CN112884222B - Time-period-oriented LSTM traffic flow density prediction method - Google Patents

Time-period-oriented LSTM traffic flow density prediction method Download PDF

Info

Publication number
CN112884222B
CN112884222B CN202110185310.5A CN202110185310A CN112884222B CN 112884222 B CN112884222 B CN 112884222B CN 202110185310 A CN202110185310 A CN 202110185310A CN 112884222 B CN112884222 B CN 112884222B
Authority
CN
China
Prior art keywords
traffic flow
flow density
time
data
neural network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110185310.5A
Other languages
Chinese (zh)
Other versions
CN112884222A (en
Inventor
曾昀敏
陈佩仪
张衡
胡栋
孙艺菲
姚剑
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN202110185310.5A priority Critical patent/CN112884222B/en
Publication of CN112884222A publication Critical patent/CN112884222A/en
Application granted granted Critical
Publication of CN112884222B publication Critical patent/CN112884222B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/01—Detecting movement of traffic to be counted or controlled
    • G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions

Abstract

The invention discloses a time-slot-oriented LSTM traffic flow density prediction method. The invention creatively provides a concept of traffic flow density, obtains a data set of the traffic flow density of each vehicle in the area, and effectively reflects the actual traffic jam condition in the urban traffic hot spot area. And local extension or shortening is carried out on a time axis by utilizing a dynamic time warping method, so that the similarity among the traffic flow density sequence slices is more accurately measured. And classifying the traffic flow density slice sequences according to the similarity by using a K-Means clustering method, so as to be beneficial to reflecting the typical mode of the traffic flow density. And respectively training neural networks aiming at traffic flow density data with similar patterns, and performing weighted summation on the output predicted values of the neural networks according to the time labels or the distances of the data clustering centers, thereby fully utilizing the characteristics of the sequence structure of the data. And finally, performing similar day correction on the weighted summation result by using the past data with the same type of the working day, so as to improve the algorithm prediction precision.

Description

Time-period-oriented LSTM traffic flow density prediction method
Technical Field
The invention belongs to the field of traffic prediction, and particularly relates to a time-slot-oriented LSTM traffic flow density prediction method.
Background
With the promotion of the society and economy development of China and the steady progress of urbanization construction, the physical and living conditions of people are improved, the holding capacity of motor vehicles and the road traffic volume are increased rapidly, and the traffic jam phenomenon of cities, particularly big cities, becomes a focus problem at present. The problems of traffic network retardation, disordered traffic order and the like are gradually highlighted, and huge impact and pressure are brought to urban traffic management. The problems of commuting time waste, energy and economic loss, air quality reduction and the like caused by the phenomenon are increasingly highlighted. At present, the problem of urban road traffic congestion not only restricts the further development of social economy, but also is an important reason influencing the trip experience of people and reducing the quality of life of people. Therefore, research on urban road vehicle motion prediction methods must be strengthened to further complete the task of solving urban traffic congestion.
In view of the problem of predicting the traffic flow of urban roads, experts and scholars in the related field have conducted many studies, and therefore prediction methods with higher complexity and better performance are continuously proposed, or the existing prediction methods are improved from different angles. Generally, these prediction methods can be broadly divided into two categories: the method is supported by classical mathematical and physical theory knowledge and is suitable for processing small-scale, simple and single traffic flow data, but the method cannot meet the requirement of actual precision along with the complication of a traffic system; the other type of prediction model is formed by taking modern scientific technology and methods as main research means, and comprises a nonparametric regression model, a multidimensional fractal-based method, a spectrum analysis method, a neural network model and the like, wherein the methods pay attention to improving the fitting effect on the real traffic flow phenomenon, do not pursue strict mathematical derivation and clear physical significance, and show advantages in processing more complex traffic flow data. A Long-Short Term Memory neural network (Long-Short Term Memory) is one of the commonly used neural network models, and has a relatively good performance in traffic flow prediction in recent years. However, since the simulation is simply performed according to the time correlation of the traffic flow data and the objective law of actual road traffic is not considered, there is still room for improvement in terms of improving the accuracy.
Disclosure of Invention
The invention aims to provide a time period-oriented LSTM (long short term memory neural network) traffic flow density prediction method by combining four modules of clustering, neural network, ensemble learning and similar day correction, because the current deep learning method for predicting a time sequence aiming at a certain time period or time label is very limited and a clustering idea is rarely introduced into a prediction process. The traffic flow density is defined by the number of track points left by the vehicle in unit time, and the traffic jam condition can be reflected better.
The technical scheme of the invention is a time slot-oriented LSTM traffic flow density prediction method, which comprises a data processing stage, a training stage, a parameter adjusting stage and a testing stage;
the data processing stage specifically comprises: selecting a research area, selecting a part which has intersection with the research area from the acquired vehicle track information, calculating according to a traffic flow density formula to obtain a traffic flow density data set of each vehicle in the area, and dividing the traffic flow density data set into a training set, a verification set and a test set;
the training phase comprises the following steps:
s11, segmenting traffic flow density data in a training set according to a given length to obtain time sequence slices, recording time labels of the time sequence slices, calculating similarity among the time sequence slices by using a dynamic time warping method, and classifying the traffic flow density slices according to the similarity by using a K-Means clustering method to obtain N clustering clusters and corresponding time labels;
Step S12, constructing a time-slot-oriented LSTM neural network, and performing time-series sliding window processing, splicing processing and normalization processing on input data of the neural network according to the specified sliding window length;
step S13, taking the data of N clusters, and respectively training N LSTM neural networks;
the parameter adjusting stage comprises the following steps:
step S21, preprocessing the verification set data, specifically including:
setting the length of a sliding window as seq _ len, performing time series sliding window processing on each sequence in the verification set, wherein the time series window size is seq _ len, the sliding window step length is 1, and then performing splicing processing and normalization processing;
step S22, inputting the processed verification set data into L neural networks according to given standards, wherein L is less than N, and performing denormalization operation and weighted summation on L prediction results;
step S23, the weighted summation result is corrected by using the past data, and the final prediction result is output;
step S24, adjusting parameters, carrying out clustering and neural network training by using the training set data again, and then repeating S22 and S23 and carrying out error analysis;
step S25, repeating step S24 for a plurality of times, and finally reserving parameters which enable the three indexes to be relatively best in performance as model parameters;
The testing stage comprises the following steps:
step S31, performing the same preprocessing operation as step S21 on the test set data;
step S32, inputting the processed test set data into L neural networks according to a given standard, wherein L is less than N, performing denormalization operation and weighted summation on L prediction results, correcting the weighted summation result by using the past data, and outputting a final prediction result;
step S33, the training set data is reused to carry out clustering and neural network training, and then step S32 is carried out;
and step S34, repeating the step S33M times, and calculating the mean value and standard deviation of the MAE, the MSE and the MAPE of the corrected predicted value and the true value.
Further, the vehicle track information is derived from a network data set;
the research area is an urban traffic hot spot area;
the acquired vehicle track information is the motor vehicle track points acquired after the interval time;
accumulating the number of track points in a selected research area within a specified interval time to obtain a traffic flow density data set of the specified interval time, wherein the specified interval time is formulated according to a prediction demand;
if the specified interval time is delta t and the frequency of the occurrence frequency of the vehicle track points in the research area in the interval time is n, the traffic flow density data is defined as follows:
Figure BDA0002942843750000031
And dividing the traffic flow density data set into a training set, a verification set and a test set.
Further, the specific implementation manner of obtaining N clusters in step S11 is as follows;
firstly, the similarity among time series slices is calculated by utilizing a dynamic time warping method, the specific implementation mode is as follows,
the dynamic time warping method carries out local extension or shortening on a time axis to calculate the similarity between time sequences, if two time sequences needing to calculate the similarity are X and Y, the lengths are | X | and | Y |, i represents the traffic flow density at the i moment in X, j represents the traffic flow density at the j moment in Y, the rounding path W is defined as follows:
W＝w1,w2,...,wk,Max(|X|,|Y|)≤k≤|X|+|Y|
in the formula: w is a1＝(1,1)，wk＝(|X|,|Y|)；
The cumulative rounding distance D (i, j) from (1,1) to (i, j) is defined as:
D(i,j)＝Dist(i,j)+Min[D(i-1,j),D(i,j-1),D(i-1,j-1)]
in the formula: dist (i, j) represents Xi,YjThe distance between the two points;
solving by using dynamic programming to obtain a normalized path distance D (| X |, | Y |);
then, classifying the traffic flow density slice sequences according to the similarity by a K-Means clustering method to obtain various typical sequences of the traffic flow sequences, and integrating N clustering clusters and clustering centers.
Further, the specific implementation manner of performing time-series sliding window processing, splicing processing and normalization processing on the input data according to the specified length in step S12 is as follows;
The time series sliding window processing comprises the following steps:
if the traffic flow density slice data in the cluster is expressed as S1,S2,...,SnAnd (5) setting the length of a traffic flow density slice as time _ span and the length of a sliding window as seq _ len, and firstly, Si(i-1.. multidot.n) is spliced with the first seq _ len-1 traffic flow density values which are logically corresponding to the time in the non-sliced traffic flow density data set and recorded as
Figure BDA0002942843750000041
Then to
Figure BDA0002942843750000042
Carrying out time sequence sliding window processing with the time sequence window size of seq _ len and the sliding window step length of 1;
the splicing treatment comprises the following steps:
recording x as a traffic flow density slice cut by the sliding window, and recording y as a traffic flow density value at the next moment of the slice;
then a traffic flow density sliding window data set is obtained and expressed as:
{(xi,yi)}1≤i≤n×time_span
the normalization process is: for { x0,...,xnAnd (5) defining the normalized traffic flow density as follows:
Figure BDA0002942843750000043
wherein xk∈{x0,...,xn}；
The normalized traffic flow density sliding window data set is the data received in the neural network
Figure BDA0002942843750000046
Further, the given criteria in step S22 are divided into two criteria, one is a time stamp as a criterion, and the other is a distance from the cluster center as a criterion;
the time label is taken as a standard, namely, the time label of input data is judged firstly, then the neural network with the same time label is selected, and if the training set of the neural network does not contain the time label, the data is not input into the neural network;
Calculating the DTW distance between the input data and the clustering center of the clustering cluster corresponding to each neural network by taking the distance between the input data and the clustering center as a standard, and selecting the neural network for prediction when the distance is smaller than a given threshold value;
the denormalization operation being the prediction of the result
Figure BDA0002942843750000044
Conversion to { x ] according to the following formula0,...,xn}：
Figure BDA0002942843750000045
Wherein xk∈{x0,...,xn}；
The weighted summation is also divided into two types, and the weighted summation corresponds to the given standard in two input types respectively;
setting the L denormalized prediction results as y by taking the time label as a standard1,...,yLIn the cluster corresponding to the kth neural network, the slice having the same time stamp as the input data has N in totalkAnd if the total number of slices containing the time label in the training set is N, the weight of the predicted value output by the kth neural network is as follows:
Figure BDA0002942843750000051
using the distance from the cluster center as the standard, if the input data and the cluster center
Figure BDA0002942843750000056
The DTW distance of (1) is decreased in sequence, and the corresponding prediction result is y1,...,yLAnd then:
c1≤...≤cL
c1+...+cL＝1
finally, the predicted values obtained by weighting are:
Figure BDA0002942843750000052
further, the specific implementation manner of step S23 is as follows;
if the input data of the neural network is in the designated correction time period, n past time sequence data { p) with the same type as the working day of the input data of the neural network exist 0,...,pn-1The weighted traffic flow density predicted value is OwAnd defining the corrected traffic flow density predicted value O as:
Figure BDA0002942843750000053
in the formula: α represents the similar day correction coefficient, and d represents the correction day range.
Further, the error analysis in step S24 includes the calculation of three indexes MAE, MAPE and MSE, which are defined as follows:
Figure BDA0002942843750000054
Figure BDA0002942843750000055
Figure BDA0002942843750000061
in the formula: y represents the true value of the traffic flow density, O represents the predicted value of the traffic flow density, and N represents the number of the predicted values of the traffic flow density. Compared with the prior art, the invention has the following advantages and beneficial effects:
the concept of traffic flow density is creatively provided, the traffic flow density data set of each vehicle in the area is obtained, and the actual traffic jam condition in the urban traffic hot spot area can be effectively reflected.
Local extension or shortening is carried out on a time axis by utilizing a dynamic time warping method, so that the similarity among the traffic flow density sequence slices is more accurately measured.
And classifying the traffic flow density slice sequences according to the similarity by using a K-Means clustering method, so as to be beneficial to reflecting the typical mode of the traffic flow density.
An initial normalization method is provided, and experiments prove that the method is more stable.
Respectively training neural networks aiming at traffic flow density data with similar patterns, and weighting and summing the output predicted values of the neural networks according to the time labels or the distances of the data clustering centers in the integrated learning step, thereby fully utilizing the characteristics of the sequence structure of the data.
And performing similar day correction on the weighted summation result by using the same type of past data of the working day, and further improving the algorithm prediction accuracy by using the long-time dependency of the traffic flow density.
And calculating the mean absolute error MAE of the three indexes, performing error analysis on the mean absolute percentage error MAPE and the mean square error MSE, and describing the deviation and the stability of the algorithm in multiple angles.
Drawings
FIG. 1 is a general flow chart of a time slot-oriented LSTM traffic flow density prediction method;
FIG. 2 is a flow chart of a training phase of the LSTM traffic flow density prediction method facing a time period;
FIG. 3 is a flow chart of a validation phase of the time slot-oriented LSTM traffic flow density prediction method;
fig. 4 is a flow chart of the testing stage of the LSTM traffic flow density prediction method facing the time slot.
Detailed Description
The model is further described in detail below with reference to the accompanying drawings and implementation examples. The specific embodiments discussed herein are merely illustrative of the present model and do not delimit the present model.
As shown in fig. 1, the invention provides a time slot-oriented LSTM traffic flow density prediction method, which is divided into a data processing stage, a training stage, a parameter adjusting stage and a testing stage, wherein the data processing stage specifically includes:
Selecting a research area, selecting a part which has intersection with the area from the acquired vehicle track information, calculating a traffic flow density sequence of vehicles in the area according to a traffic flow density definition formula, and dividing the sequence into a training set, a verification set and a test set.
The vehicle trajectory information is derived from a network data set.
The research area is an urban traffic hot spot area.
The acquired vehicle track information is the number of motor vehicle track points acquired after the interval time.
And accumulating the number of the track points in the selected study area within the specified interval time to obtain a traffic flow density data set of the specified interval time, wherein the specified interval time can be set according to the prediction requirement (for example, 2 minutes).
If the specified interval time is delta t and the frequency of the occurrence frequency of the vehicle track points in the research area in the interval time is n, the traffic flow density data is defined as follows:
Figure BDA0002942843750000071
different from the traditional traffic flow definition, the traffic flow density is the number of track points in unit time, and the actual congestion condition can be reflected better. In practical situations, a phenomenon that traffic flow is large but traffic is smooth may occur, vehicle track points are a two-dimensional time sequence formed by acquiring longitude and latitude data of vehicles every time a fixed time interval passes, and if the number of track points left by all vehicles in a certain area is large in unit time, the probability is that each vehicle forms a stagnation in the area.
And dividing the traffic flow density data set into a training set, a verification set and a test set. Specifically, if the acquired vehicle track information contains T days in total, the traffic flow density data set of the previous T-2 days is used as a training set, the data set of the T-1 day is used as a verification set, and the data set of the T day is used as a test set.
As shown in fig. 2, the training phase includes the following steps:
and step S11, dividing the traffic flow density data in the training set according to a given length to obtain time series slices, and recording the time labels of the slices according to the positions of the slices in the day. And calculating the similarity among the time sequence slices by using a dynamic time warping method, and classifying the traffic flow density slices according to the similarity by using a K-Means clustering method to obtain N clustering clusters, wherein the slices in each cluster have respective time labels.
The given length may be established according to predicted demand (e.g. 3 hours).
The Dynamic Time Warping (Dynamic Time Warping) method is a sequence measurement method, and local extension or shortening is performed on a Time axis to calculate the similarity between Time sequences.
The dynamic time warping method carries out local extension or shortening on a time axis to calculate the similarity between time sequences, if two time sequences needing to calculate the similarity are X and Y, the lengths are | X | and | Y |, i represents the traffic flow density at the time i in X, j represents the traffic flow density at the time j in Y, the rounding path W is defined as:
W＝w1,w2,...,wk,Max(|X|,|Y)≤k≤|X|+|Y|
In the formula: w is a1＝(1,1)，wk＝(|X|,|Y|)；
The cumulative rounding distance D (i, j) from (1,1) to (i, j) is defined as:
D(i,j)＝Dist(i,j)+Min[D(i-1,j),D(i,j-1),D(i-1,j-1)]
in the formula: dist (i, j) denotes Xi,YjThe distance between the two points;
solving by using dynamic programming to obtain a normalized path distance D (| X |, | Y |);
the K-Means clustering classifies the traffic flow density slices according to similarity (dynamic time warping measurement) to obtain various typical sequences of traffic flow sequences, and integrates N clustering clusters and clustering centers. The value of N may be formulated based on predicted demand.
And step S12, constructing an LSTM neural network facing a time period, and performing time series sliding window processing, splicing processing and normalization processing on input data of the neural network according to the specified sliding window length.
The time series sliding window processing comprises the following steps:
if the traffic flow density slice data in the cluster is expressed as S1,S2,...,SnAnd (5) setting the length of a traffic flow density slice as time _ span and the length of a sliding window as seq _ len, and firstly, Si(i 1.. said, n) is spliced with the first seq _ len-1 traffic flow density values which are logically corresponding to the time in the non-sliced traffic flow density data set and recorded as
Figure BDA0002942843750000081
Then to
Figure BDA0002942843750000082
Performing time series sliding window processing with the time series window size of seq _ len and the sliding window step length of 1;
the splicing treatment comprises the following steps:
recording x as a traffic flow density slice cut by the sliding window, and recording y as a traffic flow density value at the next moment of the slice;
Then the traffic flow density sliding window data set is obtained and expressed as:
{(xi,yi)}1≤i≤n×time_span
the normalization process is: for { x0,...,xnIntersection after normalizationThe through-flow density is defined as:
Figure BDA0002942843750000083
wherein xk∈{x0,...,xn}；
The normalized traffic flow density sliding window data set is the data received in the neural network
Figure BDA0002942843750000084
And step S13, taking the data of the N clusters, and respectively training N LSTM neural networks.
As shown in fig. 3, the parameter adjusting stage includes the following steps:
step S21, preprocessing the verification set data, specifically including:
and setting the length of the sliding window as seq _ len, performing time series sliding window processing with the time series window size of seq _ len and the sliding window step length of 1 on each sequence in the verification set, and then performing splicing processing and normalization processing which are the same as S12.
And step S22, inputting the processed data into L neural networks (L < N) according to given standards, and performing denormalization operation and weighted summation on L prediction results.
The neural network consists of 1 or 2 layers (adjustable parameters) of LSTM layers, a Dropout layer behind the LSTM layer and a last layer of full-connection layer, the activation function of the full-connection layer is a linear function, the network uses an Adam optimizer, and 1 traffic flow density value is output as a predicted value at the next moment when the traffic flow density of seq _ len length is received.
The given criteria are one based on time labels and one based on distances from the cluster center.
The time labels are used as the standard, namely the time labels of input data are firstly judged, and then the neural networks with the same time labels are selected. If the training set of a neural network does not contain the time tag, no data is input into the neural network.
And calculating the DTW distance between the input data and the clustering center of the clustering cluster corresponding to each neural network by taking the distance between the input data and the clustering center as a standard, and selecting the network for prediction when the distance is smaller than a given threshold value.
The data input process selects the neural network trained by the data similar to the data mode to be predicted, makes full use of the data structure characteristics of the traffic flow density sequence, and has strong pertinence; and the neural network with weak pattern similarity is omitted, so that the calculation amount is saved.
The denormalization operation being the prediction of the result
Figure BDA0002942843750000091
Conversion to { x ] according to the following formula0,...,xn}：
Figure BDA0002942843750000092
Wherein xk∈{x0,...,xn}；
There are also two ways of weighted summation, corresponding to the given criteria at the two inputs.
Setting the L denormalized prediction results as y by taking the time label as a standard1,...,yLIn the cluster corresponding to the kth neural network, the slice having the same time stamp as the input data has N in total kAnd if the total number of slices containing the time label in the training set is N, the weight of the predicted value output by the kth neural network is as follows:
Figure BDA0002942843750000093
using the distance from the cluster center as the standard, if the input data and the cluster center
Figure BDA0002942843750000094
The DTW distance of (1) is decreased in sequence, and the corresponding prediction result is y1,...,yLAnd then:
c1≤...≤cL
c1+...+cL＝1
finally, the predicted values obtained by weighting are:
Figure BDA0002942843750000101
in step S23, the weighted sum result is corrected using the past data, and the final prediction result is output.
If the input data of the neural network is in the specified correction time period, n past time sequence data { p) with the same type as the working day of the input data of the neural network exist0,...,pn-1The weighted traffic flow density predicted value is OwAnd defining the corrected traffic flow density predicted value O as:
Figure BDA0002942843750000102
in the formula: alpha represents the correction coefficient of the similar day, and d is the range of the correction days.
And step S24, adjusting parameters, carrying out clustering and neural network training by using the training set data again, and then repeating S22 and S23 and carrying out error analysis.
The error analysis includes the calculation of three indices, MAE, MAPE and MSE, defined as follows:
Figure BDA0002942843750000103
Figure BDA0002942843750000104
Figure BDA0002942843750000105
in the formula: y represents the true value of the traffic flow density, O represents the predicted value of the traffic flow density, and N represents the number of the predicted values of the traffic flow density.
And step S25, repeating the step S24 for a plurality of times, and finally reserving parameters which enable the three indexes to be relatively best in performance as model parameters.
As shown in fig. 4, the test phase performs a test to analyze the predicted result and the actual value, and exemplifies the effectiveness of the model. The method comprises the following specific steps:
step S31, preprocessing the test set data in the same way as S21;
step S32, inputting the processed test set data into L neural networks (L < N) by taking a time label or a distance from a cluster center as a standard, performing denormalization operation and weighted summation on L prediction results, correcting the weighted summation result by using past data, and outputting a final prediction result;
step S33, clustering and neural network training are carried out by using the training set data again, and then step S32 is carried out;
and step S34, repeating the step S33M times, and calculating the Mean value (Mean) and standard deviation (Std) of the MAE, the MSE and the MAPE of the corrected predicted value and the real value so as to simulate the prediction effect of the model under the real condition.
The effectiveness of the present invention can be further exemplified by the following simulation experiments. It is noted that the parameters used in the experiments do not influence the generality of the present invention.
1) Simulation conditions are as follows:
8G memory, 128G hard disk, Intel Core i3-8750H processor, operating system Windows, simulation software SQL, SPSS and Jupyter Notebook.
2) Simulation content:
analytical modeling was performed using the KDD CUP 2020 dataset of the dropletin cover data open platform. Selecting local areas of the metropolis, recording original vehicle track data at intervals of 2-4 seconds within the time range of 2016 (11 months, 1 day) to 11 months, and recording the original vehicle track data at intervals of 2 minutes, and normalizing the data by using the method. In the experiment, traffic flow density data from 11 month 1 to 11 month 28 days are used as a training set, traffic flow density data from 11 month 29 days is used as a verification set, and traffic flow density data from 11 month 30 days is used as a test set.
The model proposed by the method of the invention (denoted as TPO) was compared with a classical time series prediction model, including the recurrent neural network LSTM (denoted as C). TPO and C are subjected to parameter adjustment work on a verification set, and respective optimal parameters are obtained.
The traffic flow density in the local area was predicted in 11 months and 30 days, and error analysis was performed, and the results are shown in table 1.
Compared with the C algorithm, the MAE of the invention is 3.3 multiplied by 10 less on the mean value46X 10 less standard deviation7(ii) a MAPE is 0.09 less in mean value and 0.014 less in standard deviation, and the accuracy and stability of traffic flow density prediction are improved.
TABLE 1 comparison of Performance of the prediction models
Figure BDA0002942843750000111
The foregoing describes a specific embodiment of the present invention. It should be understood that the above description is only a specific embodiment of the present invention, and is not intended to limit the present invention, and those skilled in the art may make various changes or modifications within the scope of the appended claims without affecting the spirit of the present invention.

Claims (5)

1. A time slot-oriented LSTM traffic flow density prediction method is characterized by comprising the following steps: a data processing stage, a training stage, a parameter adjusting stage and a testing stage;
the data processing stage specifically comprises: selecting a research area, selecting a part which has intersection with the research area from the acquired vehicle track information, calculating a traffic flow density data set of each vehicle in the area according to a traffic flow density formula, and dividing the traffic flow density data set into a training set, a verification set and a test set;
the vehicle track information is derived from a network data set;
the research area is an urban traffic hot spot area;
the acquired vehicle track information is the motor vehicle track points acquired after the interval time;
accumulating the number of track points in the selected research area within the specified interval time to obtain a traffic flow density data set of each road segment within the specified interval time, wherein the specified interval time is made according to the prediction requirement;
If the specified interval time is delta t and the frequency of the occurrence frequency of the vehicle track points in the research area in the interval time is n', the traffic flow density data is defined as follows:
Figure FDA0003609452510000011
dividing a traffic flow density data set into a training set, a verification set and a test set; the training phase comprises the following steps:
s11, segmenting traffic flow density data in a training set according to a given length to obtain time sequence slices, recording time labels of the time sequence slices, calculating similarity among the time sequence slices by using a dynamic time warping method, and classifying the traffic flow density slices according to the similarity by using a K-Means clustering method to obtain N clustering clusters and corresponding time labels;
step S12, constructing a time-slot-oriented LSTM neural network, and performing time-series sliding window processing, splicing processing and normalization processing on input data of the neural network according to a specified length;
step S13, taking the data of N clusters, and respectively training N LSTM neural networks;
the parameter adjusting stage comprises the following steps:
step S21, preprocessing the verification set data, specifically including:
setting the length of a sliding window as seq _ len, performing time series sliding window processing on each sequence in the verification set, wherein the time series window size is seq _ len, the sliding window step length is 1, and then performing splicing processing and normalization processing;
Step S22, inputting the processed verification set data into L neural networks according to given standards, wherein L is less than N, and performing denormalization operation and weighted summation on L prediction results;
step S23, the weighted summation result is corrected by using the past data, and the final prediction result is output;
the specific implementation of step S23 is as follows;
if there are m past time sequence data { p) with the same type as the working day of the input data0,…pj ,…,pm-1The weighted traffic flow density predicted value is OwAnd defining the corrected traffic flow density predicted value O as:
Figure FDA0003609452510000021
in the formula: alpha represents a similar day correction coefficient, and d is a correction day range;
step S24, adjusting parameters, carrying out clustering and neural network training by using the training set data again, and then repeating S22 and S23 and carrying out error analysis;
step S25, repeating step S24 for multiple times, and finally reserving parameters which enable the best relative performance in error analysis as model parameters;
the testing phase comprises the following steps:
step S31, the preprocessing operation of the step S21 is carried out on the test set data;
step S32, inputting the processed test set data into L neural networks according to a given standard, wherein L is less than N, performing inverse normalization operation and weighted summation on L prediction results, correcting the weighted summation result by using past data, and outputting a final prediction result;
Step S33, the training set data is reused to carry out clustering and neural network training, and then step S32 is carried out;
and step S34, repeating the step S33M times, and predicting the traffic flow density.
2. The time slot-oriented LSTM traffic flow density prediction method according to claim 1, characterized in that: the specific implementation manner of performing time series sliding window processing, splicing processing and normalization processing on the input data according to the specified length in step S12 is as follows;
the time series sliding window processing comprises the following steps:
if the traffic flow density slice data in the cluster is expressed as S1,S2,...,SnThe length of a traffic flow density slice is time _ span, the length of a sliding window is seq _ len, and S is firstly carried outiThe first seq _ len-1 traffic flow density values which are logically corresponding to the time in the non-sliced traffic flow density data set are spliced and recorded as
Figure FDA0003609452510000022
Then to
Figure FDA0003609452510000023
Performing time-series sliding window processing with a time-series window size seq _ len and a sliding window step size of 1, wherein i is 1, …, n;
the splicing treatment comprises the following steps:
recording x as a traffic flow density slice cut by the sliding window, and recording y as a traffic flow density value at the next moment of the slice;
then a traffic flow density sliding window data set is obtained and expressed as:
{(xa,ya)}1≤a≤n×time_span
the normalization process is: for { xa,1,...,xa,seq_lenNormalizing, wherein the normalized traffic flow density is defined as:
Figure FDA0003609452510000031
Wherein l 1., seq _ len;
the normalized traffic flow density sliding window data set is the data received in the neural network
Figure FDA0003609452510000032
3. The time slot-oriented LSTM traffic flow density prediction method according to claim 2, characterized in that: the given criteria in step S22 are divided into two criteria, one is a criterion of time stamp, and the other is a criterion of distance from the cluster center;
the time labels are taken as the standard, namely, the time labels of input data are judged firstly, then the neural networks with the same time labels are selected, and if the training set of one neural network does not contain the time labels, the data are not input into the neural network;
calculating the DTW distance between the input data and the clustering center of the clustering cluster corresponding to each neural network by taking the distance between the input data and the clustering center as a standard, and selecting the neural network for prediction when the distance is smaller than a given threshold value;
the normalization operation is to perform normalization operation on the prediction result to obtain y1,...,yLThe denormalization operation is in the opposite way to the normalization operation;
the weighted summation is also divided into two types, and the weighted summation corresponds to the given standard in two input types respectively;
setting the L denormalized prediction results as y by taking the time label as a standard 1,...,yLIn the cluster corresponding to the kth neural network, the slice having the same time stamp as the input data has N in totalkAnd if the total number of slices containing the time label in the training set is N, the weight of the predicted value output by the kth neural network is as follows:
Figure FDA0003609452510000033
wherein k is 1-L;
using the distance from the cluster center as the standard, if the input data and the cluster center
Figure FDA0003609452510000034
The DTW distance of (1) is decreased in sequence, and the corresponding prediction result is y1,...,yLAnd then:
c1≤...≤ck≤...≤cL
c1+...+ck+...+cL＝1
finally, the predicted values obtained by weighting are:
Figure FDA0003609452510000035
wherein k is 1-L.
4. The time slot-oriented LSTM traffic flow density prediction method according to claim 1, characterized in that: the error analysis in step S24 includes the calculation of three indices MAE, MAPE and MSE, defined as follows:
Figure FDA0003609452510000041
Figure FDA0003609452510000042
Figure FDA0003609452510000043
in the formula: y represents the true value of the traffic flow density, O represents the predicted value of the traffic flow density, and N represents the number of the predicted values of the traffic flow density.
5. The time slot-oriented LSTM traffic flow density prediction method according to claim 1, characterized in that: in step S34, the mean and standard deviation of the MAE, MSE, MAPE of the corrected predicted value and the true value are calculated to simulate the prediction effect under the true condition.
CN202110185310.5A 2021-02-10 2021-02-10 Time-period-oriented LSTM traffic flow density prediction method Active CN112884222B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110185310.5A CN112884222B (en) 2021-02-10 2021-02-10 Time-period-oriented LSTM traffic flow density prediction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110185310.5A CN112884222B (en) 2021-02-10 2021-02-10 Time-period-oriented LSTM traffic flow density prediction method

Publications (2)

Publication Number Publication Date
CN112884222A CN112884222A (en) 2021-06-01
CN112884222B true CN112884222B (en) 2022-06-14

Family

ID=76056434

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110185310.5A Active CN112884222B (en) 2021-02-10 2021-02-10 Time-period-oriented LSTM traffic flow density prediction method

Country Status (1)

Country Link
CN (1) CN112884222B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113570004B (en) * 2021-09-24 2022-01-07 西南交通大学 Riding hot spot area prediction method, device, equipment and readable storage medium
CN115457764B (en) * 2022-08-24 2023-07-18 华南理工大学 Road section traffic density estimation method, device and medium based on vehicle track data

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
CN109377752A (en) * 2018-10-19 2019-02-22 桂林电子科技大学 Short-term traffic flow variation prediction method, apparatus, computer equipment and storage medium
CN110889546A (en) * 2019-11-20 2020-03-17 浙江省交通规划设计研究院有限公司 Attention mechanism-based traffic flow model training method
CN111292534A (en) * 2020-02-13 2020-06-16 北京工业大学 Traffic state estimation method based on clustering and deep sequence learning

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
CN109377752A (en) * 2018-10-19 2019-02-22 桂林电子科技大学 Short-term traffic flow variation prediction method, apparatus, computer equipment and storage medium
CN110889546A (en) * 2019-11-20 2020-03-17 浙江省交通规划设计研究院有限公司 Attention mechanism-based traffic flow model training method
CN111292534A (en) * 2020-02-13 2020-06-16 北京工业大学 Traffic state estimation method based on clustering and deep sequence learning

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
Cluster-Based LSTM Network for Short-Term Passenger Flow Forecasting in Urban Rail Transit;Jinlei Zhang，etc;《IEEE》;20190917;第7卷;全文 *
Short-Term Traffic Flow Prediction using DTW-BiGRU Model;Hanyi Zou，etc;《2020 35th Youth Academic Annual Conference of Chinese Association of Automation (YAC)》;20210205;全文 *
基于LSTM的船舶航迹预测模型;权波等;《计算机科学》;20181115;全文 *
基于卷积循环神经网络的城市区域车流量预测模型;薛佳瑶等;《信息工程大学学报》;20190415(第02期);全文 *
基于时空关联度加权的LSTM短时交通速度预测;刘易诗等;《地理信息世界》;20200225(第01期);全文 *

Also Published As

Publication number Publication date
CN112884222A (en) 2021-06-01

Similar Documents

Publication Publication Date Title
CN107610464B (en) A kind of trajectory predictions method based on Gaussian Mixture time series models
CN110097755B (en) Highway traffic flow state identification method based on deep neural network
CN110570651A (en) Road network traffic situation prediction method and system based on deep learning
CN112884222B (en) Time-period-oriented LSTM traffic flow density prediction method
CN111105104A (en) Short-term power load prediction method based on similar day and RBF neural network
CN113902011A (en) Urban rail transit short-time passenger flow prediction method based on cyclic neural network
CN112396234A (en) User side load probability prediction method based on time domain convolutional neural network
CN112966871A (en) Traffic jam prediction method and system based on convolution long-short term memory neural network
CN117977568A (en) Power load prediction method based on nested LSTM and quantile calculation
CN114548592A (en) Non-stationary time series data prediction method based on CEMD and LSTM
CN105913078A (en) Multi-mode soft measurement method for improving adaptive affine propagation clustering
CN116187835A (en) Data-driven-based method and system for estimating theoretical line loss interval of transformer area
CN115496257A (en) Short-term vehicle speed prediction based on space-time fusion
CN113436433A (en) Efficient urban traffic outlier detection method
CN115269758A (en) Passenger-guidance-oriented road network passenger flow state deduction method and system
Wibawa et al. Long Short-Term Memory to Predict Unique Visitors of an Electronic Journal
Yang et al. A carbon price hybrid forecasting model based on data multi-scale decomposition and machine learning
CN113253363A (en) Lightning activity path prediction method and system
CN112330051A (en) Short-term load prediction method based on Kmeans and FR-DBN
Dhole et al. An ensemble approach to multi-source transfer learning for air quality prediction
Guo et al. Mobile user credit prediction based on lightgbm
Mao et al. Naive Bayesian algorithm classification model with local attribute weighted based on KNN
Mead Hybrid CNN and LSTM Model (HCLM) for Short-Term Traffic Volume Prediction
CN113657533A (en) Multi-element time sequence segmentation clustering method for space-time scene construction
CN114997366A (en) Protein structure model quality evaluation method based on graph neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant