In recent years, battery swapping services have experienced rapid growth, positioning themselves as a significant alternative to traditional charging stations. One determinant of their success lies in the ability to continuously refine the battery swapping station network, i.e., redistributing batteries across stations, to meet the fluctuating demands of electric scooters. However, achieving such network refining is far from straightforward and presents two major challenges. First, for regions where stations have been recently deployed, the lack of historical data makes it difficult to accurately predict user demand. Secondly, effective battery scheduling is inherently complex, as it requires meticulous coordination across all stations to dynamically balance users’ fluctuating long-term demand. To tackle these challenges, we propose dRefine, a novel data-driven battery swapping station network refinement system. Specifically, to address the first challenge, we develop a station-level spatiotemporal representation-guided conditional diffusion model, which leverages data from regions with established networks to predict demand in regions lacking historical data. For the second challenge, we develop a demand-oriented deep reinforcement learning model that dynamically optimizes battery scheduling strategies. By continuously learning from real-time demand patterns and operational feedback, it ensures efficient and adaptive battery redistribution across the entire network. We evaluate dRefine using a real-world dataset encompassing 388 stations, 41358 batteries, and 108574 electric scooter users. Extensive experimental results demonstrate that our method consistently outperforms state-of-the-art approaches by an average of 29.28%.

Recently, Massive Open Online Courses (MOOCs) have become a major online learning methodology for millions of people worldwide. However, the dropout rates from several current MOOCs are high. Usually, dropout prediction aims to predict whether a learner will exhibit learning behaviors during several consecutive days in the future. Therefore, the information related to the learning behaviors of a learner in several consecutive days should be considered. After in-depth analysis of the learning behavior patterns of the MOOC learners, this study reports that learners often exhibit similar learning behaviors on several consecutive days, i.e., the learning status of a learner for the subsequent day is likely to be similar to that for the previous day. Based on this characteristic of MOOC learning, this study proposes a new simple feature matrix for keeping information related to the local correlation of learning behaviors and a new Convolutional Neural Network (CNN) model for predicting the dropout. Extensive experimental validations illustrate that the local correlation of learning behaviors should not be neglected. The proposed CNN model considers this characteristic and improves the dropout prediction accuracy. Furthermore, the proposed model can be used to predict dropout temporally and early when sufficient data are collected.