WSSNet: A Multivariate Time Series Forecasting Framework Incorporating Serial Correlations

As a hot topic of the time series analysis field, multivariate time series forecasting has always abstracted extensive research interest. Recent studies concentrate to explore the complex dependencies within time series data, especially the intra-series and inter-series correlations. However, exploring multivariate correlations within a target time series across different time scales is still an under-researched area. To bridge this gap, this paper introduces WSSNet (a deep learning model designed to capture the window, series, and subsequence correlations). This model uses frequency domain analysis, multi-head attention mechanism, and two graph convolution modules to capture complex dependencies in time series, including the multivariate intra-series, inter-series and inter-subsequence correlations across different time scales. WSSNet adopts FFT (Fast Fourier Transform) to effectively extract the periodic features of time series and break them down into various time scales. The model uses a multi-head attention mechanism to accurately reveal intra-series correlations. Additionally, WSSNet introduces two adaptive graph convolution modules to consider the correlations between different series and between subsequences at various time scales. To further uncover hidden patterns in the time series data, this paper introduces a decomposition-based model architecture to separately extract and do with the seasonal and trend features of the data. We conduct extensive experiments with the WSSNet model on eight real-world datasets. Compared with state-of-the-art methods, WSSNet demonstrates its effectiveness in multivariate time series forecasting.