Abstract

Effective on-street parking is key to reduce urban traffic and pollution in densely populated cities. Thus, researchers have focused on forecasting future occupancy values depending on factors like time, space, or weather. This approach shows high average performances, but fails in predicting congested scenarios, actually the most critical.
This work proposes a data-driven parking level of service (LOS) predictor that outperforms traditional methods, solving its inherent class imbalance issue by means of Random Undersampling Boost classifiers. We trained and validated the LOS classifiers using 13 months of data collected from the smart parking system in the city of Madrid, Spain. Results display average recall values above 0.94 and 0.87 at prediction horizons up to 10 and 60 minutes respectively. We compare these results with traditional regression-based occupancy predictors showing that our classifier clearly outperforms the formers predicting the minority classes, which carry the most significant information for drivers and parking managers.
We further analyze the impact of performance on temporal and spatial features, revealing mid-term temporal data as the most relevant forecasting information, and low correlations between parking behaviors in bordering neighborhoods. In the light of these results, we believe that the proposed data-driven parking LOS classification has the potential to open a novel perspective