Traffic-flow prediction via granular computing and stacked autoencoder

Accurate traffic-flow prediction is essential to traffic management. Traffic data collected in very short intervals normally contain high variability, while common preprocessing approaches applied within a window such as simple average or median operator are unable to obtain sufficient latent information from original data. Moreover, the prediction performance of shallow neural network is not satisfying, since its capacity to represent the temporal-spatial correlation of mass traffic data is insufficient, and its adaptation capacity to noisy data is relatively poor. In this paper, fuzzy information granulation (FIG) and deep neural network are combined to solve these two issues. To be specific, FIG is utilized to process original data series and extract granules, which denote the trend and fluctuation range of each time window. Then, stacked autoencoder is combined to obtain the predictive results based on processed granules, especially, a multi-output mechanism is designed to predict all granulation parameters simultaneously, which makes better use of the correlation of diverse inputs. A real-world traffic volume data set is applied to conduct an empirical study, and the experimental results illustrate that based on the proposed method, the interval prediction of the traffic-flow fluctuation range is realized, and superior traffic trend prediction performance is achieved.