Forecasting Parking Search Times Using Big Data

Searching for parking (also known as cruising) is one of the key contributors to urban congestion and consequently air pollution. Providing information about parking availability is already present in many cities, efficiently guiding drivers to vacant parking spaces, either based on past data or using technologies for real-time information. Informing drivers about the expected cruising time at their destination is an additional pathway, which can impact their departure time and mode choices, and consequently improve the overall mobility system performances. One of the main challenge of cruising time estimation lies within how cruising itself is detected. This study examines the case of detecting cruising using GPS traces from trajectories. A new parametric detection method is proposed establishing speed and acceleration/deceleration related conditions. A sensitivity analysis to the method is presented along with a comparison against existing and similar GPS-based cruising detection methods and validation against labelled data. After the cruising detection, about 800,000 GPS trips were used to estimate and validate an offline machine learning algorithm to forecast the cruising time in three different urban areas in the City of Copenhagen, Denmark, with clear distinct parking conditions. Neighborhoods were divided into spatial cells for which hourly cruising times were estimated. Feed-Forward Neural Network (FFNN) and eXtreme Gradient Boosting (XGBoost) architectures were tested as machine learning algorithms and outperformed a simple moving average (RMSE gains from 62.01 to 52.57 s). The present study paves the ground for the exploration of large datasets with GPS trajectories in urban areas for tackling the lack of information on parking search. Despite the improved overall prediction power, the potential errors from the cruising detection method, lack of data needed to capture patterns when cruising time is high or the existence of many missing values due to aggregation of data could be the reason for the observed algorithm’s inability to predict the larger values of cruising time.