ACENet: Adaptive correlation-enhanced network for multivariate time series forecasting

Abstract

A multitude of practical applications necessitate the utilization of multivariate time series forecasting techniques, including the issuance of extreme weather warnings and the formulation of energy consumption plans. However, time series data frequently display intricate intra- and inter-series correlations, rendering modelling and forecasting particularly challenging due to these complex dependencies. The comprehension and representation of these multi-level interactions represent a fundamental research challenge, one that is also of paramount importance in numerous application domains. The extant literature has a restricted focus on capturing correlations within periodic time intervals at disparate time scales and between these intervals. To address these challenges, we propose the Adaptive Correlation-Enhanced Network (ACENet). The model begins by extracting multiple significant period lengths through Fast Fourier Transform (FFT) and segmenting the time series accordingly. At each temporal scale, three dedicated correlation matrices - capturing feature-wise correlations within periods, timestamp-wise correlations within periods, and cross-period correlations respectively - work in concert to enhance periodic pattern learning. The framework then employs an adaptive weighting mechanism to dynamically balance intra-period and inter-period correlations, ultimately generating the final prediction through this hierarchical integration of multi-scale temporal dependencies. Finally, experiments on several real-world datasets demonstrate the effectiveness of ACENet on MST datasets.