Forecasting Methods based on fuzzy C-mean algorithm magnitude of traffic flow cluster and error feedback convolutional neural networks Download PDFInfo
- Publication number
- CN109711640A CN109711640A CN201910064570.XA CN201910064570A CN109711640A CN 109711640 A CN109711640 A CN 109711640A CN 201910064570 A CN201910064570 A CN 201910064570A CN 109711640 A CN109711640 A CN 109711640A
- Authority
- CN
- China
- Prior art keywords
- traffic flow
- time
- data
- prediction
- flow
- 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
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Traffic Control Systems (AREA)
Abstract
It is a kind of based on the fuzzy C-mean algorithm magnitude of traffic flow cluster and error feedback convolutional neural networks Short-time Traffic Flow Forecasting Methods belong to traffic forecast field.The present invention carries out mode division to traffic flow using the thought of fuzzy clustering, compensates for the deficiency that traditional rigid clustering algorithm carries out mode division to the magnitude of traffic flow.Error feedback convolutional neural networks structure is proposed simultaneously, the disadvantages of space-time traffic information can not be made full use of by solving traditional neural network, and insensitive to Sudden Anomalies flow.And the two is combined into building built-up pattern, in the prediction of short-term traffic flow.This method improves the recognition capability to changes in flow rate caused by emergency event in the flow for allowing prediction model to predict future time unit more accurately.
Description
Technical field
The invention belongs to traffic forecast fields, more particularly to one kind is based on fuzzy C-mean algorithm magnitude of traffic flow cluster and error
Feed back the Short-time Traffic Flow Forecasting Methods of convolutional neural networks.
Background technique
As the rate of economic development was constantly accelerated in recent years, the quantity of private car is growing day by day, road traffic congestion and friendship
Interpreter thus wait traffic problems gradually at the global topic paid close attention to jointly.Since the eighties in last century, the government of many countries
Urban transportation planning of science activities is brought into schedule, intelligent transportation system (Intelligent Transportation System,
ITS) just developed gradually.Intelligent transportation system mainly utilizes advanced data communication technology and sensor technology,
Traffic data is integrated and is analyzed, to provide reasonable traffic guidance for urban highway traffic, road network is improved and passes through energy
Power reduces traffic accident.Herein wherein, accurately timely the prediction of short-term traffic flow not only can provide data for traffic programme
On support, reliable basis can also be provided for the road network development in future.Therefore short-time traffic flow forecast is for intelligent friendship
Way system has great significance.
Traffic forecast field achieved research achievement abundant in recent years, including based on linearly or nonlinearly system
The prediction technique of theory is based on Dynamic traffic assignment model prediction model and Artificial Neural Network Prediction Model.With new skill
The rise of art, more and more scholars start the relevant algorithm of machine learning being applied to solving road traffic problems.However
The prediction of many traffic flows remains in many aspects can be with improved place.Firstly, the data on flows between road is deposited
In certain relevance, this relevance is the reflection of road topology structure, if the pass between road can reasonably be excavated
Connection, then can improve the forecasting accuracy of model to a certain extent.Secondly, how to improve model to different caused by emergency event
The precision of prediction of normal flow, and the key point of research.
In order to further enhance the predictive ability of neural network, researchers are by itself and other intelligent methods or statistical method
It combines, constructs Comprehensive Model.Since these models often have higher precision of prediction relative to single model, because
This they become current mainstream research trend, and be gradually applied to forecasting traffic flow field.In addition, there are many more will be neural
The collective model that network technology is combined with other field advanced method, such as the combination of neural network and fuzzy logic, nerve
The combination of network and theory of heredity and flock of birds algorithm combines neural network with modified particle swarm optiziation, to wavelet neural
The Optimal improvements of network, BP neural network and the combination of Elman method etc..Since these models often have relative to single model
There is higher precision of prediction, therefore they become current research tendency, and is gradually applied to forecasting traffic flow field.
Simultaneously, convolutional neural networks image recognition, video detection and in terms of all obtained it is wide
General application.The design of convolutional neural networks is derived from the special nature of image data, we are generally acknowledged that in image that space joins
System is that local pixel connection is closer, and longer-distance pixel interdependence is weaker.And traffic flow data and image data
Very similar, the data on flows in some place is mainly the data influence by similar time point and upstream and downstream section.Pass through volume
A series of processing such as lamination and pond layer can preferably extract the space-time characterisation of traffic data, improve mould to a certain extent
The prediction levels of precision of type.
To sum up, the extensive use in view of built-up pattern in forecasting traffic flow and convolutional neural networks are in traffic flow data
The advantage for handling aspect, the invention proposes clustered using obscure idea to traffic flow and construct combined prediction mould on this basis
The Short-time Traffic Flow Forecasting Methods of type, herein wherein, in terms of the building of prediction model, the present invention is in order to solve Classical forecast mould
Type cannot using traffic flow space time information and the disadvantage insensitive to Sudden Anomalies traffic data, convolutional neural networks original
Have and carried out some improvement in structure, on the one hand model can be made to possess higher precision of prediction, aspect is but also model energy
More travel pattern variations are enough adapted to, the robustness of model is improved.
Summary of the invention
The contents of the present invention:
1. proposing a kind of friendship in short-term based on fuzzy C-mean algorithm magnitude of traffic flow cluster and error feedback convolutional neural networks
Through-flow prediction technique pre-processes data using the mode division of the magnitude of traffic flow, and prediction model building is combined to hand in short-term
Through-flow combination forecasting.
2. being improved traditional convolutional neural networks, it is added to error feedback layer, building error feedback convolution mind
Through network, and the prediction of application and short-term traffic flow.
This method is a kind of Short-time Traffic Flow Forecasting Methods based on built-up pattern, and traditional method directlys adopt original friendship
Logical data carry out the training of model, and noise big in view of the data volume of traffic data is more, and this method takes the magnitude of traffic flow fuzzy
The mode clustered in advance carries out mode division, the data difference for reducing training data concentration of high degree to the magnitude of traffic flow in advance
It is different, while each traffic data and each travel pattern relationship are remained again.
Designed error feeds back convolutional neural networks in this method, can receive the traffic of two dimensions of time and space
Stream information, at the same newly added error feedback layer can use before prediction error in set time step-length to prediction later
As a result it is adjusted, keeps model abnormal flow for caused by emergency event more sensitive, improve model to a certain extent
Precision of prediction and robustness.
To achieve the above object, it discusses and repeatedly practises after study, the present invention adopts the following technical scheme that:
Step 1. carries out mode division to the magnitude of traffic flow using fuzzy C-mean algorithm.The traffic data sample point of acquisition is expressed as X
={ x1,x2,...,xn, target clusters number c, maximum number of iterations T are set as 1000, and iteration threshold ε is set as 10-4, simultaneously
Initialize c cluster centre point P={ p1,p2,...,pcAnd subordinated-degree matrix U, subordinated-degree matrix U in element uijIt indicates
Degree of membership of j-th of sample point for i-th of flow rate mode.Cost function in iteration can indicate are as follows:Wherein dijIndicate that the Euclidean distance of sample and j-th of flow rate mode central point, m are
Cluster weighting parameters greater than 1.Specific step is as follows for step 1:
Step 1.1 randomly chooses c cluster centre P={ p1,p2,...,pc, while recording current iteration number t=0；
Step 1.2 is according to subordinating degree function iterative equationUpdate each member in subordinated-degree matrix U
Element, whereinIndicate the degree of membership updated value after the t times iteration, dkjIndicate sample point xjWith flow rate mode central point pk's
Euclidean distance.Updated subordinated-degree matrix can be expressed as Ut+1；
Step 1.3 is according to cluster centre iterative equationCluster centre is updated, wherein
Indicate the updated value that arrives of i-th kind of flow rate mode after the t times iteration,Indicate sample point xjWith i-th of flow rate mode in t
The secondary updated degree of membership of iteration.Updated cluster centre set can be expressed as Pt+1；
Step 1.4 calculates value formulaWherein Jt+1(Ut+1,Pt+1) indicate
The value of formula is worth after the t times iteration.
Step 1.5 judges whether current iteration number t is greater than or equal to maximum number of iterations T, or | | Ut+1-Ut| | <
ε, | | | | representing matrix norm, i.e. cost function updated value are less than preset update threshold value, terminate iteration if meeting, and enter
Step 2, t=t+1 is enabled if being unsatisfactory for, and is jumped to step 1.2 and continued iteration.
The space-time matrix of step 2. building traffic flow.Since convolution operation needs to extract the information of traffic flow two-dimensional matrix,
Therefore it first has to binding time and two, space dimension constructs the space-time matrix data of traffic flow, design pattern can indicate
Are as follows:
Longitudinal time series data for indicating l time step of traffic flow space-time matrix, it is lateral then indicate that n is a different
The Space expanding of monitoring point.Element x so in matrixn,t-lThen indicate that n-th of monitoring point is examined on the t-l time point
The data on flows measured.
Step 3. constructs error and feeds back convolutional neural networks.It mainly includes following three portions that error, which feeds back convolutional neural networks,
Point:
1) characteristic extraction part.Feature extraction network portion is close with existing convolutional neural networks, passes through convolutional layer, pond
Change layer, developer layer and connect layer entirely, connects the information of a dimension to extract traffic flow time and space.This part input data format
For the magnitude of traffic flow two-dimensional matrix in step 2, output result is a feature vector.
2) error feedback fraction.The main main function of error feedback network portions is to receive chronomere between model
Prediction error value, to identify that abnormal flow makes adjustment in time to final prediction result.The received data of error feedback layer point
For two parts:
A) result that feature extraction network portion is predicted.Assuming that the input data of this part is characterized vector v, then miss
The output of this part of poor feedback layer can then indicate are as follows:
pC=δ (wCv+bC)
Wherein wCFor connection weight, bCFor output biasing, δ () is excitation function, uses Relu function in this method.
B) before existing model l time step prediction error data: i.e. the input data that receives of this part can be with
It is expressed as the form of error vector, the length l of vector, the error that l time step is predicted before each element represents.Prediction misses
Difference is the vector as composed by the difference of predicted value and true value in l step-length before, can be indicated are as follows:
et=[y (t-1)-o (t-1) ..., y (t-l)-o (t-l)]
The truthful data and prediction data of l time step before wherein y (t-l) and o (t-l) are illustrated respectively in.With mistake
Last point of processing mode of poor feedback layer is identical, can indicate in the output of t moment, this part are as follows:
WhereinFor connection weight, bEFor output biasing, δ () is excitation function, uses Relu function in this method.
3) output fusion part.This part receives two parts output of error feedback layer, i.e. p respectivelycAnd pE, then by two
Part is merged, using result as final output.Calculating process can indicate are as follows:
O=f (wOPpC+wOEpE+bO)
Wherein wOP,wOEAnd bOConnection weight and biasing for output layer neuron, f () are the excitation function of output layer,
Use ReLU function as the excitation function of output layer, o is the final prediction result of model.
Step 4. defines loss function, training pattern.Loss function is defined as follows:
Wherein n is number of samples, and o is the predicted value of prediction model, and y is real traffic data.According to the above loss function,
Optimal solution is sought to model parameter by back-propagation algorithm.
The prediction of step 5. realization short-term traffic flow.The real-time traffic flow data of acquisition will make the traffic flow of t moment
Prediction, then prediction steps are as follows:
Step 5.1 constructs traffic flow space-time matrix as data according to step 2 and is input to combined error feedback forecasting model
In, available prediction output valve
Step 5.2 is searched subordinated-degree matrix according to predicted time point t, in the subordinated-degree matrix that obtains in step 1 and is found
The membership vector u of corresponding time point sample and each magnitude of traffic flow modet={ ut1,ut2,...,utc, wherein utcIndicate t
The degree of membership of flow rate mode in the traffic flow data and c at moment.
Step 5.3 is by step 5.2 membership vector utAs the prediction output valve O in 5.1tWeighted value, calculate weighting
With as final prediction result, calculating process be may be expressed as:
Detailed description of the invention
Fig. 1 error feeds back convolutional neural networks structure；
Fig. 2 error feedback layer internal structure chart；
Fig. 3 is based on error and feeds back convolutional neural networks combination forecasting overall structure；
Specific embodiment
1. the fuzzy clustering of the magnitude of traffic flow
In the magnitude of traffic flow mode division stage, it is necessary first to a mean daily flow data are obtained, thus in flow dimension
It is clustered, re-maps and obtain corresponding flow rate mode on time dimension.The method that the present invention takes is comprehensive daily first
In each moment road total flow, then each moment is carried out again to be averaged operation.D days flows are shared in data set
Data can collect the magnitude of traffic flow at T moment daily, then will on d days t moment roads all monitoring point flow set expressions
For Nt d, then the average lane flow of t moment can indicate are as follows:
The data set divided for traffic flow pattern can be expressed as
Fuzzy C-Means Cluster Algorithm is a kind of clustering algorithm based on division, and thought is constantly to be obtained by objective function
To each sample point to the degree of membership at all class clusters center, thus determine the generic of sample point with reach automatically to sample data into
The purpose of row classification.Traditional clustering algorithm is rigid for the division of data, and Fuzzy C-Means Cluster Algorithm is being gathered
The association between data can also be retained while class.Sample in Fuzzy C-Means Cluster Algorithm iterative process is for each point
The cost function of class cluster can indicate are as follows:
Wherein pi(i=1,2 ..., c) indicate ith cluster center, uijFor j-th of sampleTo i-th of cluster of classifying
Degree of membership, and degree of membership uijMeet:
M is the weighting parameters greater than 1, dijTo be a kind of apart from norm, using Euclidean distance, as shown in formula (4),
Middle xjThat indicate is sample vector, pjIndicate class cluster center vector.
Algorithm initially initializes subordinated-degree matrix U in requisition for the random number used between 0 to 1, and user inputs mesh
Mark cluster numbers c, iteration threshold ε, maximum number of iterations T and weighting parameters m.Cluster centre p in iterative processiPass through formula
(5) it calculates, wherein t is just in the number of iteration, while the degree of membership U of element and each cluster uses formula (6) to be updated, repeatedly
Each step in generation is calculated according to cluster cost function, if the updated value result of cost function is less than iteration set by user
Threshold epsilon or the number of iterations reach default iteration upper limit T, then can stop iteration, export cluster centre and subordinated-degree matrix
As cluster result.
Based on above description, Fuzzy C-Means Cluster Algorithm can be summarized as following steps:
By way of the above fuzzy clustering, the magnitude of traffic flow of different moments can be clustered, obtain each moment
With the membership of flow rate mode, supported so that the Combined model forecast model for after provides data.
2. the building of the traffic flow matrix based on space-time characteristic
The magnitude of traffic flow has space-time characterisation: on time dimension, an observation point can be handed in a period of time persistent collection
Through-current capacity data obtain the sequence that a magnitude of traffic flow changes over time；It, can be according to road distribution situation on Spatial Dimension
Multiple monitoring points are set in different roads or crossing, to obtain the monitoring number similar to upstream section and downstream road section
According to.Therefore original traffic flow data can be formatted as the traffic flow matrix with time-space attribute on time and Spatial Dimension
The tectonic style of magnitude of traffic flow space-time matrix is as follows:
Longitudinal time series data for indicating l time step of traffic flow space-time matrix, it is lateral then indicate that n is a different
The Space expanding of monitoring point.Element x so in matrixn,t-lThen indicate that n-th of monitoring point is examined on the t-l time point
The data on flows measured.
3. the building of error feedback convolutional neural networks prediction model
Convolutional neural networks model structure is fed back for error as shown in Figure 1, is divided into 3 parts, feature extraction unit in model
Divide, error feedback fraction and output merge part.Characteristic extraction part.
3.1 feature extraction network portions can extract the information of two dimensions of traffic flow time and space, with existing volume
Product neural network is close, wherein comprising with flowering structure:
3.1.1 convolutional layer: identical as the convolution operation in existing convolutional neural networks, convolutional layer uses the side of two-dimensional convolution
Formula is calculated, and the process of convolution not only can handle time series represented by row vector, while can also be in column vector
Space characteristics extract.The calculation expression of convolutional layer is as follows:
WhereinIndicate n-th of magnitude of traffic flow eigenmatrix in l layers of convolutional layer,For convolution kernel function,For biasing
Parameter, f () indicate the activation primitive of neuron, use activation primitive of the Relu function as convolutional layer in this method.
3.1.2 pond layer:, the input data that pond layer receives identical as the pond operation in existing convolutional neural networks
For the output dimensional matrix data of convolutional layer, input data is compressed in the case where not influencing feature.It is adopted in this method
With average pond method.
3.1.3 it developer layer: by the operation of convolution sum pond it is found that two-part output data is the form of two-dimensional matrix, cares for
The format for being converted into one-dimensional vector is output it using developer layer, expression formula is as follows:
F=Flat (Y, size) (9)
Wherein Flat is expanded function, and Y is input data, and size is transformation result dimension, such as Flat (Y, (1*5)) is then
It is that input data Y is converted to an one-dimensional vector with 5 attributes.
3.2 error feedback fractions.As shown in Fig. 2, the main main function of error feedback network portions be receive model it
Between chronomere prediction error value, to identify that abnormal flow makes adjustment in time to final prediction result.Error feedback layer
Received data are divided into two parts:
3.2.1 the result that feature extraction network portion is predicted.Assuming that the input data of this part is characterized vector v,
Then the output of this part of error feedback layer can then indicate are as follows:
pC=δ (wCv+bC) (10)
Wherein wCFor connection weight, bCFor output biasing, δ () is excitation function, uses Relu function in this method.
3.2.2 before existing model l time step prediction error data: the input data that i.e. this part receives can
In the form of being expressed as error vector, the length l of vector, the error that l time step is predicted before each element represents.Prediction
Error is the vector as composed by the difference of predicted value and true value in l step-length before, can be indicated are as follows:
et=[y (t-1)-o (t-1) ..., y (t-l)-o (t-l)] (11)
The truthful data and prediction data of l time step before wherein y (t-l) and o (t-l) are illustrated respectively in.With mistake
Last point of processing mode of poor feedback layer is identical, can indicate in the output of t moment, this part are as follows:
WhereinFor connection weight, bEFor output biasing, δ () is excitation function, uses Relu function in this method.
3.3 output fusion parts.This part receives two parts output of error feedback layer, i.e. p respectivelycAnd pE, then by two
Part is merged, using result as final output.Calculating process can indicate are as follows:
O=f (wOPpC+wOEpE+bO) (12)
Wherein wOP,wOEAnd bOConnection weight and biasing for output layer neuron, f () are the excitation function of output layer,
Use ReLU function as the excitation function of output layer, o is the final prediction result of model.
4. example: short-time traffic flow forecast
4.1 choose experimental data
Original traffic experiment number contains the flow detection number for the test point 22 days of fastlink 9 that a length is 28km
According to the time recorded daily is 0:00 to 21:35, and intra-record slack byte is 5 minutes.Preceding 20 days traffic flow datas are chosen as training
Collection, rear 2 days data are as predictive data set.
4.2 parameters determine
The stage is divided in traffic flow flow rate mode, it is 3 that main parameter, which is set as target cluster numbers, i.e., by the friendship in one day
Through-current capacity is divided into high, medium and low Three models, iteration threshold 10-4, maximum number of iterations 1000.
Error is fed back in the building process of convolutional neural networks, and convolution layer number is set as 2 layers, and every layer of convolution nuclear volume is
5, size 3*3, the quantity of the pond number of plies are 1, and pond window size is 2*2.Feedback error step is received in error feedback layer
It is long to be set as 5, that is, receive the prediction error of preceding 5 chronomere's models.
4.3 prediction result
Based on the resulting optimized parameter of training, the available experimental result for test data set after model construction.
This method uses three kinds of indexs and carrys out the performance of assessment prediction model, respectively mean absolute error (MAE), average
Absolute percent error (MAPE) and root-mean-square error (RMSE), calculation formula can indicate are as follows:
WhereinIndicate prediction result, θ (i) indicates that the actual flow monitored, N indicate the total quantity of data.
It is as shown in table 1 for the prediction statistical result in nine sections:
Claims (5)
1. a kind of short-time traffic flow forecast side based on fuzzy C-mean algorithm magnitude of traffic flow cluster and error feedback convolutional neural networks
Method, which comprises the following steps:
Step 1. uses FCM Algorithms, and the magnitude of traffic flow in one day is divided into the different flow rate mode of C kind；
Step 2. constructs traffic flow two-dimensional matrix in time dimension and Spatial Dimension according to the space-time characterisation of traffic data；
Step 3. constructs error feedback convolutional neural networks in the structure basis of traditional convolutional neural networks；
Step 4. defines loss function, training pattern；Define loss functionWherein n is sample
Number, o are the predicted value of prediction model, and y is real traffic data；According to loss function, model is joined by back-propagation algorithm
Number seeks optimal solution；
Step 5. feeds back volume and auditing the network combination forecasting according to error, realizes the prediction of short-term traffic flow.
2. for the friendship in short-term based on fuzzy C-mean algorithm magnitude of traffic flow cluster and convolutional neural networks described in claim 1
Through-flow prediction technique, which is characterized in that in step 1, X={ x is expressed as the traffic data sample point of acquisition1,x2,...,
xn, target clusters number c, maximum number of iterations T are set as 1000, and iteration threshold ε is set as 10-4, while initializing c and gathering
Class central point P={ p1,p2,...,pcAnd subordinated-degree matrix U, subordinated-degree matrix U in element uijIndicate j-th of sample point
For the degree of membership of i-th of flow rate mode；Cost function in iteration indicates are as follows:
Wherein dijIndicate that the Euclidean distance of sample and j-th of flow rate mode central point, m are the cluster weighting parameters greater than 1, the tool of cluster
Steps are as follows for body:
Step 1.1 randomly chooses c cluster centre P={ p1,p2,...,pc, while recording current iteration number t=0；
Step 1.2 is according to subordinating degree function iterative equationEach element in subordinated-degree matrix U is updated,
InIndicate the degree of membership updated value after the t times iteration, dkjIndicate sample point xjWith flow rate mode central point pkIt is European away from
From；Updated subordinated-degree matrix is expressed as Ut+1；
Step 1.3 is according to cluster centre iterative equationCluster centre is updated, whereinIt indicates
The updated value that arrives of i-th kind of flow rate mode after the t times iteration,Indicate sample point xjIt changes with i-th of flow rate mode at the t times
For updated degree of membership；Updated cluster centre set expression is Pt+1；
Step 1.4 calculates value formulaWherein Jt+1(Ut+1,Pt+1) indicate at the t times
The value of formula is worth after iteration；
Step 1.5 judges whether current iteration number t is greater than or equal to maximum number of iterations T, or | | Ut+1-Ut| | < ε, | |
| | representing matrix norm, i.e. cost function updated value are less than preset update threshold value, terminate iteration if meeting, enter step 2,
T=t+1 is enabled if being unsatisfactory for, and is jumped to step 1.2 and continued iteration.
3. for the friendship in short-term based on fuzzy C-mean algorithm magnitude of traffic flow cluster and convolutional neural networks described in claim 1
Through-flow prediction technique, which is characterized in that in step 2, in order to which convolution operation extracts the information of traffic flow two-dimensional matrix, first have to tie
It closes the time and two, space dimension constructs the space-time matrix data of traffic flow, design pattern indicates are as follows:
Longitudinal time series data for indicating l time step of traffic flow space-time matrix, it is lateral then indicate a different monitorings of n
The Space expanding of point；Element x so in matrixn,t-lThen indicate that n-th of monitoring point detects on the t-l time point
Data on flows.
4. for the friendship in short-term based on fuzzy C-mean algorithm magnitude of traffic flow cluster and convolutional neural networks described in claim 1
Through-flow prediction technique, which is characterized in that the error feedback convolutional neural networks of step 3 include following three parts:
3.1 characteristic extraction part；Feature extraction network portion is close with existing convolutional neural networks, passes through convolutional layer, Chi Hua
Layer and connects layer at developer layer entirely, to extract the information of traffic flow time and space even dimension；This part input data format is
Magnitude of traffic flow two-dimensional matrix in step 2, output result are a feature vector；
3.2 error feedback fractions；The main main function of error feedback network portions is the pre- of chronomere between reception model
Error amount is surveyed, to identify that abnormal flow makes adjustment in time to final prediction result；The received data of error feedback layer are divided into
Two parts:
3.2.1 the result that feature extraction network portion is predicted；Assuming that the input data of this part is characterized vector v, then miss
The output of this part of poor feedback layer then indicates are as follows: pC=δ (wCv+bC), wherein wPFor connection weight, bpFor output biasing, δ
() is excitation function, and excitation function uses Relu function；
3.2.2 before existing model l time step prediction error data；The input data that this part receives is expressed as missing
The form of difference vector, the length l of vector, the error that l time step is predicted before each element represents；Predict that error is by it
Predicted value in preceding l step-length and vector composed by the difference of true value indicate are as follows:
et=[y (t-1)-o (t-1) ..., y (t-l)-o (t-l)], l before wherein y (t-l) and o (t-l) are illustrated respectively in
The truthful data and prediction data of time step；It is identical as last point of processing mode of error feedback layer, in t moment, this portion
The output divided indicates are as follows:WhereinFor connection weight, bEFor output biasing, δ () is excitation
Function, excitation function Relu function；
3.3 output fusion parts；This part receives two parts output of error feedback layer, i.e. p respectivelycAnd pE, then by two parts
It is merged, using result as final output；Calculating process indicates are as follows: o=f (wOPpC+wOEpE+bO), wherein wOP,wOEAnd bO
Connection weight and biasing for output layer neuron, f () are the excitation function of output layer, use ReLU function as output layer
Excitation function, o is the final prediction result of model.
5. for the friendship in short-term based on fuzzy C-mean algorithm magnitude of traffic flow cluster and convolutional neural networks described in claim 1
Through-flow prediction technique, which is characterized in that in step 5, the prediction steps made prediction to the traffic flow of t moment are as follows:
Step 5.1 constructs traffic flow space-time matrix as data according to step 2 and is input in combined error feedback forecasting model, obtains
To prediction output valve
Step 5.2 searches subordinated-degree matrix according to predicted time point t, in the subordinated-degree matrix that obtains in step 1 and finds correspondence
The membership vector u of time point sample and each magnitude of traffic flow modet={ ut1,ut2,...,utc, wherein utcIndicate t moment
Traffic flow data and c in flow rate mode degree of membership；
Step 5.3 is by step 5.2 membership vector utAs the prediction output valve O in 5.1tWeighted value, calculate weighted sum make
For final prediction result, calculating process is indicated are as follows:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910064570.XA CN109711640A (en) | 2019-01-23 | 2019-01-23 | A kind of Short-time Traffic Flow Forecasting Methods based on fuzzy C-mean algorithm magnitude of traffic flow cluster and error feedback convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910064570.XA CN109711640A (en) | 2019-01-23 | 2019-01-23 | A kind of Short-time Traffic Flow Forecasting Methods based on fuzzy C-mean algorithm magnitude of traffic flow cluster and error feedback convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109711640A true CN109711640A (en) | 2019-05-03 |
Family
ID=66262792
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910064570.XA Pending CN109711640A (en) | 2019-01-23 | 2019-01-23 | A kind of Short-time Traffic Flow Forecasting Methods based on fuzzy C-mean algorithm magnitude of traffic flow cluster and error feedback convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109711640A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110139299A (en) * | 2019-05-14 | 2019-08-16 | 鹰潭泰尔物联网研究中心 | The clustering method of base station flow in a kind of cellular network |
| CN110766066A (en) * | 2019-10-18 | 2020-02-07 | 天津理工大学 | FNN-based tensor heterogeneous integrated internet of vehicles missing data estimation method |
| CN110907155A (en) * | 2019-12-02 | 2020-03-24 | 吉林松江河水力发电有限责任公司 | Fault monitoring method for rotating shaft of water turbine |
| CN110991775A (en) * | 2020-03-02 | 2020-04-10 | 北京全路通信信号研究设计院集团有限公司 | Deep learning-based rail transit passenger flow demand prediction method and device |
| CN111242268A (en) * | 2019-09-05 | 2020-06-05 | 中国科学院计算技术研究所 | Method for searching convolutional neural network |
| CN111882114A (en) * | 2020-07-01 | 2020-11-03 | 长安大学 | Short-term traffic flow prediction model construction method and prediction method |
| CN112085947A (en) * | 2020-07-31 | 2020-12-15 | 浙江工业大学 | Traffic jam prediction method based on deep learning and fuzzy clustering |
| CN112561146A (en) * | 2020-12-08 | 2021-03-26 | 哈尔滨工程大学 | Large-scale real-time traffic flow prediction method based on fuzzy logic and depth LSTM |
| CN113011397A (en) * | 2021-04-27 | 2021-06-22 | 北京工商大学 | Multi-factor cyanobacterial bloom prediction method based on remote sensing image 4D-FractalNet |
| CN115953186A (en) * | 2023-02-24 | 2023-04-11 | 北京化工大学 | Network appointment demand pattern recognition and short-time demand prediction method |
| CN116599779A (en) * | 2023-07-19 | 2023-08-15 | 中国电信股份有限公司江西分公司 | IPv6 cloud conversion method for improving network security performance |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102693633A (en) * | 2012-06-07 | 2012-09-26 | 浙江大学 | Short-term traffic flow weighted combination prediction method |
| 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 |
| WO2016095708A1 (en) * | 2014-12-16 | 2016-06-23 | 高德软件有限公司 | Traffic flow prediction method, and prediction model generation method and device |
| US20160314686A1 (en) * | 2013-12-30 | 2016-10-27 | Fudan University | Method for traffic flow prediction based on spatio-temporal correlation mining |
| 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 |
| CN108647834A (en) * | 2018-05-24 | 2018-10-12 | 浙江工业大学 | A kind of traffic flow forecasting method based on convolutional neural networks structure |
-
2019
- 2019-01-23 CN CN201910064570.XA patent/CN109711640A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102693633A (en) * | 2012-06-07 | 2012-09-26 | 浙江大学 | Short-term traffic flow weighted combination prediction method |
| US20160314686A1 (en) * | 2013-12-30 | 2016-10-27 | Fudan University | Method for traffic flow prediction based on spatio-temporal correlation mining |
| WO2016095708A1 (en) * | 2014-12-16 | 2016-06-23 | 高德软件有限公司 | Traffic flow prediction method, and prediction model generation method and device |
| 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 |
| 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 |
| CN108647834A (en) * | 2018-05-24 | 2018-10-12 | 浙江工业大学 | A kind of traffic flow forecasting method based on convolutional neural networks structure |
Non-Patent Citations (1)
| Title |
|---|
| 杨世坚等: "基于模糊C均值聚类和神经网络的短时交通流预测方法", 《系统工程》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110139299A (en) * | 2019-05-14 | 2019-08-16 | 鹰潭泰尔物联网研究中心 | The clustering method of base station flow in a kind of cellular network |
| CN111242268A (en) * | 2019-09-05 | 2020-06-05 | 中国科学院计算技术研究所 | Method for searching convolutional neural network |
| CN110766066B (en) * | 2019-10-18 | 2023-06-23 | 天津理工大学 | Tensor heterogeneous integrated vehicle networking missing data estimation method based on FNN |
| CN110766066A (en) * | 2019-10-18 | 2020-02-07 | 天津理工大学 | FNN-based tensor heterogeneous integrated internet of vehicles missing data estimation method |
| CN110907155A (en) * | 2019-12-02 | 2020-03-24 | 吉林松江河水力发电有限责任公司 | Fault monitoring method for rotating shaft of water turbine |
| CN110991775A (en) * | 2020-03-02 | 2020-04-10 | 北京全路通信信号研究设计院集团有限公司 | Deep learning-based rail transit passenger flow demand prediction method and device |
| CN111882114A (en) * | 2020-07-01 | 2020-11-03 | 长安大学 | Short-term traffic flow prediction model construction method and prediction method |
| CN111882114B (en) * | 2020-07-01 | 2023-10-31 | 长安大学 | Short-time traffic flow prediction model construction method and prediction method |
| CN112085947A (en) * | 2020-07-31 | 2020-12-15 | 浙江工业大学 | Traffic jam prediction method based on deep learning and fuzzy clustering |
| CN112085947B (en) * | 2020-07-31 | 2023-10-24 | 浙江工业大学 | Traffic jam prediction method based on deep learning and fuzzy clustering |
| CN112561146B (en) * | 2020-12-08 | 2023-04-18 | 哈尔滨工程大学 | Large-scale real-time traffic flow prediction method based on fuzzy logic and depth LSTM |
| CN112561146A (en) * | 2020-12-08 | 2021-03-26 | 哈尔滨工程大学 | Large-scale real-time traffic flow prediction method based on fuzzy logic and depth LSTM |
| CN113011397A (en) * | 2021-04-27 | 2021-06-22 | 北京工商大学 | Multi-factor cyanobacterial bloom prediction method based on remote sensing image 4D-FractalNet |
| CN113011397B (en) * | 2021-04-27 | 2024-03-29 | 北京工商大学 | Multi-factor cyanobacterial bloom prediction method based on remote sensing image 4D-Fractalnet |
| CN115953186A (en) * | 2023-02-24 | 2023-04-11 | 北京化工大学 | Network appointment demand pattern recognition and short-time demand prediction method |
| CN115953186B (en) * | 2023-02-24 | 2023-05-16 | 北京化工大学 | Network appointment vehicle demand pattern recognition and short-time demand prediction method |
| CN116599779A (en) * | 2023-07-19 | 2023-08-15 | 中国电信股份有限公司江西分公司 | IPv6 cloud conversion method for improving network security performance |
| CN116599779B (en) * | 2023-07-19 | 2023-10-27 | 中国电信股份有限公司江西分公司 | IPv6 cloud conversion method for improving network security performance |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109711640A (en) | A kind of Short-time Traffic Flow Forecasting Methods based on fuzzy C-mean algorithm magnitude of traffic flow cluster and error feedback convolutional neural networks | |
| Hong et al. | HetETA: Heterogeneous information network embedding for estimating time of arrival | |
| CN106251625B (en) | Three-dimensional urban road network global state prediction technique under big data environment | |
| Ren et al. | TBSM: A traffic burst-sensitive model for short-term prediction under special events | |
| Liu et al. | Data fusion for multi-source sensors using GA-PSO-BP neural network | |
| CN110070713A (en) | A kind of traffic flow forecasting method based on two-way nested-grid ocean LSTM neural network | |
| CN103871246B (en) | Based on the Short-time Traffic Flow Forecasting Methods of road network spatial relation constraint Lasso | |
| 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 | |
| CN111860951A (en) | Rail transit passenger flow prediction method based on dynamic hypergraph convolutional network | |
| Yan et al. | Spatial-temporal chebyshev graph neural network for traffic flow prediction in iot-based its | |
| Akgüngör et al. | An artificial intelligent approach to traffic accident estimation: Model development and application | |
| CN110390349A (en) | Bus passenger flow volume based on XGBoost model predicts modeling method | |
| CN110503104B (en) | Short-time remaining parking space quantity prediction method based on convolutional neural network | |
| Li et al. | Evaluation of urban green space landscape planning scheme based on PSO-BP neural network model | |
| CN109508360A (en) | A kind of polynary flow data space-time autocorrelation analysis method of geography based on cellular automata | |
| CN110288202A (en) | A kind of Urban Park Green Space frastructure state Evaluation and Optimization | |
| CN109785618A (en) | Short-term traffic flow prediction method based on combinational logic | |
| CN106910199A (en) | Towards the car networking mass-rent method of city space information gathering | |
| CN111523706B (en) | Section lane-level short-term traffic flow prediction method based on deep learning combination model | |
| CN109191849A (en) | A kind of traffic congestion Duration Prediction method based on multi-source data feature extraction | |
| CN107862877A (en) | A kind of urban traffic signal fuzzy control method | |
| CN106384507A (en) | Travel time real-time estimation method based on sparse detector | |
| CN115204477A (en) | Bicycle flow prediction method of context awareness graph recursive network | |
| CN110287995B (en) | Multi-feature learning network model method for grading all-day overhead traffic jam conditions |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190503 |
|
| WD01 | Invention patent application deemed withdrawn after publication |