Fleet Management Optimisation with Spatio-Temporal Demand Forecasting in MaaS for Free-floating Micro-mobility
Ladda ner
Typ
Examensarbete på kandidatnivå
Program
Publicerad
2021
Författare
Almström, Oscar
Carlsson, Erik
Cronqvist, Daniel
Karlsson, Max
Lilliecreutz, Fredrik
Viala Bellander, Alexander
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
In recent years, micro-mobility services have grown rapidly. Companies such as
Voi, Bolt and Lime are front figures in this development. Utilising their resources
efficiently through fleet management has been an essential factor in reaching profitability. Gathering data is becoming more critical for companies, enabling them to
use data science and mathematical models to leverage their business. This report
aims to examine the opportunity to use different data-driven models to predict future demand for micro-mobility services. Voi Technology provided real market data
for this paper, limited to Gothenburg. All geospatial data was aggregated to geospatial units defined by Uber’s H3 spatial index. Furthermore, the research focused on
the following methods; Markov properties, Time Series Analysis and Poisson processes. These methods were implemented in the programming language Python.
The Markov and Prophet models provided solid demand predictions. However, due
to sparse data representation, for some areas of the city, the data was clustered to
give more accurate forecasts at the cost of geospatial granularity.
Comparing the two, the Prophet models gave better results alone, while the
Markov models instead gave insight into how the supply moves around in the market. With further work, the Markov model could prove favourable for a fleet operator
during rebalancing due to its ability to track the vehicle flow and predict lack and
surplus of supply. The data indicated significant seasonality effects and sporadic
behaviour. FBprophet showed decent results in analysing those characteristics, using a sliding window technique as a significant contributor. Supporting FBprophet
with several features improved the prediction, where rain stood out as the most impactful feature. The Poisson proccess model interprets demand as non-homogeneous
and stochastic with inherent temporal randomness. While the theory behind the
Poisson process model seems relevant, the results remain inconclusive. FBprophet
performed the best demand prediction in contrast to the other models. However,
further research could contradict this, and there is room for deeper exploration.
Onwards, Neural Networks and Deep Learning are also exciting subjects for further
research in demand forecasting.
Beskrivning
Ämne/nyckelord
data analysis , micro-mobility , MaaS , demand forecasting , Markov chains , time series analysis , fleet optimisation , FBProphet , Poisson processes