Predicting Free Parking Slots via Deep Learning in Short-Mid Terms Explaining Temporal Impact of Features

Abstract

Looking for available parking slots has become a serious issue in urban mobility, since it influences traffic and emissions. This paper presents a set of metrics and techniques to predict the number of available parking slots in off-street parking facilities. This study deals with deep learning model solutions according with a mid-term prediction of 24 hours, every 15 minutes. Such a mid-term prediction can be useful for citizens who need to plan a car transfer well in advance and to reduce as much as possible any computational effort. Since most solutions in literature are focused on 1-hour ahead prediction, the proposed solution has been also tested in these conditions. The proposed solution is based on Convolutional Bidirectional LSTM models. Results have been compared in terms of precision metrics based both on occupancy and free slots. The paper also provides a framework to pass from an assessment model based on occupancy to models based on free slots and vice-versa. The obtained results have improved those already available in literature. A formal study has been conducted to perform feature relevance analysis by using explainable AI technique based on gradient and integrated gradient and proposing new heatmaps which highlighted the difference from LSTM and Bidirectional LSTM, feature relevance (base line, weather, traffic, etc.) and the impact of seasonality on predictions, namely the temporal relevance of features. The comparison has been performed on the basis of data collected in garages in the area of Florence, Tuscany, Italy by using Snap4city platform and infrastructure.

Keywords

• Smart city
• available parking lots
• prediction model
• machine learning
• deep learning
• explainable AI