Ebike Sharing vs. Bike Sharing: Demand Prediction Using Deep Neural Networks and Random Forests

Background: Conventional bike sharing systems are frequently adding electric bicycles. A major question now arises: Does the bike sharing system have a sufficient number of ebikes available, and are there customers who prefer to use an ebike even though none are available? Methods: Trip data from three different bike sharing systems (Indego in Philadelphia, Santander Cycles in London, and Metro in Los Angeles and Austin) have been used in this study. To determine if an ebike was available at the station when a customer departed, an algorithm was created. Using only those trips that departed while an ebike was available, a random forest classifier and deep neural network classifier were used to predict whether the trip was completed with an ebike or not. These models were used to predict the potential demand for ebikes at times when no ebikes were available. Results: For the system with the highest prediction accuracy, Santander Cycles in London, between 21% and 27% of the trips were predicted to have used an ebike if one had been available. The most important features were temperature, distance, wind speed, and altitude difference. Conclusion: The prediction methods can help bike sharing operators to estimate the current demand for ebikes.

Automated summary

This study uses data from three different bike sharing systems to predict the demand for electric bicycles using an algorithm and classifiers. The system with the highest prediction accuracy, Santander Cycles in London, predicts that approximately 21% to 27% of trips would have used an electric bike if one had been available. Temperature, distance, wind speed, and altitude difference are the most important features.