/traffic_flow_code

short term traffic forecasting using deep learning architectures

Primary LanguageMATLAB

Traffic Flow Forecasting using Deep learning and a Probabilistic Loss Function

Welcome. This repository contains the data and scripts comprising the 'Traffic Flow Forecasting using Deep learning and a Probabilistic Loss Function' (TFF with PLF). This work is a real-time automatic system for predicting short-term traffic speeds within freeways.

Included are the tools to allow you to easily run the code.

This readme is a brief overview and contains details for setting up and running TFF with PLF. Please refer to the following:

Running TFF with PLF


Initial requirements

  1. To the code, the environment needed is Matlab. So you need to install Matlab.
  2. To run TFF with PLF using Convolutional Neural Networks (CNN), the MatConvNet Toolbox has already been downloaded and compiled for you. So, you don't need to install and compile MatConvNet. But, if you have your own version of MatConvNet, you can do so by replacing the MatConvNet folder within 'traffic_flow_code_CNN/' directory by your own.

Usage

There are several use cases for TFF with PLF:
  1. You can train and test Convolutional Neural Networks (CNNs) by going into 'traffic_flow_code_CNN/' directory and running the file 'proj_traffic_flow_prediction_10wStr.m'.

You can choose to train and test the CNN (see example 1 below ):

  • among different prediction time interval (5-, 10-, 15-, and 20-min forecasting)
  • with either the L2 Loss function (lambda=0) or the probabilistic loss function (lambda=1)
  • on one of the following networks (freeways): H101_North_D7 / I5_North_D7 / I5_South_D7 / I5_North_D11 / I450_North_D7 / I210_West_D7

  1. You can choose to test the CNN using a an available CNN which was already trained using traffic data of 'H101_North_D7' freeway (see example 3 below) If you wich to test using a CNN trained on another network (freeway), you first train the CNN using the desired network (use case 1) then call it for testing (use case 2).

  2. You can train and test Deep Belief Networks (DBNs) by going into 'traffic_flow_code_DBN/examples/' directory and running the file 'proj_traffic_flow_prediction_DBN.m'

  3. You can choose to test the DBN using a an available DBN which was already trained using traffic data of 'H101_North_D7' freeway (see example 2 below). If you wich to test using a DBN trained on another network (freeway), you first train the CNN using the desired network (use case 3) then call it for testing (use case 4).

PS. If you want to compare CNN and DBN performances with existant methods, you can try :

  • the Support Vector Machin (SVM) by going into 'traffic_flow_code_SVM/' directory and running the file 'proj_traffic_flow_prediction_SVM.m'
  • the ARIMA by y going into 'traffic_flow_code_ARIMA/' directory and running the file 'proj_traffic_flow_prediction_ARIMA.m'
  • the HW-ExpS by y going into 'traffic_flow_code_HW-exp/' directory and running the file 'main.m'

Examples for training and/or testing our models :

1. Exampe of training and testing the CNN model:

In this example, we want to predict the exact speed at 15-min forecasting (i.e., using regression) by training the CNN based on the probabilistic loss function.

The default data is the 'US101-North District 7' freeway (i.e., H101-North-D7) from september 1 to september 30 (2017) from 6AM to 8:55PM. 3/4th of the data is used for training and 1/4th is for testing.

The measure of performance is RMSE which gives the error in miles/hour.

To do so, follow these steps:

  1. run proj_traffic_flow_prediction_10wStr.m
  2. select the following:

  • Please forecasting for which you wish to predict speed (1)for 5-min, (2)for 10-min, (3)...): 1

  • Please select the prediction type: (c)classification / (r)regression r

  • Please enter the k-fold (k-1 for training & 1 for testing)_(0 for testing): 4

  • Please select the number of days (15, 21, 27, 30 or 59): 30

  • Please enter the loss (0)L2 loss, (1)P loss: 1

  • Please select the freeway: H101_North_D7 / H101_South_D7 / I5_North_D7 / I5_South_D7 / I5_North_D11 / I450_North_D7 / I210_West_D7 H101_North_D7

The code will:

  • display a plot of the train RMSEs and test RMSEs per epoch.
  • output the lowest test RMSE.
  • display the weights of the 1st convolutional layer filters.

2. Example of testing the DBN model:

In this example, we want to predict the exact speed for network points of 'I5_North_D7' freeway at 15-min forecasting by testing/applying a DBN which was previously trained based on the probabilistic loss function using data of 'US101-North District 7' freeway (i.e., H101-North-D7) .

To do so, follow these steps:

  1. Go to 'traffic_flow_code_DBN' directory and run the file 'proj_traffic_flow_prediction_DBN.m'.
  2. select the following:

  • Please enter the k-fold (k-1 for training & 1 for testing)_(0 for testing): 0

  • Please enter the loss (0)L2 loss, (1)P loss: 1

  • Please select the number of days (15, 21, 27, 30 or 59): 30

  • Enter prediction point: 3

  • Please select the testing freeway: H101_North_D7 / I5_North_D7 / I5_South_D7 / I5_North_D11 / I450_North_D7 / I210_West_D7: I5_North_D7

  • Please select the input freeway used for training: H101_North_D7 / I5_North_D7 / I5_South_D7 / I5_North_D11 / I450_North_D7 / I210_West_D7: H101_North_D7

The displayed result is : Classification error (testing): 5.53

3. Example of testing the CNN model:

In this example, we want to predict the exact speed for network points of 'I5_North_D11' freeway at 15-min forecasting by testing/applying a CNN which was previously trained based on the probabilistic loss function using data of 'US101-North District 7' freeway (i.e., H101-North-D7) .

To do so, follow these steps:

  1. Go to 'traffic_flow_code_CNN' directory and run the file 'proj_traffic_flow_prediction_10wStr.m'.
  2. select the following:

  • Please forecasting for which you wish to predict speed (1)for 5-min, (2)for 10-min, (3)...): 3

  • Please select the prediction type: (c)classification / (r)regression r

  • Please enter the loss (0)L2 loss, (1)P loss: 1

  • Please select the freeway used for training: H101_North_D7 / I5_North_D7 / I5_South_D7 / I5_North_D11 / I450_North_D7 / I210_West_D7: H101_North_D7

  • Please enter the k-fold (k-1 for training & 1 for testing)_(0 for testing): 0

  • Please select the testing freeway: H101_North_D7 / I5_North_D7 / I5_South_D7 / I5_North_D11 / I450_North_D7 / I210_West_D7: I5_North_D11

  • Please select the number of days (15, 21, 27, 30 or 59): 30

The displayed result is : Lowest validation error is 3.821451 in epoch 1