2passformer: 강한 상관관계의 다변량 시계열 데이터를 위한 효과적인 예측 모델

초록

The widespread deployment of IoT devices have led to an increase in generation and collection of multivariate time series. As these time series become more correlated—due to sensors recording inter-dependent events—forecasting models are required to effectively capture inter-variable dependencies and provide accurate predictions. In this work, we propose 2passformer, a novel forecasting model that incorporates (1) a decomposition mechanism to disentangle complex input sequences and (2) variable-wise bias vectors to discriminate correlated variables. Extensive experiments demonstrate that 2passformer achieves state-of-the-art performance on strongly correlated multivariate time series datasets, showing clear advantages in effectively modeling inter-variable relationships with high accuracy.

키워드