MixNet: A scale-adaptive method for multivariate time series forecasting

Abstract

Time series forecasting is a critical task with widespread applications in industrial domains and daily life, including weather prediction, long-term energy consumption planning, and marketing analysis. Nevertheless, effectively extracting salient temporal patterns and exploring dependencies within multivariate time series remains a challenge. This paper focuses on multivariate time series forecasting, a common and pivotal issue in numerous analytical tasks. To address the complexity and high variability inherent in multivariate time series, we propose a scale-adaptive multi-head attention mechanism based on a hybrid mixture of experts network. Building on this mechanism, we develop MixNet, a novel architecture designed to achieve flexible feature extraction across diverse types of time series data. Furthermore, to tackle the difficulty in capturing inter-variable dependencies, we introduce a dedicated multivariate time series embedding (MTSE) scheme integrated with learnable positional encoding. This approach aims to comprehensively model the dependencies among variables, thereby enhancing overall forecasting performance. Experimental results demonstrate that MixNet outperforms several state-of-the-art methods on seven benchmark datasets from primary domains.