Traffic Speed Forecasting: Comparison of Modeling Approaches

Abstract

Over the past two decades, building an intelligent transportation system has become a popular and challenging research topic. As a key role in such a system, accurate traffic speed forecasting is critical. Although many powerful prediction methods have been proposed, they have not considered the application of models in real situations, that is, on various types of roads with different characteristics. So, we apply some representative and state-of-the-art methods on different types of roads to help people select the appropriate prediction method to construct an intelligent transportation system. First, we use the traffic data of 214 roads in Guangzhou in 61 days as the data set, and select four typical roads according to their characteristics of the roads. Then we use feature engineering to enhance the quality of the data set. After that, we apply Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing, short-term and long-term memory (LSTM) neural networks, and Informer on selected roads to make comparisons. The performance of models varies significantly in different types of roads: The Mean Absolute Error (MAE) for low mean and low variance roads is around 2, but the MAE for high mean and high variance roads is about 5. Notably, the Holt-Winters model shows the best performance in short-period prediction, and the Informer model offers the best performance in long-period prediction in the benchmarking.