Traffic flow forecasting on data-scarce environments using ARIMA and LSTM Networks

Abstract

Traffic flow forecasting has been in the mind of researchers for the last decades, remaining a challenge mainly due to its stochastic nonlinear nature. In fact, producing accurate traffic flow predictions would be extremely useful not only for drivers but also for those more vulnerable in the road, such as pedestrians or cyclists. With a citizen-first approach in mind, forecasting models can be used to help advise citizens based on the perception of outdoor risks, dangerous behaviors and time delays, among others. Hence, this work develops and evaluates the accuracy of different ARIMA and LSTM based-models for traffic flow forecasting on data-scarce and non-data-scarce environments. The obtained results show the great potential of LSTM networks while, in contrast, expose the poor performance of ARIMA models on large datasets. Nonetheless, both were able to identify trends and the cyclic nature of traffic.