Floating Car Data–Based Short-Term Travel Time Forecasting with Deep Recurrent Neural Networks Incorporating Weather Data

Abstract

The prediction of future traffic conditions represents a main building block for traffic management. With the advent of multiple traffic and environmental sensors, diverse data for predictions are available and models may incorporate not only traffic data but also additional aspects such as local weather conditions. A review of the state-of-the art methods in short-term traffic forecasting presented in this paper reveals that machine learning (ML) algorithms from the field of deep learning are occasionally used for forecasts based on historical traffic data, but not for traffic predictions including exogenous factors. Weather conditions represent an exogenous factor, which, for example, may affect travel times. This paper investigates how short-term travel time predictions may be improved by applying deep recurrent neural networks that incorporate weather data. Therefore, two hypotheses are formulated. Hypothesis 1 tests the prediction quality of the recurrent neural network (RNN) models, long short-term memory (LSTM), and gated recurrent unit (GRU) compared to the autoregressive moving average (ARMA) prediction method. The respective results indicate that the RNN models using historical traffic data and weather data show significant improvement compared to the ARMA model using only historical traffic data. Hypothesis 2 tests the prediction quality of LSTM and GRU compared to ML-based forecast models already in place in the field of traffic predictions, namely k-nearest neighbor (kNN), support vector regression (SVR), and neural networks (NN). In this context, the LSTM and GRU models using historical traffic data and weather data show significant improvement compared to the models kNN, SVR, and NN that also consider weather data. Despite the results presented in this work, there is still further potential for improvement. Thus, further research focusing on hyperparameter tuning of the RNN algorithms and the optimized selection of (additional) input variables with significant influence on travel times can contribute to further improvements of the forecast quality.