Data-Segmentation Prompt Based Continual Learning Framework for Online Spatio-Temporal Prediction

The rapid proliferation of physical sensors continuously generates massive spatio-temporal data streams, supporting critical applications such as transportation management and environmental monitoring. However, in realworld settings, the sensor network often evolves dynamically, where both the data distribution and graph structure change over time. These dynamics pose fundamental challenges to online spatio-temporal prediction, including catastrophic forgetting, high retraining cost, and efficiency degradation. To address these issues, we propose DepCL, a data-segmentation prompt–based continual learning framework from the perspective of spatio-temporal data management. DepCL partitions nodes into clusters according to distribution discrepancies and associates each cluster with lightweight learnable prompts, enabling efficient model adaptation through prompt retrieval and fine-tuning. Furthermore, a Gradient Projection–based Knowledge Balance mechanism theoretically rectifies gradient updates to preserve historical representations while maintaining model plasticity. Additionally, to mitigate representation conflicts caused by node imbalance, we introduce a Density-Aware Sample Alignment strategy that adaptively strengthens learning on newly arrived nodes. Extensive experiments on three real-world datasets demonstrate that DepCL achieves superior accuracy and efficiency compared with state-of-the-art methods.