An interpretable model for short term traffic flow prediction

Predicting short term traffic flow to improve traffic control is a research problem attracting increased attention over the past 30 years. With increasing number of traffic data acquisition equipments coming into usage, it provides an opportunity to use deep neural network (DNN) to predict short-term traffic flow. Behind its considerable success, the DNN is weighed down by some problems, and here we focus on: 1. how to justify the number of input nodes employed by DNN; 2. how to explain the causality between the historical spatiotemporal information and the future traffic condition. In this paper, we propose a deep polynomial neural network combined with a seasonal autoregressive integrated moving average model. The new model has superior predicting accuracy as well as enhanced clarity on the spatiotemporal relationship in its deep architecture. Experimental results indicate that the proposed model has better explanation power and higher accuracy compared with the LSTM based model.