A 2-Step Approach to Improve Data-driven Parking Availability Predictions

Knowing where to park in advance is a most wished feature by many drivers. In recent years, many research efforts have been spent to analyse massive amount of parking information, to learn availability trends and thus to predict, within a Parking Guidance and Information (PGI) system, where there is the highest chance to find free parking spaces. The most of these solutions exploits raw data coming from stationary sensors or crowd-sensed by mobile probes. In both the cases, these massive amounts of data present a high level of noise. In this paper we propose a 2-step approach to predict parking space availability with the twofold goal to handle the noise in the data and to significantly reduce the space needed to store these models. In particular, in the first step, we smooth the raw parking data by using Support Vector Regressions (SVR) in combination with a specifically defined technique to tune the SVR parameters. In the second step, on top of this smoothed trend curve, we train a multidimensional SVR model, representing parking space availability, and suitable for parking predictions. The proposal has been empirically evaluated on a real-world dataset of on-street parking information from the SFpark project, and compared against a standard, one-step SVR model with different settings. Results show that the predictions obtained with the proposed approach are always by far more accurate, with a statistically significant difference, while requiring a fraction of the storage normally used for raw data.