Multi-task supply-demand prediction and reliability analysis for docked bike-sharing systems via transformer-encoder-based neural processes

With the rise of sharing economy, bike-sharing systems (BSSs) have gained heated attention, and their operations require accurate prediction of bike usage. Although many deep learning methods have been exploited to predict bike usage, they generally provide point predictions of average bike usage, neglecting the stochasticity in BSSs. Due to the analytically explainable properties and linear computational costs with respect to data size, neural processes (NPs) have recently attracted increasing interest. An NP model learns a Gaussian process (GP) by mapping the input–output observations to a probabilistic distribution over functions. Each function is a distribution of the outputs given an input, conditioned on the arbitrary size of observed data. NPs provide probabilistic confidence in predicted results, which overcomes the point prediction issue faced by other models and provides insights for operational strategies in stochastic scenarios. This paper originally proposes a transformer-encoder-based NP (TENP) model to fit the distribution of bike usage in BSSs. To the best of our knowledge, this work is among the first to incorporate transformer encoders into NPs, enhancing the capability of extracting relevant information in a targeted manner. Based on the Citi Bike datasets in New York City, the TENP method is adopted in a multi-task learning task that simultaneously fits the number of pickups and returns. The proposed TENP model outperforms the conventional NP method and its extensions and prevalent machine learning models in terms of prediction accuracy. Armed with the probabilistic confidence provided by the TENP, reliability analysis is conducted, and thoughtful guidance is provided for bike-sharing operations, such as dynamic bike rebalancing.