Exploring the Influence of Data Aggregation in Parking Prediction

Abstract

Parking occupancy is influenced by many external factors that make the availability prediction task difficult. We want to investigate how information from different data sources, such as events, weather and geographical entities interrelate in affecting parking prediction and thereby form a knowledge graph for the parking prediction problem. In this paper, we try to tackle this problem by answering the following questions; What is the effect of the external features on different models? Is there a correlation between the amount of historical training data and external features? These questions are evaluated by applying three well-known time series forecasting models; long short term memory, convolutional neural network and multilayer perceptron. Additionally we introduce gradient boosted regression trees with handcrafted features. Experimental results on two real-world datasets showed that external features have a significant effect throughout the experiments and that the extent of the effectiveness varies across training histories and tested models. The findings show that the models are able to outperform recent work in the parking prediction literature. Furthermore, a comparison of the feature-engineered gradient boosted decision trees to other potential models has shown its advantage in the field of time series forecasting.