Scalable data-driven short-term traffic prediction

Short-term traffic prediction has a lot of potential for traffic management. However, most research has traditionally focused on either traffic models - which do not scale very well to large networks, computationally - or on data-driven methods for freeways, leaving out urban arterials completely. Urban arterials complicate traffic predictions, compared to freeways, because the non-linear effects of traffic are more pronounced on short links and with the presence of more crossings, more modalities and non-necessarily conservation of flow (parking). In this paper we consider several data-driven methods and their prediction performance for various situations, including both freeways and urban arterials, for prediction horizons from five minutes up to one week. The focus lies on predicting the traffic flow and speed given available measured data on a certain location. Thus challenges regarding data fusion, state estimation, and other methods providing complete temporal and spatial data is not addressed. The methods evaluated include several naive, parametric and non-parametric methods. For the evaluation of the prediction performance several weeks of data of various locations were used. Performance indicators contained the Root Mean Square Error, Mean Absolute Percentage Error and Mean Absolute Error. Because evaluating average performance might ignore the performance for non-regular traffic conditions, the evaluation also focused on non-regular traffic conditions. Especially these conditions are important in practice, because in these situations the need for accurate information is the highest. Real-world applicability of traffic prediction requires not only accurate results, but also an indication of the accuracy for each prediction. Earlier research has mostly ignored this, leaving this up to the intuition of users of these predictions. This paper introduces a simple way to calculate confidence intervals, applicable to any traffic prediction method. For comparing these confidence intervals between different prediction horizons, an error measure for the accuracy of a confidence interval is defined. Two methods, SARIMA (Seasonal Auto-Regressive Integrated Moving Average) and NLM (Neighborhood Link Method), proved to be the best. The results also indicate key features necessary for accurate traffic prediction with data-driven methods. The results also show reasonably accurate confidence intervals, with those intervals able to adapt well to different traffic situations.