Demand Prediction for Bikeshare Using BERT

Abstract

In this paper, we propose a method for predicting hourly bike availability for the coming month at bikeshare stations, by utilizing a large language model (LLM). Our approach begins by converting historical bikeshare station data into text. The training dataset includes information such as station ID, date, time, timeslot, day of the week, weekday/holiday classification, national holidays, temperature, weather conditions, station location, days since establishment, station penetration in the area, comfort level, wind speed, precipitation, user registration type, movement between stations, and bike inflow/outflow at each station. These elements are combined into text format, with the number of available bikes per station per hour serving as the target label. We then fine-tune BERT, an LLM, to predict these labels. Our method achieved an approximate 3.1% improvement in average RMSE compared to machine learning models trained on text data, and an approximate 16.4% improvement in average RMSE compared to machine learning models trained on tabular data. These results demonstrate the effectiveness of converting historical bikeshare data into text and fine-tuning LLMs for demand prediction.