Spatio-Temporal Clustering for Optimal Real-Time Parking Availability Estimation

Urban transport systems represent a major infrastructure asset in contemporary cities, enabling many millions of people to commute and travel every day. Transport systems are increasingly complex because of rapid urbanization and rising vehicle ownership. Effectively predicting parking availability across a city means more efficient parking management, better urban planning, smoother traffic flow, lower fuel wastage and, ultimately, less environmental pollution. To this end, several research studies have proposed predictive approaches to parking space availability based on supervised machine learning, mainly regression algorithms. But if real-time information on parking space availability is lacking, these approaches become useless, since their driving force is historical data. What’s more, many city zones simply don’t require exact information on parking space occupancy; all that’s needed is a global view of whether there are any spaces available or not. That’s why, in this paper, we outline an approach to predicting parking availability that combines both clustering and classification. Our aim is to model parking availability through a number of typical days. A typical day is defined by a profile characterizing parking space occupancy. We start by determining precisely how many typical days per spatial cluster are required to map parking availability, and then cluster data to form groups typified by each typical day, using six clustering algorithms of ascending difficulty. Obtained results show that this number varies between 3 and 4, depending on which algorithm is applied. Thus, model evaluation reveals that distance between cluster elements is short and separation between clusters is high, in other words, clusters are far apart and not very dispersed.