You only adapt once: An adaptive transformer for dynamic multivariate time series forecasting across time-varying topologies and multi-patterns

Shuai Zhang, Jiyuan Xu, Wenyu Zhang, Yixiang Zhang, Chengjie Ni

Abstract: Highlights•Novel dynamic multivariate time series forecasting model is explored.•Dynamic encoder adapts seamlessly to time-varying topologies and missing values.•Dual-pattern pool captures spatiotemporal multi-patterns of periodical change.•Multi-scale decoder enhances forecasting precision via multi-pattern analysis.•New node-wise curriculum learning method boosts model training efficiency.