Multi-Scale Periodic Residual State Space Model for Time Series Forecasting

Jiqing Wang, Nianping Qin, Yao Zhou

Abstract: Time series forecasting is vital across applications such as weather prediction and electricity theft detection. However, the complexity of time series data, characterized by varied patterns across sampling scales and intricate spatiotemporal dependencies, presents significant forecasting challenges. To address these issues, we propose MPRS, a multi-scale periodic residual state-space model for time series forecasting. MPRS introduces a Multi-scale Temporal Feature Learning Block for effective feature extraction, composed of a spatiotemporal learning layer (TSL) and a dimension expansion layer (DimExpand). The TSL utilizes Mamba (a selective state space model) and CNN to capture spatial dependencies and local temporal features, while the DimExpand extends the time dimension to enhance the learning of global temporal information. Experimental results on benchmark datasets validate the effectiveness of MPRS in enhancing forecasting performance comparing to other competitive approaches, demonstrating its potential as a robust solution for time series forecasting tasks.

External IDs:dblp:conf/icic/WangQZ25