Predicting Parking Occupancy with Deep Learning on Noisy Empirical Data

Parking is a contributor to urban traffic. According to a survey by the International Parking Institute (IPI), over 30% of cars cruising in the cities are looking for parking space, which highly contributes to urban congestion. Besides, an increase or decrease of parking occupancy is an indicator for decreasing or increasing car travel demand. Therefore, predicting urban parking occupancy can be beneficial for routing and urban traffic management. However, precision parking prediction solutions often require, e.g., investments in real-time detection technologies or access to big volumes of floating car data. However, parking data is rarely available in the form in which it is required. In this paper, we work with data of a small city in Germany that contains the occupancy of eleven parking lots recorded over one year. We implement a Long-Short-Term Memory (LSTM) model to predict parking occupancy, focusing on the following topics: challenges and limitations of the given data, performance and sensitivity to meteorological and event features and time sequence selection for training. The highest accuracy is reached when choosing two days data - the day to predict for and the corresponding day of the preceding week - within one week time window for parking prediction.