Short-term Crowdedness Predictions for Locations in Amsterdam

Abstract

The aim of this thesis was to investigate multiple models for short-term crowdedness predictions based on mobile phone data. For this, we used data of three different locations in the city of Amsterdam (a square, park and market). We examined the contribution of various external factors (such as the weather and COVID-19 regulations) and we compared different modelling techniques (such as regression and LSTM) when predicting crowdedness two hours ahead. We found that regression models obtained the highest prediction accuracy when used in combination with an oversampling technique to account for the sparsity of crowded samples. Furthermore, we found that historical values of the ground truth data (e.g. crowdedness of the previous time step) and information on temporal aspects (e.g. time of day and day of the week) were most influential in the prediction models. These prediction models could be used to support crowd management by providing the expected crowdedness for the near future.