From a multi-period perspective: A periodic dynamics forecasting network for multivariate time series forecasting

Abstract

To achieve more accurate multivariate time series (MTS) forecasting, it is crucial to extract temporal features of individual univariate series and interdependencies among multiple variables that evolve over time (i.e., dynamic interdependencies). Many MTS manifest inherent periodic fluctuations, which are highly significant for effective modeling. However, existing methods for modeling periodic features mainly focus on multi-period temporal features of univariate series and have not sufficiently addressed complex interdependencies in multiple periods, including dynamic dependencies among variables within the same period (Intra-period) and across different periods (Inter-period). Furthermore, methods for modeling dynamic interdependencies are insufficient for mining the inherent multi-periodicity of MTS. Thus, collaboratively capturing both intra-period and inter-period interdependencies remains challenging. To address above challenges, this paper introduces the Periodic Dynamics Forecasting Network (PDFNet) for modeling multi-period dynamic interdependencies of MTS. We design a periodic feature extraction module that utilizes frequency domain analysis to identify the multi-period features of MTS. The multi-period temporal networks module is designed to capture temporal features within and across periods from a multi-period perspective. To capture intra-period and inter-period dynamic dependencies among multiple variables, we propose a gated periodic recurrent unit and a gated graph structure learning module to construct dynamic graphs and then effectively learn intra-period and inter-period information through dynamic graph convolution networks. Extensive experiments on multiple MTS datasets have demonstrated our superior performance compared with state-of-the-art methods.