CN112330158B - Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network - Google Patents

Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network Download PDF

Info

Publication number
CN112330158B
CN112330158B CN202011231975.7A CN202011231975A CN112330158B CN 112330158 B CN112330158 B CN 112330158B CN 202011231975 A CN202011231975 A CN 202011231975A CN 112330158 B CN112330158 B CN 112330158B
Authority
CN
China
Prior art keywords
traffic index
time sequence
data
traffic
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
CN202011231975.7A
Other languages
Chinese (zh)
Other versions
CN112330158A (en
Inventor
张学东
卢剑
张健钦
徐志洁
王家川
石瑞轩
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.)
BEIJING TRANSPORTATION INFORMATION CENTER
Beijing University of Civil Engineering and Architecture
Original Assignee
BEIJING TRANSPORTATION INFORMATION CENTER
Beijing University of Civil Engineering and Architecture
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 BEIJING TRANSPORTATION INFORMATION CENTER, Beijing University of Civil Engineering and Architecture filed Critical BEIJING TRANSPORTATION INFORMATION CENTER
Priority to CN202011231975.7A priority Critical patent/CN112330158B/en
Publication of CN112330158A publication Critical patent/CN112330158A/en
Application granted granted Critical
Publication of CN112330158B publication Critical patent/CN112330158B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • 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—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Human Resources & Organizations (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Development Economics (AREA)
  • Educational Administration (AREA)
  • Tourism & Hospitality (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Business, Economics & Management (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Game Theory and Decision Science (AREA)
  • Primary Health Care (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention relates to a method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network, which comprises the following steps: acquiring a traffic index original data set and generating a traffic index time sequence; converting the traffic index time sequence into a stable sequence, and fitting an autoregressive differential moving average algorithm model according to a Bayesian information criterion matrix to realize traffic index prediction; generating a training traffic index time sequence and a test traffic index time sequence according to the traffic index original data set, extracting traffic index feature information according to the training traffic index time sequence, acquiring an optimal convolutional neural network model, integrating the traffic index feature information into a one-dimensional feature vector by utilizing the optimal convolutional neural network model, determining the mode category of the one-dimensional feature vector according to a Softmax classifier, and further identifying the category to which the test traffic index time sequence belongs.

Description

Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network
Technical Field
The invention relates to a method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network.
Background
Traffic is the pulse of a city. With the rapid development of urban economy, traffic jam problems are also becoming more severe, and areas with more developed economy are more prominent. In order to cope with complex and changeable traffic conditions of cities and relieve urban traffic pressure, a series of traffic laws and regulations are set out by traffic management departments to restrict driving regulations, and scientific research institutions also use technologies such as the Internet of things to assist traffic management departments in monitoring road congestion. Although various industries achieve certain effects of reducing the risk of traffic jam from various aspects, the adverse effects of the traffic jam on cities cannot be eliminated at present in the face of the influence of complex environments, emergency conditions, human intervention and other factors. The traffic index is an important index for researching the running state of urban traffic, reflects the quantification result of urban congestion, and has a certain rule in the time dimension. Traffic index is generally represented by a value of 0-10, with larger values representing more congested traffic and worse traffic conditions. Therefore, if the potential characteristics of resident travel are obtained from the historical traffic indexes through the time sequence pattern recognition, the category of the resident travel is distinguished, basic data can be provided for traffic running state research and prediction, and the method has important value for city traffic jam relief.
At present, the traffic passenger flow index prediction method mainly comprises the following steps: kalman filtering recursive algorithm (Recursive algorithm of Kalman filter), gray Theory (Grey Theory), support vector machine (Support Vector Machine), deep Learning (Deep Learning) and other methods. Although iterative estimation models based on Kalman filter recursion algorithms have been widely used in passenger flow predictions, kalman filter recursion algorithms require a large number of matrix and vector operations, and suffer from an insufficient computational efficiency (PeiYan, biswasSwarnendu, fusseldonald S, et al, an elementary introduction to Kalman filtering [ J ]. Communications of the ACM,2019.Baptista M,Henriques E M P,De Medeiros I P,et al.Remaining useful life estimation in aeronautics:Combining data-driven and Kalman filtering [ J ]. Reliability Engineering & System Safety,2019,184 (APR.): 228-239.Yang D.On post-processing day-ahead NWP forecasts using Kalman filtering [ J ]. Solar Energy,2019,182 (APR.): 179-181.). The gray theoretical model achieves the aim of prediction by identifying the degree of dissimilarity of the development trend among the system factors, but has better effect of predicting the data only in a short period and not in a long period (Wang Q, jiang F. Integrating linear and nonlinear forecasting techniques based on grey theory and artificial intelligence to forecast shale gas monthly production in Pennsylvania and Texas of the United States [ J ]. Energy,2019,178.Huiming Duan,Di Wang,Xinyu Pang,Yunmei Liu,Suhua Zeng,A novel forecasting approach based on Multi-Kernel Nonlinear Multivariable Grey Model: a case report, journal of Cleaner Production (2020)). The support vector machine maps the input layers to a high dimensional space and solves the separation hyperplane with very large computational effort (Fan S, chen L.short-Term Load Forecasting Based on an Adaptive Hybrid Method [ J ]. IEEE Transactions on Power Systems,2006, 21 (1): p.392-401.Kalra A,Ahmad S.Using Oceanic-Atmospheric Oscillations for Long Lead Time Streamflow Forecasting [ J ]. Water Resources Research,2009,45 (3) [ Gwo-Fong, lin, et al typhoon flood forecasting using integrated two-stage Support Vector Machine approach [ J ]. Journal of Hydrology, 2013). The Long-term memory network (Long Short Term Memory, LSTM) deep learning model alleviates the gradient vanishing problem in the recurrent neural network model, but the LSTM model requires a linear layer to run in each sequential time step, while the layer requires a large amount of memory bandwidth computation, which takes a Long time to train (Azzouni A, pujolle G.A Long Short-Term Memory Recurrent Neural Network Framework for Network Traffic Matrix Prediction [ J ].2017.Gers F A,Schraudolph N N,Schmidhuber,J u-rgen. Learning Precise Timing with LSTM Recurrent Networks [ J ]. Journal of Machine Learning Research,2003,3 (1): 115-143.Selvin S,Vinayakumar R,Gopalakrishnan E A,et al.Stock price prediction using LSTM,RNN and CNN-sliding window model [ C ]//2017International Conference on Advances in Computing,Communications and Informatics (ICACCI) & IEEE, 2017.). The method for identifying the traffic passenger flow mode mainly comprises the following steps: distance-based pattern recognition and feature-based pattern recognition. Distance-based pattern recognition generally uses Euclidean distance (Euclidean distance) to measure the similarity of traffic and passenger flows, such as the K-Nearest Neighbor (KNN) algorithm (Macqueen J.Some Methods for Classification and Analysis of Multi Variate Observations [ C ]. Proc of Berkeley Symposium on Mathematical Statistics & Probality, 1965. Saroj, kavita. Review: study on Simple K Mean and Modified K Mean Clustering Technique [ J ] International Journal of Computer Science Engineering and Technology,2016,6 (7): 279-281.). Feature-based pattern recognition algorithms typically look for differential subsections to distinguish the categories to which traffic flows belong, e.g., the shape algorithm looks for the most representative consecutive subsequences in the data (Zhu L, lu C, sun y. Time Series Shapelet Classification Based Online Short-Term Voltage Stability Assessment [ J ]. Power Systems, IEEE Transactions on,2016,31 (2): 1430-1439.). Although both types of pattern recognition methods can obtain good classification results under specific conditions, the pattern recognition methods are influenced by multiparty traffic factors, and the traffic index time series data have certain distortion and deformation phenomena, so that the conventional distance-based and feature-based recognition methods have certain defects on the pattern recognition capability of the traffic index time series data.
The autoregressive differential moving average algorithm (Auto-regressive Integrated Moving Average algorithm, ARIMA) refers to a model that is built by regressing a dependent variable only on its hysteresis value and the present and hysteresis values of a random error term in converting a non-stationary time series to a stationary time series (Ensor, bennett K.Time Series Analysis and Its Applications [ J ]. Journal of the American Statistical Association,2002,97 (458): 656-657.). The basic idea of the algorithm is to consider the data formed by the predicted object with time as a random sequence, describe the sequence by using a certain mathematical model, and predict future values according to past values by using the model, wherein the sequence is composed of an autoregressive process, an averaging process and a differential process. ARIMA includes an Auto-regressive Model (AR), a moving average Model (Moving Average Model, MA) and an Auto-Regressive Moving Average Model (ARMA) Model, depending on the stationarity of the time series data and the part involved in the regression analysis. The ARIMA prediction process for time series data is shown in fig. 2.
Prior to time series data prediction, data preprocessing is required, including randomness and stationarity detection. Data can be classified into three types by detection results: a purely random sequence, a stationary non-random sequence, and a non-stationary non-random sequence. The pure random sequence is also called a white noise sequence, and the pure random sequence is not related to each item of the sequence, is completely disordered and randomly distributed, and has no analysis and research values; the mean and variance of the stationary non-random sequence are constants, and fitting is generally performed through a linear model, so that a sequence rule is extracted; the mean and variance of the non-stationary non-random sequence are not determined and need to be converted to a stationary sequence by differential operation. The mean and variance of the stationary sequence data do not vary excessively over a period of time, and the fitted curve of the time sequence maintains the existing morphological continuation for a short time in the future. The plateau sequence includes both a strict plateau and a generalized plateau. Strictly and smoothly refers to data distribution not changing with time; the generalized plateau is that the expected and correlation coefficients of the sequence data do not change. Strict stability is often over-absolute, and data distribution in real life is mostly generalized stability.
The pure randomness test is detected by constructing test statistics, calculating p values corresponding to the test statistics, and if the p values are larger than the significance level alpha, the p values are the pure random sequences. For a non-random sequence, the time sequence is a stationary sequence if the fluctuation range of the time sequence at a certain moment is limited, i.e. there is a mean and a variance, and the periodic variation of the autocovariance is equal to the autocorrelation coefficient. The nature of the time series is generally determined using a time chart and a unit root test method.
Convolutional neural networks are neural networks that implement image processing by back propagation (Zhou Feiyan, jin Linpen, dong jun. Review of Convolutional Neural Network [ J ]. Chinese Journal of Computers,2017 (6)). As shown in fig. 3, it includes an Input layer (Input layer), a convolution layer (convolutional layer), a pooling layer (pooling layer), a full-connection layer (fully connected layer), and an output layer (output layer).
In convolutional neural networks, the input layer is multidimensional data. The input layer of the one-dimensional convolutional neural network is generally a one-dimensional or two-dimensional array; the input layer of the two-dimensional convolutional neural network is a two-dimensional or three-dimensional array; the input layer of the three-dimensional convolution is a four-dimensional array. Because of the characteristics of convolutional neural networks, the algorithm is widely applied in the field of computer image processing, so that input layers in many researches are two-dimensional pixel points or RGB images, and output layers can output the size, coordinates, classification and the like of objects according to research contents.
The function of the convolutional layer is to traverse the input layer through an internal convolutional kernel (convolution kernel) to achieve feature extraction (Palmieri F an, buonano a. Belief propagation and learning in convolution multi-layer factor graphs [ C ]// International Workshop on Cognitive Information processing.ieee, 2014.). The algorithm and size of the convolution kernel is determined according to the size of the input layer, and the convolution kernel is moved a fixed unit length on the input layer each time, as shown in fig. 4. Equation (5) is an expression in which the input layer performs a convolution operation with the convolution kernel.
Wherein, the liquid crystal display device comprises a liquid crystal display device,the result is output after convolution operation; />Convolving for layer j; />Is a bias parameter; f (x) is an activation function (activation function).
In order to solve the linear inseparable problem, the result after convolution operation needs to be converted in a nonlinear way by using an activation function again. Common activation functions include Sigmoid function (equation 6), tanh function (equation 7), and ReLU function (equation 8) (Lecun Y, bottou l.gradient-based learning applied to document recognition [ J ]. Proceedings of the IEEE,1998,86 (11): p.2278-2324.). The calculated amount of the Sigmoid function and the Tanh function is large, and meanwhile, the situation of gradient disappearance or gradient explosion can be caused, so that information is lost, and the neural network training is not facilitated; and the ReLU function can effectively reduce the influence of gradient disappearance (gradient disappearance) and gradient explosion (gradient explosion), so that the calculation process is optimized, the parameter dependence is reduced, and the probability of over-fitting phenomenon is reduced.
After the convolution layer and the pooling layer are processed, the characteristic information of the input layer is extracted, and the input layer is classified by the full-connection layer. The fully-connected layer integrates the feature layers, so that local features integrate information in a high dimension, and output a feature vector (Szegedy, c., liu, w., jia, y., seranet, p., reed, s., anguelov, d., erhan, d., vanhoucke, v.and Rabinovich, a.,2015.Going deeper with convolutions.In Proceedings of the IEEE conference on computer vision and pattern recognition (pp.1-9)) integrating all input layer feature information. For the integrated feature vectors, classification is performed using a Softmax function (Ran pen, wang Ling, li Xin, liu pengwei.improved Softmax Classifier for Deep Convolution Neural Networks and Its Application in Face Recognition [ J ]. Journal of Shanghai University (Natural Science), 2018,24 (03): 352-366.Simone Totaro,Amir Hussain,Simone Scardapane.A Non-parametric Softmax for Improving Neural Attention in Time-series Forecasting [ J ]. Neurostarting, 2019 ]), the probability of each class occurrence is calculated by the formula (9), and the corresponding class with the largest probability value is determined as the classification result according to the probability.
Wherein a is j A j-th value representing a vector in the fully connected layer; a, a k Representing each bit value in the full connection layer; t represents a preset number of classification categories.
Time series data refers to a sequence of random variables that change over time, differing from ordinary data in the effect of time on the data. The time intervals in the time series may be years, months, dates, or any other time interval, depending on the time of observation. Time series data is very common in the fields of finance, transportation, etc., such as stock exchanges and traffic volumes. Assume a set of random variables x= { X 1 ，X 2 ，……，X n Time is defined as t= { T } 1 ，t 2 ，……，t n Then define X t ＝{X 1 ，X 2 ，......，X t T e T is a time sequence within time T.
Disclosure of Invention
In order to solve the technical problems, the invention provides a method for identifying traffic index time sequence based on an autoregressive differential moving average-convolution neural network, which comprises the following steps:
0000, acquiring a traffic index original data set, and generating a traffic index time sequence;
step 1000, converting the traffic index time sequence into a stable sequence, fitting an autoregressive differential moving average algorithm model according to a Bayesian information criterion matrix, and realizing traffic index prediction by using the autoregressive differential moving average algorithm model;
And 2000, generating a training traffic index time sequence and a test traffic index time sequence according to the traffic index original data set, extracting traffic index feature information according to the training traffic index time sequence, acquiring an optimal convolutional neural network model, integrating the traffic index feature information into a one-dimensional feature vector by using the optimal convolutional neural network model, determining the mode category of the one-dimensional feature vector according to a Softmax classifier, and further identifying the category to which the test traffic index time sequence belongs.
Further, the method for generating the traffic index time sequence in the step 0000 includes:
step 0200: and calculating the average traffic index at each moment according to the traffic index original data set.
Step 0400: and selecting a first continuous time sequence and/or a second continuous time sequence, wherein the first continuous time sequence is a continuous time sequence with the average traffic index being greater than a first threshold value, and the second continuous time sequence is a continuous time sequence with the average traffic index being greater than a second threshold value and less than the first threshold value.
Step 0600: and dividing the traffic index original data set according to the traffic index original data set, the first continuous time sequence and/or the second continuous time sequence.
Step 0800: and generating a traffic index time sequence according to the division result of the step 0600.
Further, step 1000 includes:
step 1200: and (3) carrying out stability test on the traffic index time sequence, if the traffic index time sequence is a non-stable time sequence, turning to step 1400, otherwise turning to step 1600.
Step 1400: and (3) carrying out stabilization treatment on the traffic index time sequence, and turning to step 1200.
Step 1600: and selecting a Bayesian information criterion matrix to determine values of parameters p and q in each period according to the traffic index time sequence, and constructing an ARIMA model, wherein p is the autoregressive term number, and q is the moving average term number.
Step 1800: and fitting the traffic index time sequence through the ARIMA model.
In particular, the smoothing method in step 1400 is:
let { x } t T epsilon T is a group of time series data, B is a backward operator, namely Bx t ＝x t-1 RecordingCalled a difference operator, at which time,
generally there are
And (3) performing differential operation on the non-stationary time series data according to the formula (1) and the formula (2) to obtain a stationary sequence.
Further, the method of stationarity check in step 1200 is the ADF check.
In particular, the ARIMA model in step 1600 includes an autoregressive operation and an average movement process.
In step 2000, the CNN framework is formed by two groups of convolution pooling layers, where the convolution pooling layers include a convolution layer, a ReLU activation function, and a maximum pooling layer, where the convolution kernel of the convolution layer is 5*5, and the maximum pooling layer is 2×2.
The invention can accurately predict the traffic index time sequence data and the capability of identifying the mode, and can assist traffic management related departments to conduct traffic dispersion in advance so as to relieve urban traffic pressure.
Drawings
Fig. 1 is a flowchart of a first embodiment of the present invention.
FIG. 2 is a flowchart of ARIMA prediction in the present invention.
Fig. 3 is a schematic diagram of a convolutional neural network according to the present invention.
Fig. 4 is a schematic diagram of convolution operation in the present invention.
FIG. 5 (a) is a training traffic index time series of 05:00-07:15 time series data according to the present invention.
FIG. 5 (b) is a chart of time series data of 07:30-09:45 of training traffic index time series according to the present invention.
FIG. 5 (c) is a 10:00-12:15 time series of training traffic index time series data according to the present invention.
FIG. 5 (d) is a 12:30-14:45 time series of training traffic index time series data according to the present invention.
FIG. 5 (e) is 15:00-17:15 time series data of training traffic index time series in the present invention.
FIG. 5 (f) is a 17:30-19:45 time series of training traffic index time series of the present invention.
FIG. 5 (g) is a 20:00-23:00 time series of training traffic index time series data according to the present invention.
FIG. 6 (a) is an autocorrelation of a time series of 05:00-07:15 training traffic index in the present invention.
FIG. 6 (b) is an autocorrelation chart of the time series 07:30-09:45 of the training traffic index time series of the present invention.
FIG. 6 (c) is a 10:00-12:15 time series autocorrelation chart of the training traffic index time series in the present invention.
FIG. 6 (d) is a 12:30-14:45 time series autocorrelation graph of a training traffic index time series in the present invention.
FIG. 6 (e) is a 15:00-17:15 time series autocorrelation graph of a training traffic index time series in the present invention.
FIG. 6 (f) is a 17:30-19:45 time series autocorrelation graph of a training traffic index time series in the present invention.
FIG. 6 (g) is a 20:00-23:00 time series autocorrelation graph of training traffic index time series in the present invention.
FIG. 7 (a) is a graph showing the time series index difference between 05:00 and 07:15 of the time series of training traffic indexes according to the present invention.
FIG. 7 (b) is a graph showing the time series index differences between 07:30 and 09:45 of the training traffic index time series according to the present invention.
FIG. 7 (c) is a graph showing the time series index difference between 10:00 and 12:15 of the time series of training traffic indexes according to the present invention.
FIG. 7 (d) is a graph showing the time series index difference between 12:30 and 14:45 of the training traffic index time series in the present invention.
FIG. 7 (e) is a 15:00-17:15 time series index difference chart of the training traffic index time series in the present invention.
FIG. 7 (f) is a graph showing the time series index difference between 17:30 and 19:45 of the training traffic index time series in the present invention.
FIG. 7 (g) is a 20:00-23:00 time series index difference chart of the training traffic index time series in the present invention.
FIG. 8 (a) is a graph of the differential autocorrelation of the time series 05:00-07:15 of the training traffic index time series of the present invention.
FIG. 8 (b) is a differential autocorrelation chart of the time series 07:30-09:45 of the training traffic index time series of the present invention.
FIG. 8 (c) is a 10:00-12:15 time series differential autocorrelation diagram of a training traffic index time series in the present invention.
FIG. 8 (d) is a 12:30-14:45 time series differential autocorrelation graph of training traffic index time series in the present invention.
FIG. 8 (e) is a 15:00-17:15 time series differential autocorrelation graph of training traffic index time series in the present invention.
FIG. 8 (f) is a 17:30-19:45 time series differential autocorrelation graph of training traffic index time series in the present invention.
FIG. 8 (g) is a 20:00-23:00 time series differential autocorrelation graph of training traffic index time series in the present invention.
FIG. 9 (a) is a graph of differential partial autocorrelation of a time series of 05:00-07:15 training traffic index time series in the present invention.
FIG. 9 (b) is a differential partial autocorrelation chart of the time series 07:30-09:45 of the training traffic index time series in the present invention.
FIG. 9 (c) is a 10:00-12:15 time series differential partial autocorrelation chart of the training traffic index time series in the present invention.
FIG. 9 (d) is a graph of differential partial autocorrelation of the time series 12:30-14:45 of the training traffic index time series according to the present invention.
FIG. 9 (e) is a 15:00-17:15 time series differential partial autocorrelation graph of training traffic index time series in the present invention.
FIG. 9 (f) is a graph of differential partial autocorrelation of the time series 17:30-19:45 of the training traffic index time series of the present invention.
FIG. 9 (g) is a 20:00-23:00 time series differential partial autocorrelation graph of training traffic index time series in the present invention.
FIG. 10 is a graph showing the comparison of the predicted results of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings. The description herein describes specific embodiments, by way of illustration and not limitation, consistent with the principles of the present invention, which are described in sufficient detail to enable those skilled in the art to practice the invention, other embodiments may be utilized and the structure of elements may be changed and/or replaced without departing from the scope and spirit of the invention. The following detailed description is, therefore, not to be taken in a limiting sense.
As shown in fig. 1, according to a first aspect of the present invention, there is provided a method for identifying traffic index time series based on an autoregressive differential moving average-convolutional neural network, including:
0000, acquiring a traffic index original data set, and generating a traffic index time sequence;
and step 1000, converting the traffic index time sequence into a stable sequence, fitting an autoregressive differential moving average algorithm model according to a Bayesian information criterion matrix, and realizing traffic index prediction by using the autoregressive differential moving average algorithm model.
Further, the method for generating the traffic index time sequence in the step 0000 includes:
step 0200: and calculating the average traffic index at each moment according to the traffic index original data set.
The traffic index original data set is { X } ij ∈[X min ,X max ]|i∈[1,n 1 ]，j∈[1,n 2 ]，n 1 ，n 2 ∈N}，X min ,X max Respectively minimum and maximum values of traffic index, X ij For the traffic index at the j-th moment on the i-th day, the average traffic index at the j-th moment
Step 0400: selecting a first continuous time sequence, a first threshold k 1 ∈[l 1 *X，X max ]，l 1 Is a natural number, preferably 3,is the average of all traffic indexes.
Step 0600: and dividing the traffic index original data set according to the traffic index original data set and the first continuous time sequence.
If the number of times of the first continuous time sequence is n t Dividing the traffic index data of the ith day into n 0 For a period of time, preferably, n 0 Is n 2 And n t More preferably, the function of (c) is, representing a rounding down. For example, if the first continuous time series of traffic index data having 20 moments per day includes 5 moments, then n t ＝5，n 2 The daily traffic index data was divided into 4 time periods =20.
Step 0800: and generating a traffic index time sequence according to the division result of the step 0600. The specific method is that traffic index data in the same time period of each day are sequentially arranged to generate a traffic index time sequence.
For example, the traffic index raw data set comprises 10 days of data, the data of each day comprises 20 time periods, the traffic index data of each day is divided into 4 sections, then the data of the first time period from 1 st day to 10 th day are sequentially arranged to generate a traffic index time sequence of the first time period, and the traffic index time sequences of the second time period, the third time period and the fourth time period can be obtained by analogy, and the traffic index time sequences of the total 4 traffic index time sequences comprise 50 traffic index data.
The traffic index time sequence generation method has the advantages that the time with higher traffic index has higher probability and can be divided into a time period, so that the peak time period of traffic can be conveniently identified, and countermeasures can be actively carried out. In addition, through the division of time periods, the dimension is reduced for the subsequent data processing, and the operation efficiency is improved.
In particular, step 1000 further comprises:
step 1200: and (3) carrying out stability test on the traffic index time sequence, if the traffic index time sequence is a non-stable time sequence, turning to step 1400, otherwise turning to step 1600.
Step 1400: and (3) carrying out stabilization treatment on the traffic index time sequence, and turning to step 1200.
Step 1600: and selecting a Bayesian information criterion matrix to determine values of parameters p and q in each period according to the traffic index time sequence, constructing an ARIMA model, wherein p is the autoregressive term number, and q is the moving average term number.
Step 1800: fitting to the time series by the ARIMA model.
Further, the method of stationarity check in step 1200 is the ADF check.
In particular, the smoothing method in step 1400 is:
let { x } t T epsilon T is a group of time series data, B is a backward operator, namely Bx t ＝x t-1 RecordingCalled a difference operator, at which time,
generally there are
In particular, the ARIMA model in step 1600 includes an autoregressive operation and an average movement process.
Step 2000, generating a training traffic index time sequence and a test traffic index time sequence according to the traffic index original data set, extracting traffic index feature information according to the training traffic index time sequence, obtaining an optimal convolutional neural network model, integrating the traffic index feature information into a one-dimensional feature vector by using the optimal convolutional neural network model, determining the mode category of the one-dimensional feature vector according to a Softmax classifier, and further identifying the category to which the test traffic index time sequence belongs.
In step 2000, the CNN framework is formed by two groups of convolution pooling layers, where the convolution pooling layers include a convolution layer, a ReLU activation function, and a maximum pooling layer, where the convolution kernel of the convolution layer is 5*5, and the maximum pooling layer is 2×2.
According to a second aspect of the present invention, there is also provided a specific embodiment of a method for identifying a traffic index time series based on an autoregressive differential moving average-convolutional neural network:
the method for generating the traffic index time sequence in the step 0000 includes:
step 0200: and calculating the average traffic index at each moment according to the traffic index original data set.
Step 0400: a first continuous time sequence and a second continuous time sequence are selected.
Wherein the second threshold k 2 ∈(l 2 *X，l 1 *X)，l 2 Is a positive real number and l 2 1 or more, preferably, l 2 1.5.
Step 0600: and dividing the traffic index original data set according to the traffic index original data set, the first continuous time sequence and the second continuous time sequence.
Step 0600 further comprises:
step 0620, if the number of times of the first continuous time sequence is n t Dividing the traffic index data of the ith day into n 0 For a period of time, preferably, n 0 Is n 2 And n t More preferably, the function of (c) is, representing a rounding down.
Alternatively, if there are a plurality of first continuous time sequences, n t Is the average number of instants of the first consecutive time series.
Alternatively, ifThe remainder of the third threshold k is greater than 3 ，k 3 Is n t Is more preferably, ++>Then the traffic index data on the i-th day is divided into n 0 +1 time period, redundant data as nth 0 +1 time period data, otherwise divided into n 0 The redundant data is divided into the last time period data.
For example, if the first continuous time series of traffic index data having 23 moments per day includes 5 moments, then n t ＝5，n 2 =23, dividing the daily traffic index data into 5 time periods, the first four time periods including data of 5 moments, and the fifth time period including data of 3 moments; if there are 22 times per day, n t ＝5，n 2 =22, dividing the daily traffic index data into 4 time periods, the 4 th time period comprising 6 momentsData.
In step 0640, the traffic index raw data set is divided into a first data subset, a second data subset …, and a Z data subset, Z being a natural number.
Preferably, the traffic index raw data set is divided into monday data subsets, monday data subset … sunday data subsets.
The benefit of using this partitioning method is that the traffic index is generally relatively fixed with the cycle of the week, so partitioning the traffic index raw data set into monday data subsets, the monday data subset … sunday data subset, facilitates finding the traffic index change law within the week. In addition, the ARIMA model is good at short-time sequence prediction, and the calculation efficiency is improved by dividing the traffic index original data set.
Optionally, in step 0660, calculating an average traffic index x of a second continuous time sequence of the first data subset and the second data subset …, where the second continuous time sequence of the second data subset is the Z data subset, obtaining the first data subset, and calculating a second continuous time sequence feature vector z=α×x+β×y, where α, β is a weighting coefficient, where α+β=1, preferably α=0.27, and β=0.73, where the number of times y of the second continuous time sequence of the second data subset …, where the average traffic index is greater than the second threshold and less than the first threshold, is the second continuous time sequence feature vector y. The first data subset and the second data subset … are divided into a first mode subset and a second mode subset … and a Y mode subset according to the Z value, Y is a natural number and is less than or equal to 7.
Further, the specific method for dividing the first data subset, the second data subset … and the Z data subset into the first mode subset and the second mode subset … and the Y mode subset according to the Z value is as follows:
Preferably, z 1 、z 2 …z Y-1 、z Y Is z min And z max Wherein z is min And z max Respectively the minimum and maximum value of the z value. More preferably, the method further comprises the steps of,
optionally, the parameters of the second continuous time sequence that may be acquired further include peak size, valley size, presence or absence of a second peak, second peak-to-peak value, second peak duration, etc.
The meaning of further combining the first data subset and the second data subset … and the Z data subset is that the data with similar traffic index change trend are put together, so that the success rate of the subsequent neural network training is improved.
Step 0800: and generating a traffic index time sequence according to the division result of the step 0600.
For example, the traffic index raw data set includes 10 days of data, the data includes 30 time periods each day, step 0620 divides the data into 3 time periods, step 0660 divides the data into 2 modes, then the first time period data of the first mode is sequentially arranged to generate a first time period traffic index time sequence of the first mode, and so on to obtain a first time period traffic index time sequence of the second mode, …, and a third time period traffic index time sequence of the second mode, which includes 6 traffic index time sequences, each traffic index time sequence including 50 traffic index data.
According to a third aspect of the present invention, there is also provided a specific embodiment of a method for identifying a traffic index time series based on an autoregressive differential moving average-convolutional neural network:
step 0000, obtaining a traffic index original data set, wherein the traffic index original data set is a Beijing city traffic index of 2016, 11, 2017, 11 and 11 (without data of 2018, 11, 30 days) of three months, the time interval frequency is 15 minutes, and each day is recorded as 05:00-23: 00 traffic index, 73 values per day.
TABLE 1 Beijing urban traffic index schematic chart
Table 1 shows the partial traffic index for each time slot on each date of beijing. Traffic index time series data can show obvious differences at different times and different dates.
Step 0200: and calculating the average traffic index at each moment according to the traffic index original data set. For example, the average traffic index at time 05:00 is (1.0+0.8+ … +1.1+1.4)/89= 1.1078 (four significant digits, the following).
Step 0400: a first continuous time sequence and a second continuous time sequence are selected. Calculated that the average of all traffic indexes is 1.3273, taking l 1 =3, having k 1 ∈[3.9819，8.5]Preferably 4, then the first continuous time sequence is 17:30-19:45 (10 instants); l (L) 2 1.5, k 1 E (1.9909,3.9819), preferably 2, the second consecutive time series is 16:15-17:30 (7 instants).
In step 0620, the number of times of the first continuous time sequence is 10, 73 times are total per day, 73 divided by 10 is 7, and remainder is 3, preferablyThe daily traffic index data is divided into 7 time periods, the 7 th time period including 13 moments.
In step 0640, the traffic index raw data set is divided into a monday data subset, a monday data subset … sunday data subset.
In step 0660, an average traffic index x of a second continuous time sequence of the monday data subset and the sunday data subset … is calculated, the monday data subset is obtained, the number y of times when the average traffic index in the second continuous time sequence of the sunday data subset … is greater than the second threshold and less than the first threshold is calculated, and a second continuous time sequence feature vector z=α×x+β×y, α=0.27, β=0.73.
Taking the monday data subset as an example, the average traffic index x of 16:15-17:30 (7 times) of the monday data subset is 1.7382, the number y of times when the average traffic index is greater than 2 and less than 4 is 0, then the second continuous time sequence feature vector z of the monday data subset=0.27×1.7382+0.73×0= 0.4693, and the second continuous time sequence feature vector z of the monday data subset … is 2.9347, 2.8424, 2.9892, 4.6539, 3.8346, 5.7532 respectively.
From the z values, it can be seen that the z values for monday are the smallest, the z values for friday, are close, and the difference is significant, so the monday data is subset, the sunday data subset … sunday data subset is divided into a monday pattern subset, a mid-week pattern subset, a friday pattern subset, a Saturday pattern subset, and a sunday pattern subset.
TABLE 2 traffic index characterization
Alternatively, as shown in Table 2, the present invention divides traffic index patterns from the factors of peak size, valley size, presence or absence of a second peak, second peak-to-peak value, and second peak duration, etc., into: monday mode, mid-week mode, friday mode, saturday mode, and sunday mode.
Step 0800: the experimental data of the invention is 1 x 73-dimensional time series data, the data time span is larger, and according to the dividing method of the step 0660, the invention divides one day of data into 7 time periods: 05:00-07:15 (10 moments), 07:30-09:45 (10 moments), 10:00-12:15 (10 moments), 12:30-14:45 (10 moments), 15:00-17:15 (10 moments), 17:30-19:45 (10 moments), and 20:00-23:00 (13 moments). As shown in fig. 5 (a) -5 (g), the present invention obtains the corresponding index predicted values from the index data of 7 time periods through ARIMA algorithm, and then combines the predicted values of 7 time periods into a predicted index of one day according to time sequence. Taking the monday pattern data prediction as an example, the method comprises the following steps of using 2016, 11, 2017, 11, 6, 2017, 11, 13, etc predicting 9 days of data in total from 2017, 11, 20, 27, 5, and 12, and finally, comparing the precision with the actual value.
TABLE 3 ADF test values for time periods
Step 1200: and (3) carrying out stability test on the traffic index time series, and drawing an autocorrelation chart of each other according to each time series data predicted by the Monday mode data, as shown in fig. 6 (a) -6 (g). A root-by-root test (ADF) was performed for 7 time periods of monday mode, and the test results are shown in table 3.
Based on the time series plot, the autocorrelation plot, and the ADF detection values, 7 sets of time series data are found to be non-stationary data, and step 1400 is performed.
TABLE 4 ADF test values after differentiation
Step 1400: and (3) carrying out stabilization treatment on the traffic index time sequence, namely carrying out differential operation on seven groups of data, wherein the operation result is shown in fig. 7, and an autocorrelation chart and a partial autocorrelation chart are drawn, and as shown in fig. 8 and 9, the step 1200 is carried out.
TABLE 5 parameter p and parameter q values for each period
Step 1200: the traffic index time series was subjected to a stability test, and the test results are shown in table 4. According to the differential exponential distribution diagram, the autocorrelation diagram, the partial correlation diagram and the result of the ADF inspection table, the p value is smaller than 0.05, and it can be determined that the non-stationary time series data is converted into a stationary sequence through differential operation, and the step 1600 is shifted.
Step 1600: and selecting a Bayesian information criterion matrix to determine the values of parameters p and q in each period according to the traffic index time sequence, and constructing an ARIMA model as shown in a table 5. Where the parameters p and q are the order of the ARIMA model.
Step 1800: fitting (results retain 3 significant digits) to the time series was performed by the ARIMA model as shown in table 6.
TABLE 6 prediction results for each period
7 groups of time series prediction results are integrated to form a 1-73-dimensional traffic index prediction result of a whole day, and the prediction result is compared with an actual traffic index (the traffic index of 11 months 19 days in 2018: [1.3,1.2,1,1,1,1.1,1,1.5,2.2,3.5,5.3,6.5,7.1,7.7,7.7,7.6,7.2,6.8,5.9,6.2,5,4.5,4.1,3.7,3.3,2.9,2.7,2.2,2.1,2.1,1.9,1.8,1.7,1.7,1.7,1.8,2,2.4,2.5,2.6,2.6,2.6, 2.5,2.5,2.5,2.5,2.7,2.8,3,3.4,4,5.8,6.5,7.2,7.3,7.1,6.7,6,4.6,3.5,2.9,2.4,2.1,2.1,1.9,2,1.7,1.8,1.7,1.5,1.6,1.5,1.5 ]) as shown in fig. 10, the overall change of the prediction index is consistent with the actual situation, and the prediction result accords with the Monday mode characteristics.
The invention adopts continuous data of 11 months in 3 years, takes 2016 years 11 months, 2017 years 11 months and 2018 years 11 months 1-15 days as training sets, takes 2018 years 11 months 16-29 days as test sets, marks according to the real category, expressed by 0 and 1, for example, (1, 0) represents a monday pattern, (0, 1, 0) represents a mid-week pattern, (0, 1, 0) represents a friday mode, (0, 1, 0) represents a Saturday mode, and (0,0,0,0,1) represents a sunday mode. The daily data is supplemented with a value of 0 at the first digit of the time series data to form a 10 x 10 dimensional matrix, and labels are marked at the tail of the matrix to distinguish the daily modes, so that an input layer of the convolutional neural network is formed. For example, the traffic index of 2018, 11, and 19 may be expressed as:
The invention adopts a convolutional neural network as a deep learning model, and realizes the experimental process in TensorFlow by using a CPU version under Python language. Even-numbered convolution kernels result in loss of image boundary information and shift of position information, and thus, convolution kernels are typically present in an odd-numbered form. The size of the convolution kernel is determined by the size of the input layer, common dimensions include: 3*3, 5*5 and 7*7, an increase in the convolution kernel side length can lead to a steep increase in the computation effort. In the invention, the receptive field range of the convolution kernel with the size of 7*7 is too large, the calculated amount is large, and the difference of the calculated results is not obvious; 5*5 is larger than 3*3 in the range of the convolution kernel receptive field, so that more characteristic information can be extracted, and in conclusion, the convolution kernel with the size of 5*5 is selected. Based on the operational characteristics of the mean pooling layer and the maximum pooling layer, the mean pooling layer is more prone to preserving background information, while the maximum pooling layer is better at extracting texture characteristics of the neighborhood. The traffic index time sequence analysis method and the traffic index time sequence analysis system mainly divide the modes according to the change rule of the traffic index time sequence, and the value with the most obvious characteristic difference can effectively reflect the change rule of the traffic index, so that the background information characteristics of the traffic index time sequence analysis method and the traffic index time sequence analysis system can be ignored, and the characteristic information of the traffic index can be characterized by searching the most value in the neighborhood. The operation mode of the maximum pooling layer just meets the experimental requirement. The invention uses 2 x 2 maximum pooling layer to realize local feature compression and extraction, and the pooling layer with larger size can cause loss of layer information and reduce robustness of extracted features.
The overall framework adopted by the mode classification part of the invention consists of CNN and Softmax classifiers. The CNN framework consists of two groups of convolution layers, a ReLU activation function and a maximum pooling layer; the Softmax classifier consists of a Softmax function classification layer; the full connectivity layer connects the CNN and Softmax classifiers. Because the traffic index data sample size is smaller, in order to avoid the phenomenon of overfitting, the experiment adopts two groups of convolution kernels and a pooling layer to extract the characteristics. Firstly, performing feature extraction on a 10 x 10 dimensional matrix of an input layer by utilizing a 5*5 convolution kernel, and performing nonlinear conversion by adopting a ReLU activation function after convolution treatment, thereby solving the problems of gradient disappearance, buffer occurrence of a fitting phenomenon and the like; the results after nonlinear conversion of the activation function are processed with a maximum pooling of 2 x 2. The second group of convolution pooling processing is the same as the first, the output layer after the two processing realizes multidimensional data unidimensional through the full-connection layer, the multidimensional data unidimensional is used as an input layer of Softmax function logistic regression, the probability of each mode is output, and the mode corresponding to the maximum probability is taken as the category of the time sequence data mode of the day.
The test set data identification results are shown in table 7. The Monday pattern (11 month 19 day, 11 month 22 day, 11 month 26 day, 11 month 27 day), the Sunday pattern (11 month 20 day, 11 month 21 day, 11 month 28 day, 11 month 29 day), the Sunday pattern (11 month 16 day, 11 month 23 day), the Saturday pattern (11 month 17 day, 11 month 24 day) and the Sunday pattern (11 month 18 day, 11 month 25 day) can be distinguished by the convolution algorithm. However, the classification accuracy of the test set is up to 85.7% when the test set is mistakenly divided into a week mode from the 11 month 22 day and the 11 month 27 day.
TABLE 7 test set experimental results
And similarly, the prediction result is processed to form a 10 x 10 dimensional matrix as an input layer, and the matrix is classified by using a convolutional neural network to obtain a Monday mode which accords with the actual situation. Experimental results show that the ARIMA-CNN model constructed by the invention can be used for realizing the prediction and classification of traffic index time sequence data; and meanwhile, the classification result is used as a judgment standard, a mode threshold line corresponding to the corresponding category and the prediction result are called for judgment, and the abnormal point exceeding the threshold value is used as a key attention point. Traffic departments can conduct traffic control according to the abnormal value as the basis of auxiliary decision, and the risk of urban traffic jam is reduced.
In order to make up for the defect of the traditional traffic jam analysis capability, the invention provides an ARIMA-CNN model for realizing traffic index prediction and traffic index type classification, and searching for abnormal conditions to be generated in traffic so as to achieve the purpose of prediction and early warning. According to the invention, the traffic index data of Beijing city in 2016-2018 is taken as a research object, the traffic index data of Beijing city is checked to be a non-stable sequence by using an ADF, the non-stable sequence is converted into a stable sequence through differential operation, and the traffic index is predicted according to ARIMA fitted by a BIC matrix; and extracting the time sequence mode characteristics of the traffic index by using a CNN algorithm, so as to realize the mode identification of the predicted traffic index. The experimental result shows that the ARIMA-CNN model accurately predicts the Beijing city traffic index data and identifies the prediction result as a Monday mode which is in accordance with the actual situation. The model has the capability of accurately predicting time sequence data and identifying modes, and can assist traffic management related departments to conduct traffic dispersion in advance so as to relieve urban traffic pressure.
In addition, in future research, the influence of factors such as traffic emergency, bad weather, major events and the like on model weight can be considered, and the prediction and decision are synchronously carried out by combining the internet of things technology, so that the risk of urban traffic jam is reduced as much as possible.
Furthermore, other implementations of the invention will be apparent to those skilled in the art from consideration of the specification of the invention disclosed. Embodiments and/or aspects of embodiments may be used in the systems and methods of the present invention alone or in any combination. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.

Claims (6)

1. A method of identifying a time series of traffic indexes based on an autoregressive differential moving average-convolutional neural network, comprising:
step 0200: calculating the average traffic index at each moment according to the traffic index original data set, wherein the traffic index original data set is that，/>，/>Respectively, minimum and maximum of traffic index, < ->Is->Day->Traffic index at each moment, then +.>Average traffic index at each moment；
Step 0400: selecting a first continuous time sequence and or a second continuous time sequence, wherein the first continuous time sequence is a continuous time sequence with the average traffic index being larger than a first threshold value, and the second continuous time sequence is a continuous time sequence with the average traffic index being larger than a second threshold value and smaller than the first threshold value; first threshold value ，/>Is natural number (i.e.)>Average over all traffic indexes; second threshold->，Is positive real number and +.>；
Step 0600: dividing the traffic index raw data set according to the traffic index raw data set and the first continuous time sequence:
if the number of time instances of the first continuous time sequence isWill be->Traffic index data of dayIs divided into->Time period(s)>Is->And->Function of->，/>Representing a downward rounding; if->The remainder of the first signal is greater than a third threshold，/>Then->Traffic index data of day is divided into +.>+1 time period, redundant data as +.>+1 time period data, otherwise split into +.>The redundant data are divided into the data of the last time period;
step 0640, dividing the traffic index original data set into a first data subset, a second data subset … and a Z data subset;
step 0660, calculating an average traffic index x of a second continuous time sequence of the first data subset, the second data subset … and the Z data subset, obtaining the first data subset, calculating a number y of times when the average traffic index is greater than a second threshold and less than the first threshold in the second continuous time sequence of the second data subset … and the Z data subset, and calculating a second continuous time sequence feature vector ，/>For the weighting factor>The method comprises the steps of carrying out a first treatment on the surface of the Dividing the first data subset, the second data subset … and the Z data subset into a first mode subset, a second mode subset … and a Y mode subset according to the Z value, wherein Y is a natural number；
；
、/> 、/>Is->And->Wherein>And->Minimum and maximum value of z value, respectively, +.>；
Step 0800: generating a traffic index time sequence according to the division result of the step 0660;
step 1000, converting the traffic index time sequence into a stable sequence, fitting an autoregressive differential moving average algorithm model according to a Bayesian information criterion matrix, and realizing traffic index prediction by using the autoregressive differential moving average algorithm model;
step 2000, generating a training traffic index time sequence and a test traffic index time sequence according to the traffic index original data set, extracting traffic index feature information according to the training traffic index time sequence, obtaining an optimal convolutional neural network model, integrating the traffic index feature information into a one-dimensional feature vector by using the optimal convolutional neural network model, determining the mode category of the one-dimensional feature vector according to a Softmax classifier, and further identifying the category to which the test traffic index time sequence belongs.
2. The method of identifying traffic index time series based on an autoregressive differential moving average-convolutional neural network of claim 1, wherein step 1000 further comprises:
step 1200: performing stability test on the traffic index time sequence, if the traffic index time sequence is a non-stable time sequence, turning to step 1400, otherwise turning to step 1600;
step 1400: performing stabilization treatment on the traffic index time sequence, and turning to step 1200;
step 1600: according to the traffic index time sequence, selecting a Bayesian information criterion matrix to determine values of parameters p and q in each period, and constructing an autoregressive differential moving average algorithm model, wherein p is the number of autoregressive terms, and q is the number of moving average terms;
step 1800: fitting the time sequence through the autoregressive differential moving average algorithm model.
3. The method for identifying traffic index time series based on autoregressive differential moving average-convolutional neural network according to claim 2, wherein the smoothing processing method in step 1400 is as follows:
for a set of time series data, B is the shift operator,>record->=1-B, there is
(1)
Then
(2)
And (3) performing differential operation on the non-stationary time series data according to the formula (1) and the formula (2) to obtain a stationary sequence.
4. A method of identifying traffic index time series based on an autoregressive differential moving average-convolutional neural network according to any one of claims 1-3, wherein the convolutional neural network framework in step 2000 is composed of two sets of convolutional pooling layers including a convolutional layer, a ReLU activation function, and a max pooling layer.
5. A method of identifying traffic index time series based on an autoregressive differential moving average-convolutional neural network according to any one of claims 1-3, wherein the convolution kernel of the convolutional layer is 55。
6. A method of identifying traffic index time series based on an autoregressive differential moving average-convolutional neural network according to any one of claims 1-3, wherein the maximum pooling layer is 22。
CN202011231975.7A 2020-11-06 2020-11-06 Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network Active CN112330158B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011231975.7A CN112330158B (en) 2020-11-06 2020-11-06 Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011231975.7A CN112330158B (en) 2020-11-06 2020-11-06 Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network

Publications (2)

Publication Number Publication Date
CN112330158A CN112330158A (en) 2021-02-05
CN112330158B true CN112330158B (en) 2023-10-13

Family

ID=74317130

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011231975.7A Active CN112330158B (en) 2020-11-06 2020-11-06 Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network

Country Status (1)

Country Link
CN (1) CN112330158B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113159128B (en) * 2021-03-24 2023-03-31 西北大学 Method for judging traffic accidents based on classification algorithm for rapidly extracting shapelets
CN113706867B (en) * 2021-08-27 2022-08-30 东北大学 Method for evaluating traffic anomaly probability of adjacent edge equipment nodes of highway
CN114511767B (en) * 2022-02-11 2023-12-01 电子科技大学 Rapid state prediction method for time sequence diagram data

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107742420A (en) * 2017-09-22 2018-02-27 北京交通大学 It is a kind of to be used for the method that road traffic flow is predicted under emergent traffic incident
CN108877225A (en) * 2018-08-24 2018-11-23 交通运输部规划研究院 Magnitude of traffic flow index determines method and device
CN109544911A (en) * 2018-10-30 2019-03-29 中山大学 A kind of city road net traffic state prediction technique based on LSTM-CNN
CN109887290A (en) * 2019-03-30 2019-06-14 西安电子科技大学 Traffic flow forecasting method based on equilibrium index exponential smoothing and stack self-encoding encoder
CN110390425A (en) * 2019-06-20 2019-10-29 阿里巴巴集团控股有限公司 Prediction technique and device
CN110517488A (en) * 2019-08-19 2019-11-29 南京理工大学 The Short-time Traffic Flow Forecasting Methods with Recognition with Recurrent Neural Network are decomposed based on timing
CN111339978A (en) * 2020-03-02 2020-06-26 北京建筑大学 Method for recognizing traffic index time series mode by using convolutional neural network model

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107742420A (en) * 2017-09-22 2018-02-27 北京交通大学 It is a kind of to be used for the method that road traffic flow is predicted under emergent traffic incident
CN108877225A (en) * 2018-08-24 2018-11-23 交通运输部规划研究院 Magnitude of traffic flow index determines method and device
CN109544911A (en) * 2018-10-30 2019-03-29 中山大学 A kind of city road net traffic state prediction technique based on LSTM-CNN
CN109887290A (en) * 2019-03-30 2019-06-14 西安电子科技大学 Traffic flow forecasting method based on equilibrium index exponential smoothing and stack self-encoding encoder
CN110390425A (en) * 2019-06-20 2019-10-29 阿里巴巴集团控股有限公司 Prediction technique and device
CN110517488A (en) * 2019-08-19 2019-11-29 南京理工大学 The Short-time Traffic Flow Forecasting Methods with Recognition with Recurrent Neural Network are decomposed based on timing
CN111339978A (en) * 2020-03-02 2020-06-26 北京建筑大学 Method for recognizing traffic index time series mode by using convolutional neural network model

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
基于AＲIMA 和小波神经网络组合模型的交通流预测;成云 等;计算机技术与发展;第27卷(第1期);第169-172页 *
基于改进模式序列匹配的交通运行指数预测研究;郭小刚， 张健钦， 卢剑， 陆浩， 李卓航;北京建筑大学学报;第35卷(第4期);第20-27页 *
时空序列预测方法综述;黎维，陶蔚，周星宇，潘志松;计算机应用研究;第37卷(第10期);第2281-2286页 *

Also Published As

Publication number Publication date
CN112330158A (en) 2021-02-05

Similar Documents

Publication Publication Date Title
CN112330158B (en) Method for identifying traffic index time sequence based on autoregressive differential moving average-convolution neural network
CN111242208B (en) Point cloud classification method, segmentation method and related equipment
Gamboa Deep learning for time-series analysis
Dudek Neural networks for pattern-based short-term load forecasting: A comparative study
CN114220271B (en) Traffic flow prediction method, equipment and storage medium based on dynamic space-time diagram convolution circulation network
CN109993100B (en) Method for realizing facial expression recognition based on deep feature clustering
Lian et al. A novel data-driven tropical cyclone track prediction model based on CNN and GRU with multi-dimensional feature selection
Yang et al. A deep multiscale pyramid network enhanced with spatial–spectral residual attention for hyperspectral image change detection
CN112270355A (en) Active safety prediction method based on big data technology and SAE-GRU
CN115410088B (en) Hyperspectral image field self-adaption method based on virtual classifier
Yang et al. CEMA-LSTM: Enhancing contextual feature correlation for radar extrapolation using fine-grained echo datasets
Tang et al. Missing traffic data imputation considering approximate intervals: A hybrid structure integrating adaptive network-based inference and fuzzy rough set
Xu et al. A novel hyperspectral image clustering method with context-aware unsupervised discriminative extreme learning machine
Ying et al. License plate detection and localization in complex scenes based on deep learning
Dangut et al. Rescaled-LSTM for predicting aircraft component replacement under imbalanced dataset constraint
Asadi et al. Clustering of time series data with prior geographical information
CN113392876B (en) Small sample image classification method based on graph neural network
Savyanavar et al. Star Galaxy Classification Using Machine Learning Algorithms and Deep Learning
Berkani et al. Spatio-temporal forecasting: A survey of data-driven models using exogenous data
Patel et al. Regularized cnn for traffic sign recognition
Cao et al. A multi-label classification method for vehicle video
Mohanty et al. Fraud detection using outlier analysis: a survey
Li Introduction and Perspective of Deep Metric Learning
Wang et al. CNN Network for Head Detection with Depth Images in cyber-physical systems
Chaozhi et al. Financial time series prediction model based recurrent 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
CB03 Change of inventor or designer information

Inventor after: Zhang Xuedong

Inventor after: Lu Jian

Inventor after: Zhang Jianqin

Inventor after: Xu Zhijie

Inventor after: Wang Jiachuan

Inventor after: Shi Ruixuan

Inventor before: Zhang Xuedong

Inventor before: Lu Jian

Inventor before: Zhang Jianqin

Inventor before: Xu Zhijie

Inventor before: Wang Jiachuan

Inventor before: Shi Ruixuan

CB03 Change of inventor or designer information