Traffic Congestion Prediction Using Categorized Vehicular Speed Data

Abstract

Traffic congestion has become a problem which is associated with our day-to-day life. Industrialization, urbanization, and the increasing population of cities have the main role in this problem. Transportation agencies in almost all the countries are trying their best to alleviate the traffic congestion problem. This study develops a traffic congestion evaluation method, which is categorized into four states named low, medium, congestion, and serious congestion. Convolutional long short-term memory (Conv-LSTM) neural network (NN) was trained to learn multivariate features as the inputs like categorized vehicular speed, time, and day of the week with respect to the traffic speed as an output of the whole stream. Trained model was used for short-term traffic speed prediction. Also, the prediction accuracy and stability of the Conv-LSTM NN has been compared with other neural network models, e.g., multi-layer perceptron (MLP), cascade forward back-propagation (CFBP), recurrent neural network (RNN), long short-term memory (LSTM) NN, and convolutional neural network(CNN). Results confirm that the Conv-LSTM NN achieves higher prediction performance than the compared models. Python 3.8.3 was used for programming purpose. Model code automatically selects 80% and 20% time intervals of the datasets in a random way for training and testing, respectively, from the traffic data. Furthermore, Conv-LSTM NN was combined with the developed congestion evaluation method and then the traffic congestion state was predicted. Study results confirm that the Conv-LSTM NN can be successfully applied for the prediction of traffic speed and traffic congestion status with non-homogeneous traffic in the Indian context.