Where is My Parking Spot?

Abstract

Parking occupancy information is central to the management of parking and traffic demand. This study proposed efficient unsupervised learning algorithms to predict parking occupancy rates. Two types of predictions were studied: (a) an offline prediction, in which next-day occupancy was predicted by using historical data along with various features (day of week, weather, seasonality), and (b) an online prediction, in which occupancy of future hours of the current day was predicted with both historical and real-time data. The two models can be applied to both off-street and on-street parking. Two data sources were used: parking payment kiosks for a visitors' parking garage and newly deployed real-time spot-by-spot parking sensors for a commuter garage. It was found that, with a proper set of features, the offline method could successfully distinguish different flow patterns, congested or underused, with intensive or mild arrival and departure rates. The offline procedure significantly outperformed both the historical and the previous day's average. The online method provided generally more accurate predictions than the offline method because it learned from the real-time occupancy data. As time progressed, the mean and maximum error rates of the online prediction decreased to a level well below both the historical average and the offline prediction error. A sharp decline of the prediction error could be obtained when sufficient real-time occupancy data were collected and the type of flow pattern was identified (around 9:00 a.m. in a case study).