Adaptive wavelet decomposition and event-aware high-frequency modeling network for multivariate time series forecasting

Abstract

Multivariate time series (MTS) forecasting has wide applications in real-world domains such as traffic, weather, and ocean monitoring. However, real-world MTS data often exhibit multiscale non-stationary patterns. These patterns arise from the dynamic coupling between long-term trends and sudden local events, making accurate forecasting highly challenging. Existing methods primarily rely on global modeling or fixed decomposition strategies. However, such approaches fail to adapt to varying spatiotemporal data and diverse task contexts. They cannot effectively disentangle trend and event sequences in accordance with the characteristics of time series data. Moreover, they cannot model unstable high-frequency events. To address these issues, we propose an Adaptive Wavelet Decomposition and Event-aware High-frequency Modeling Network (AweHF). The model employs an Adaptive Wavelet Decomposition module (AWD) to decouple the original sequence into low-frequency trends and high-frequency events in a data-driven way, avoiding the limitations of fixed wavelet bases. Subsequently, we apply a lightweight multilayer perceptron (MLP) to capture long-term dependencies in the trend component. In addition, we design a Time Aggregation Network (TAN) and Dual-Source Personalized Graph Convolution (DSPGC) to jointly model the volatility and instability of the event component. Finally, the bidirectional interaction fusion mechanism is used to integrate the trend and event components to fully exploit their complementary advantages. We conducted extensive experiments on six real-world datasets from multiple domains, and our results demonstrate that AweHF consistently outperforms all state-of-the-art baselines, achieving an average MAE reduction of more than 3.8% across the datasets. Code is available at this repository: https://github.com/WangChengci/AweHF.