An Effective Data Driven Approach to Predict Bike Rental Demand

Bike Sharing Systems (BSS) are becoming more popular since they are more space-efficient in crowded cities and promote physical fitness. By not emitting any greenhouse gases. Also protects the environment by not producing CO2. In BSS, the system operators generally distribute bikes among stations by driving trucks according to their gut instinct. However, this method may not always be accurate which results in unnecessary relocation of bikes and increases expenses. So, accurately predicting future rental demand at each station is crucial for both bike-sharing operators and users. In this paper, we investigated how various contextual information can help to predict bike rental demand in the following days. Due to various mathematical properties of machine learning algorithms, the collected contextual information may provide different predictions accuracy. Therefore, we studied four different modeling approaches: Autoregressive Integrated Moving Average, Random Forest and Category Boosting, and Long-Short Term Memory. By measuring the model's correctness of prediction, it is determined that CatBoost is the best-performing model compared to others for one day rental demand prediction. But for more than the one-day prediction, LSTM performs better. Moreover, we found that contextual information such as geography has a significant impact on the model prediction.