GIS-Based Car-Sharing Demand Prediction and Analysis

Dominik received his bachelor’s degree in Geospatial Engineering at the Institute of Cartography and Geoinformation of ETH Zurich, Switzerland. Now, he is pursuing a master’s degree in Geomatic Engineering in the same institute. His interests lie in GIS, data science, artificial intelligence, spatial development, and transportation. Dominik gained first working experience in the field of GIS at the Federal Office for the Environment in Switzerland. There, he supported the division for biodiversity and landscape by performing spatial analysis for monitoring the development of nature reserves over time, validation and quality assurance of environmental geodata, and implementing software for the automated creation of legal object sheets for the federal inventory of Swiss game reserves using ArcGIS map series

This story map is based on his bachelor thesis and a follow-up study in collaboration with his supervisors Nina Wiedemann, Dr. Yanan Xin and Prof. Dr. Martin Raubal. The scientific publication is available as a preprint [1].

Car-sharing has become increasingly popular as a sustainable alternative to car ownership. By reducing the number of cars on the road and promoting more efficient use of resources, car-sharing can help mitigate emissions and slow down climate change. Moreover, car-sharing can reduce the need for parking spaces in cities, improving mobility and making cities more livable. Improving services and lowering costs is essential to increase car-sharing usage, as car-sharing businesses are rarely profitable. Therefore, car-sharing operators must make informed decisions on business strategy, station placement, and marketing to ensure the highest level of user satisfaction and efficiency. An accurate prediction of car-sharing demand and a good understanding of how spatial and socio-demographic features influence the demand are integral parts of the informed decision-making process.

Through their possibilities in prediction and spatial analysis for solving real-world problems, modern GIS and machine learning methods have great potential to help improve car-sharing operations and support decision-making. Using a dataset from Switzerland's largest provider of car-sharing services, this work investigates the effectiveness of combining geoinformation technology and artificial intelligence to predict and analyze car-sharing demand and identify the spatial and socio-demographic factors that influence the demand.

These demand-prediction methods can serve as a decision-support tool for planning new car-sharing stations. However, the supply of vehicles at a station is an essential feature for calculating its demand. Differences in vehicle supply between stations cannot be explained solely by socio-demographic factors. Therefore, it is possible to estimate demand while varying the vehicle-supply input feature to determine demand and the optimal vehicle supply at a station. As a case study, three potential locations for new car-sharing stations in the region of Zurich were selected. The locations vary in their spatial context; one is in the city center, another in an intermediate region, and the last in the countryside. The demand for each station was predicted by varying the number of vehicles placed at each station, as shown in the figure below. While GWR is a locally-linear model in which predicted demand constantly increases with the supply of cars, Random Forest saturates for a higher number of vehicles. The demand to which the model converges highly depends on socio-demographic features, with the city location converging to a much higher demand than the rural or intermediate station. The model can simulate a realistic saturation pattern for the city location since the largest station in the neighborhood shows a very similar demand and vehicle supply. Therefore, the Random Forest model can help plan the new station's supply of cars, whether to place all vehicles at a single station or arrange several smaller stations depending on infrastructure availability and design strategies.

The research investigated the task of car-sharing demand prediction, with the ultimate goal of reducing private cars and CO ₂  emissions to slow down climate change. To achieve this, a combination of GIS, machine learning and geographically weighted methods are needed. Geoinformation technology allows the merging of various geodata sources to build the foundation for demand analysis and prediction. While powerful machine learning methods achieve high performance for car-sharing demand prediction, they often lack interpretability. Geographically weighted methods can fill this gap by providing fascinating insights into spatially varying patterns. Overall, this comprehensive approach to car-sharing demand analysis and prediction can support operators in making informed decisions to contribute to a more sustainable future mobility.

This research is based on ArcGIS Pro and Python. ArcGIS Pro proves to be a powerful tool for data processing and visualization. It is also the first available software that implemented MGWR for prediction, which allows forecasting demand for new car-sharing stations through out-of-sample prediction. Since the algorithm needs to find an optimal bandwidth for each feature, computation time is often an important consideration, especially if there are many observations and features. However, the overall MGWR analysis using ArcGIS Pro is reasonably fast given the number of observations we have. Besides, its ease of use makes it a suitable tool for analysts without experience in programming.

I would like to express my profound gratitude to my supervisors, Dr. Yanan Xin, Nina Wiedemann, and Prof. Dr. Martin Raubal, for their exceptional support during my bachelor's thesis and our collaboration on the subsequent research study. We had engaging and productive discussions, and I was able to benefit greatly from their experience. I would also like to thank the Swiss Federal Office of Energy (SFOE) for supporting the research project through Grant 1-008352.

[1] Mühlematter, D. J., Wiedemann, N., Xin, Y., & Raubal, M. (2023). Spatially-Aware Car-Sharing Demand Prediction [Preprint].  arXiv:2303.14421 

[2] Mobility Cooperative. (2022). Private Customers.  https://www.mobility.ch/en/private-customers . Retrieved March 30, 2023.

[3] Deplazes, J., Kohli, R., & Babel, J. (2020). Szenarien zur Bevölkerungsentwicklung der Schweiz und der Kantone. Federal Statistical Office (FSO), Neuchâtel.

[4] Federal Office for Spatial Development (ARE), Federal Statistical Office (FSO). (2015). Mobility and Transport Microcensus (MTMC).  https://www.are.admin.ch/are/en/home/mobility/data/mtmc.html . Retrieved March 23, 2023

[5] Federal Statistical Office. (2021a). Structural Business Statistics (STATENT).  https://www.bfs.admin.ch/bfs/en/home/services/geostat/swiss-federal-statistics-geodata/business-employment/structural-business-statistics-statent-from-2011-onwards.html . Retrieved March 23, 2023.

[6] Federal Statistical Office. (2021b). Population and Households Statistics (STATPOP).  https://www.bfs.admin.ch/bfs/en/home/services/geostat/swiss-federal-statistics-geodata/population-buildings-dwellings-persons/population-housholds-from-2010.html . Retrieved March 23, 2023.

[7] Federal Office of Spatial Development (ARE). (2022). ÖV-Güteklassen Berechnungsmethodik - Grundlagenbericht.

[8] OpenStreetMap contributors. Planet dump retrieved from  https://planet.osm.org .  https://www.openstreetmap.org , 2022.

[9] Federal Statistical Office (FSO). (2020). Szenarien zur Bevölkerungsentwicklung der Kantone 2020-2050, Referenzszenario AR-00-2020 - zukünftige Bevölkerungsentwicklung nach Kanton, Staatsangehörigkeit (Kategorie), Geschlecht und Alter - 2019-2050.  https://www.bfs.admin.ch/asset/de/12947637.  Retrieved March 30, 2023.