CN111882114B - A short-term traffic flow prediction model construction method and prediction method - Google Patents

Abstract

The application discloses a short-time traffic flow prediction model construction method and a short-time traffic flow prediction method. The influence on the prediction result of the short-time traffic flow is obvious aiming at the training set selection of the neural network, the K-Means clustering algorithm is adopted to cluster the historical short-time traffic flow data, the prediction is carried out in a targeted manner, and the accuracy of the prediction result is improved.

Description

A short-term traffic flow prediction model construction method and prediction method

Technical field

The invention belongs to the field of intelligent transportation systems, and specifically relates to a short-term traffic flow prediction model construction method and a prediction method.

Background technique

The intelligent transportation system applies modern technological achievements to transportation planning and management, and establishes an intelligent transportation management system. Accuracy, efficiency, and real-time are the significant advantages of the intelligent transportation system. Through the feedback of the intelligent transportation system, relevant departments can promptly Master the traffic conditions in areas of concern to effectively guide traffic operations, reduce or even avoid the occurrence of traffic congestion, and provide information support for urban road planning, energy conservation and emission reduction.

With the rise of artificial intelligence research, neural network models have achieved remarkable results in short-term traffic flow prediction. However, due to factors such as different road conditions, different weather, and different dates, which have a great impact on the time series distribution of short-term traffic flow, the existing neural network model cannot accurately predict traffic flow.

Contents of the invention

In view of the shortcomings in the existing technology, the purpose of the present invention is to provide a short-term traffic flow prediction method based on K-Means clustering and GRU network to solve the technical problem that the existing technology cannot accurately predict traffic flow.

In order to solve the above technical problems, this application adopts the following technical solutions to achieve:

A short-term traffic flow prediction model construction method includes the following steps:

Step 1: Obtain traffic flow data within a period of time to obtain a traffic flow data set; preprocess the obtained traffic flow data set to obtain a preprocessed traffic flow data set;

Step 2: Cluster the data in the preprocessed traffic flow data set obtained in step 1 using the K-Means algorithm to obtain short-term traffic flow pattern categories, and establish a short-term traffic flow pattern library;

Step 3: Determine the prediction date, obtain the short-term traffic flow data for the first N moments of the prediction date, determine the feature vectors of the N time short-term traffic flow patterns in sequence, and use the classification method to determine the short-term traffic flow patterns corresponding to the N time feature vectors. Time traffic flow pattern, select the traffic flow data corresponding to the traffic pattern with the highest frequency of short-term traffic flow patterns corresponding to N moments to form the training data set;

Step 4: Use the GRU neural network model to train the training data set formed in step 3. After the training is completed, the short-term traffic flow prediction model is obtained;

The GRU neural network model adopts a stack structure, including an input layer, a first GRU unit layer, a second GRU unit layer and a third GRU unit layer that are connected in sequence, and the third GRU unit layer is connected behind to prevent over-simulation. combined dropout layer.

Specifically, the preprocessing described in step 1 includes data deletion, data interpolation, data denoising and normalization.

Specifically, in step 2, the short-term traffic flow data obtained in step 1 is clustered using the K-Means algorithm to obtain short-term traffic flow categories. The composition of the short-term traffic flow category pattern library specifically includes: selecting K points as initial Clustering is performed using the cluster center, where K ranges from 2 to 7.

Specifically, the parameter settings of the GRU network described in step 3 include: prediction step size 8, number of hidden layer neurons 12, learning rate 0.02, and number of iterations 800.

A short-term traffic flow prediction method, the method includes the following steps:

Step 1. Obtain the traffic flow data set to be predicted and perform preprocessing to obtain the preprocessed traffic flow data set to be predicted;

Step 2: Input the preprocessed traffic flow data set to be predicted into the short-term traffic flow prediction model obtained by the short-term traffic flow prediction model construction method according to any one of claims 1 to 5, Obtain the short-term traffic flow category to be predicted.

Compared with the existing technology, the beneficial technical effects of the present invention are:

This invention combines the K-Means clustering method with the GRU neural network. The training set selection of the neural network has a significant impact on the prediction results of short-term traffic flow. The K-Means clustering algorithm is used to cluster historical short-term traffic flow data. Class, targeted predictions are made, improving the accuracy of prediction results.

Description of the drawings

Figure 1 is a program flow chart of the present invention;

Figure 2 is a K-means clustering result diagram obtained in Example 1 of the present invention;

Figure 3 GRU prediction model flow chart.

Figure 4 is a short-term traffic flow prediction result diagram obtained in Embodiment 1 of the present invention, in which the solid line represents the actual short-term traffic flow, and the dotted line represents the predicted value;

Figure 5 is a comparison chart of the prediction errors of the KMeans-GRU model, the traditional GRU network model, the ARIMA model and the SAEs model.

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments of the description.

Detailed ways

The overall technical concept and technical principle of the present invention are: cluster the traffic flow data within a period of time, fully extract the data information and input it into the GRU neural network, and use the short-term traffic flow pattern category as the output to train the GRU neural network model , after the training is completed, the short-term traffic flow prediction model is obtained to achieve accurate prediction of short-term traffic flow.

The definitions or conceptual connotations involved in the present invention are explained below:

The Euclidean metric (also called Euclidean distance) is a commonly used distance definition, which refers to the real distance between two points in m-dimensional space, or the natural length of a vector (that is, the distance from the point to the origin). The Euclidean distance in two- and three-dimensional space is the actual distance between two points.

Two-norm: refers to the 2-norm of matrix A, which is the square root value of the largest eigenvalue of the product of the transposed conjugate matrix of A and matrix A. It refers to the straight-line distance between two vector matrices in space.

In order to make the purpose and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the drawings and examples, and the advantages of the present invention will be reflected through the analysis of comparative examples. It should be understood that the specific embodiments described here are only used to explain the present invention and are not intended to limit the present invention.

The principle of K-means clustering algorithm is: with k as parameter, n objects are divided into k clusters, so that the similarity within the cluster is high and the similarity between clusters is low. The processing process of the K-means clustering algorithm is as follows: first, k objects are randomly selected as centroids, and each object initially represents the mean or center of a cluster; for each remaining object, according to its relationship with the center of each cluster distance, assigning it to the nearest cluster, and then the algorithm iterates between the data assignment step and the centroid update step until a stopping criterion is met (i.e., no data points change, or the sum of cluster distances is minimized, or some maximum number of iterations is reached ).

KNN classification method: The core idea of the KNN model is very simple. It measures the Euclidean distance between each test set sample point and each sample in the training set, and then takes the K points with the closest Euclidean distance (k can be manually drawn (The determination of K will affect the algorithm results), and count the category frequencies of the K training set sample points, and convert the category with the highest frequency into the predicted category of the test sample point.

The present invention also uses multiple model evaluation index comparisons to evaluate the quality of the prediction results, including square percentage error (MAPE), square absolute error (MAE), and root mean square error (RMSE).

For the squared percentage error, squared absolute error and root mean square error, the smaller the error, the better the prediction effect.

Example:

This embodiment

Step 1: Obtain historical traffic flow data within a certain period of time and short-term, and then perform data filling, deletion, or data denoising and normalization related preprocessing to obtain a preprocessed traffic flow data set.

As a specific implementation mode of the present invention, traffic flow data can be obtained by directly downloading the official database or independently detecting it. The historical traffic flow data in this example is downloaded from the US PeMS database, including traffic flow data collected every 5 minutes from January 1, 2018 to January 31, 2018, with a total of 8928 pieces of data.

Step 11: For the input short-term traffic flow historical data, first determine whether it is a null value. If it is a null value, perform a deletion operation. If the proportion of null values in the traffic flow data in a day is greater than 20%, delete the filled traffic flow. data;

Step 12: If it is not a null value, determine whether it is abnormal data according to equation (1).

Among them, q represents the short-term traffic flow, C represents the maximum road capacity, T represents the traffic data collection time interval, and f _c represents the correction coefficient. If it is abnormal data, delete it first and then use the mean of the previous value and the next value for interpolation.

Step 13: The min-max normalization method is used to process the short-term traffic flow data, as shown in Equation (2):

Among them, X _i represents the i-th sample value, X _i ′ represents the normalized value of the i-th sample, X _max represents the maximum value in the sample, and X _min represents the minimum value in the sample.

Step 2: Cluster the data in the preprocessed traffic flow data set obtained in step 1 using the K-Means algorithm to obtain short-term traffic flow pattern categories, and establish a short-term traffic flow pattern library.

Step 21: First select K points as the initial clustering center (the value range of K is generally [2-7]). In this example, K=3 is selected.

Step 22: Distribute the data in the preprocessed traffic flow dataset.

Each centroid defines a cluster. In this step, the distance between data points is calculated using the Euclidean distance based on the 2 norm, and each data point is assigned to the centroid closest to it, as shown in Equation (3). If c _i is a centroid set in set C, then all points x in the data set will be assigned to a cluster based on centroid c _i .

Among them, dist() is the Euclidean distance under 2 norm.

Step 23: Centroid Update

In this step, the centroid is updated by calculating the mean of all data points, as shown in Equation (4):

Steps 1 and 2 are iterated continuously until no data point changes the cluster, and the sum of the prime distances from each data point in the cluster reaches the minimum, or the maximum number of iterations is reached, as shown in Figure 2.

Step 3: Select the short-term traffic flow at 24 time points (first 2 hours) of the date to be predicted as the state vector, and compare the short-term traffic flow at these 24 time points with the short-term traffic flow pattern library obtained in step 2. The Euclidean distance of short-term traffic flow is used as a measure of data similarity.

Step 31: Calculate the Euclidean distance between all data points in the interval from 0:00 on January 1, 2018 to 24:00 on January 30, 2018, and the data 24 times before 0:00 on January 31, 2018.

Step 32: Arrange the calculated Euclidean distances at the 24 moments before 0:00 on January 31, 2018 from small to large.

Step 33: Determine and select the N sample points with the smallest distance from the unknown sample (the selection range is [3-7] in the implementation example).

Step 34: Calculate the frequency of occurrence of the categories to which the selected N points belong. In this embodiment, N=5 is selected.

Step 35: Use the category with the highest frequency as the category of the current sample point.

Step 4: Select the data in the category most similar to the traffic flow pattern on January 31, 2018 as the training data of the GRU network, and then use the designed GRU network to make predictions.

The GRU network includes 1 layer of input layer, 3 layers of GRU unit layer, 1 layer of dropout layer and 1 layer of output layer. The structure is shown in Figure 3. The GRU prediction model structure adopts a stack structure, and the multi-layer architecture can express the data at a deeper level. As shown in Figure 4.5, the three input parameters none, 12, and 1 in gru_1_input:InputLayer represent the number of samples, time steps and variable dimensions of the input data of the network's input layer, and the parameters of the output of the previous layer. It must be the same as the parameter dimension of the next layer input. The Dropout layer (parameter value range is 0-0.2) connected after the 3-layer GRU network is to reduce the over-fitting of the GRU network. The final Dense (fully connected) layer converts the multi-dimensional data output by the GRU into one-dimensional output.

The hyperparameter settings of the GRU network are shown in Table 1.

Table 1 GRU network hyperparameter settings

Figure 4 shows the prediction results using this method, in which the blue solid line represents the actual short-term traffic flow and the orange solid line represents the predicted value.

Figure 5 shows the error comparison chart of KMeans-GRU model, traditional GRU network model, ARIMA model and SAEs model.

Table 2: Error comparison between the prediction results of the KMeans-GRU prediction model and the traditional GRU network model, ARIMA model and SAEs model:

Table 2 Comparison table of evaluation indicators of models

Claims (3)

1. The short-time traffic flow prediction model construction method is characterized by comprising the following steps of:

step 1, acquiring traffic flow data in a period of time to obtain a traffic flow data set; preprocessing the obtained traffic flow data set to obtain a preprocessed traffic flow data set;

step 1.1: for the input short-time traffic flow historical data, firstly judging whether the data is null, if so, executing deleting operation, and if the null ratio of the traffic flow data in one day is more than 20%, deleting the whole filled traffic flow data;

step 1.2: if the value is not null, judging whether the value is abnormal data according to the formula (1);

wherein q represents short-time traffic flow, C represents maximum road traffic capacity, T represents flow data acquisition time interval, f _c Representing the correction coefficient;

if the data is abnormal data, deleting the data firstly and then interpolating by using the average value of the previous value and the next value;

step 1.3: the min-max normalization method processes the short-time traffic flow data as shown in the formula (2):

wherein X is _i Represents the i-th sample value, X _i ' represents the normalized value of the ith sample, X _max Represents the maximum value in the sample, X _min Representing the minimum value in the sample;

step 2, clustering the data in the preprocessed traffic flow data set obtained in the step 1 by a K-Means algorithm to obtain short-time traffic flow mode types, and establishing a short-time traffic flow mode library;

step 2.1: firstly, K points are selected as initial clustering centers, and the value range of K is 2-7;

step 2.2: distributing data in the preprocessed traffic flow data set;

each centroid defines a cluster; in this step, the distance between data points is calculated using euclidean distance based on 2 norms, and each data point is assigned to the centroid nearest to it, as shown in the following formula (3); if c _i Is a centroid set in set C, then points x in the dataset will all be assigned to a centroid-based C _i Is in a cluster of classes (1);

wherein dist () is a euclidean distance at 2 norms;

step 2.3: centroid update

In this step, the centroid is updated by calculating the mean of all data points, as shown in equation (4):

iterating the step 1 and the step 2 continuously until no data point changes the class cluster, and the sum of the distances from each data point in the cluster to the quality is minimum or the maximum iteration times are reached;

step 3, determining a prediction date, acquiring short-time traffic flow data of the first N moments of the prediction date, sequentially determining feature vectors of the N moment short-time traffic flow modes, respectively determining short-time traffic flow modes corresponding to the N moment feature vectors by using a classification method, and selecting traffic flow data corresponding to a traffic flow mode with highest occurrence frequency of the short-time traffic flow modes corresponding to the N moments to form a training data set;

step 4, training the training data set formed in the step 3 by using the GRU neural network model, and obtaining a short-time traffic flow prediction model after training is completed;

the GRU neural network model adopts a stack structure and comprises an input layer, a first GRU unit layer, a second GRU unit layer and a third GRU unit layer which are sequentially connected, and a dropout layer for preventing overfitting is connected behind the third GRU unit layer;

the parameter setting of the GRU network model comprises the following steps: prediction step size 8, hidden layer neuron number 12, learning rate 0.02 and iteration number 800.

2. The short-term traffic flow prediction model construction method according to claim 1, wherein the preprocessing in step 1 includes data deletion, data interpolation, data denoising and normalization.

3. A short-term traffic flow prediction method, characterized in that the method comprises the steps of:

step 1, obtaining a traffic flow data set to be predicted and preprocessing the traffic flow data set to obtain a preprocessed traffic flow data set to be predicted;

and 2, inputting the preprocessed traffic flow data set to be predicted into the short-time traffic flow prediction model obtained by the short-time traffic flow prediction model construction method according to claim 1 or 2, and obtaining the class of the short-time traffic flow to be predicted.