Comparing machine learning models for large scale prediction of parking space availability

The search for solutions to the problems related to traffic congestion has become a significant challenge for large urban cities. The analysis of traffic jams in some congested cities revealed that more than 70% of traffic congestion is due to a prolonged search for parking spaces. Predicting the availability of parking spaces in advance is a vital step to intelligently assist drivers in their research for parking spaces and allow them to find free spaces quickly. Thus, reducing traffic congestion and its negative impacts on the environment, the economy, and public health. Several solutions have been proposed to solve the problems due to traffic congestion. Among those solutions is the use of machine learning algorithms to predict parking spaces’ availabilities. However, there is a lack of comparative study to find the models that work best on a large scale and with a given prediction time period. In the present study, we propose a comparison of nine machine learning algorithms for their ability to efficiently predict long-term, large-scale parking space availability. Based on two approaches: 1) using on-street parking data alone and 2) incorporating data from external sources (such as weather data), we used automatic machine learning models to compare the performance of different algorithms according to the prediction efficiency and execution time. Our results showed that the automatic machine learning models implemented were fitted to our data. The Extra Tree and Random Forest algorithms were the most efficient models. In addition, when taking into account the execution time, we observed that the Random Forest algorithm was less greedy than the Extra Tree. This work suggests that the Random Forest algorithm is the best-suited machine learning model in terms of efficiency and execution time for predicting large-scale, long-term parking space availabilities.