A deep learning-based framework for road traffic prediction

Due to the exponential rise in the number of vehicles and road segments in cities, traffic prediction becomes more difficult, necessitating the application of sophisticated algorithms such as deep learning (DL). The models used in the literature provide accurate predictions for specific cases when the data flow is properly prepared. However, in complex situations, these approaches fail, and thus, the prediction must be developed through a process rather than a prediction calculation method. In addition to using a pure and robust DL prediction model, an efficient approach could be built by taking into account two other factors, namely the relationships between road segments and the amount and quality of the training data. The main goal of our research is to develop a three-stage framework for road traffic prediction based on statistical and deep learning modules. First, a cross-correlation prediction with a Long Short-Term Memory model (LSTM) is implemented to predict the influential road segments; second, a deep generative model (DGM)-based data augmentation is used to improve the data of the related segments; and third, we adapt a Neural Basis Expansion Analysis for interpretable Time Series (N-BEATS) architecture, to the resulting data to implement the prediction module. The framework components are trained and validated using the 6th Beijing road traffic dataset.