Evaluating Machine Learning Methods for City Bike Demand Prediction in Oslo
Abstract
A bike-sharing system is a service in which a fleet of bicycles is made available to the publicon a short-term basis through self-served docking stations. These stations are limitedin capacity and are often depleted or saturated with bikes due to sudden spikes in demand.These spikes are hard to avoid and are both detrimental to the user experience and theeffectiveness of the system. Machine learning methods have been used to forecast demandspikes at station level in similar systems successfully and would likely be a valuable toolin proactively counteracting the effect of demand spikes in the Oslo bike-sharing system.
The goal of this thesis is to evaluate common machine learning methods for demand predictionmodeling at individual bike-sharing stations in Oslo.
To accomplish this, four machine learning methods which have successfully predictedstation-level demand in similar systems are evaluated through a set of experiments. Themethods are based on a random forest, gradient boosting tree and recurrent neural networkswith either long short-term memory- or gated recurrent unit units.
Based on the experimental results, the recurrent neural network with the long short-termmemory unit is deemed to be the most suitable for the Oslo bike-sharing system, bothdue to performance and future potential. However, all methods achieved good performanceand made accurate predictions.
The results pave the road for developing a full-scale prediction system in the Oslo bike-sharingsystem, by highlighting the most promising prediction method.