MIANet: Multi-level temporal information aggregation in mixed-periodicity time series forecasting tasks

Abstract

Regular human activities generate a large number of time series with mixed periodicity that can reflect human behavior patterns and the societal working mechanism. When forecasting these time series, nonlinear neural networks often encounter some limitations, such as utilizing mixed-periodic patterns, balancing multi-level information, incorporating future vision, forecasting delays and scale insensitivity, which affect the forecasting accuracy. To address these problems, we propose the Multi-level Information Aggregation Network (MIANet), a novel neural network with four key characteristics: (i) a novel folded recurrent structure that dynamically updates the local and mini-local information at a global range in a compact manner; (ii) a new recurrent unit called Folded Convolution Aggregation Temporal Memory (FCATM) that extracts and aggregates neighbor-trends in local and mini-local data; (iii) a fusing decoder structure that promotes the sharing of forward–backward future information and adaptively adjusts relationships among adjacent points; and (iv) a new Skip-Autoregressive (SAR) linear strategy that addresses scale sensitivity issues. The SAR can be embedded as a plug-and-play component into other deep learning (DL) models. Compared with other baseline methods, MIANet obtains statistically significant improvements on six real-world datasets, as demonstrated by conducting two-sample t-tests, indicating that the MIANet can be applied to various predictive scenarios, such as road occupancy, electricity consumption, pedestrian flow and urban noise.