Optimizing Bike Sharing Demand Prediction: A Comparative Study of Multiple Linear Regression and LSTM Models based on Time and Weather Factors

Abstract

Bike sharing, as an emerging Internet product, utilizes advanced technologies like IoT, cloud computing, and big data to improve urban public transport and offer more travel options. However, poor demand forecasting has resulted in the over-placement of shared bicycles. In this paper, we explore the use of an LSTM-based approach to predict the demand for shared bikes by analyzing time and weather factors to optimize operations. The study begins by underscoring the significance of these factors in demand forecasting. After identifying and addressing outliers through the use of box plots and the DBI index, we find that bicycle usage exhibits a cyclical pattern over time with a lagged weather effect. To tackle this issue, we introduce triangular transformations of the temporal factor and lagged weather variables in our multiple linear regression model. While these adjustments enhance the model's ability to capture complex relationships, they result in poor accuracy. Consequently, we employ a Long Short-Term Memory (LSTM) neural network, which significantly outperforms the traditional multiple linear regression model. The LSTM model achieves RMSE values of 14.63 and 39.41, as well as R-squared values of 93.21% and 95.58% for casual and registered users, respectively. This highlights its superior ability to capture complex patterns in the data, providing insights for better management and optimization of shared bike systems.