Long-term & short-term bike sharing demand predictions using contextual data

Bike Sharing Systems (BSSs) have gained popularity in the last two decades, becoming an integrated part of our mobility ecosystem. One of the major issues BSS operators struggle to deal with is forecasting the mobility demand, namely how many bikes are needed, when, and where. To deal with this challenge, this paper introduces a novel deep learning architecture that combines Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and contextual data to predict hourly demand in large-scale bike-sharing systems one day ahead. Specifically, the CNN layer captures spatio-temporal trends and is used to model long-term demand patterns, while the LSTM layer focuses on short-term mobility strategies. A separate Neural Network is used to include contextual data, such as weather data or special events. Predictions are provided for the next 24 hours at a frequency of one hour. The modular structure of the model ensures that the framework can be used in different situations and that individual components can be replaced. Similarly, additional contextual information can be included with limited effort. Results show that such a modular framework, where each module captures different dynamics (weather effects, long-term mobility patterns, short-term mobility patterns) outperforms the baseline models and also reduces their computational times.