PROJECT TITLE :

On Model Selection for Scalable Time Series Forecasting in Transport Networks

ABSTRACT:

When it comes to short-term traffic predictions, up to the scale of one hour, the transport literature is quite extensive; however, when it comes to long-term traffic predictions, the transport literature is less extensive. When it comes to city-scale traffic predictions, the transport literature is similarly lacking, primarily as a result of the limited availability of data. In this paper, we report on an investigation that was carried out to determine whether or not deep learning models can be useful for the task of long-term large-scale traffic prediction, with the primary emphasis being placed on the scalability of the models. We investigate a city-scale traffic dataset in the hypercenter of Los Angeles, California, which includes 14 weeks of speed observations collected every 15 minutes over 1098 segments. We take a look at a number of different state-of-the-art machine learning and deep learning predictors for link-based predictions and investigate how such predictors can scale up to larger areas using clustering and graph convolutional approaches. We discuss how the incorporation of temporal and spatial features into deep learning predictors can be useful for long-term forecasting, whereas the performance of simpler predictors that are not based on deep learning can be very satisfactorily achieved for link-based and short-term forecasting. The trade-off is discussed not only in terms of the accuracy of the prediction versus the time horizon of the prediction, but also in terms of the amount of training time and the size of the model.