Traffic Prediction with Convolutional Long Short-Term Memory

Abstract

that underlying our data are spatio-temporal correlations, In recent years, Convolutional LSTM neural networks have which Conv-LSTM can take advantage of. We begin by demonstrated superior performance when applied to prob- building a simple FC-LSTM architecture, then continue to lems in multiple domains, including biology [1], weather gradually enhance it to yield better predictions, compared forecasting [2], and speech recognition [3]. In this work, to a Linear Regression baseline. Then, we implement a we apply this technique to the traffic domain, by measuring Conv-LSTM network, apply it to the same input, and show its accuracy in predicting speeds and flows. Our dataset that Conv-LSTM outperforms FC-LSTM in predicting both comprises of 6 months of traffic information, collected from speeds and flows. Android devices in several roads around Nørre Campus in Copenhagen, Denmark. We compare the predictive perfor- II. DATASET mance of Convolutional LSTM to several Recurrent Neural The data we use in this work consists of traffic information Network architectures which use a more "classic" Fully- for several places around Nørre Campus: a campus of the Connected LSTM. The results show that for this traffic University of Copenhagen. Each place falls into one of two forecasting problem too, Convolutional LSTM outperforms place groups: junctions and middle-of-roads (both traffic other models. directions). Because traffic tends to flow more freely in