AppoloConv: Multi-Scale Frequency-Aware Convolutions for Robust Multivariate Time Series Forecasting

YangYouLiu, Suwan Yin, Han yunxiang, Yuankai Wu

Keywords: Time Series Forecasting, Convolutional Neural Networks, Adaptive Receptive Fields

Abstract: Time series forecasting requires models that balance expressive power with computational efficiency. While convolutional neural networks offer efficient temporal modeling, their inherent translation invariance often misaligns with the recency bias and non-stationary dynamics present in real-world time series. We propose ApolloConv, a convolutional architecture that enhances temporal inductive bias through integrated time–frequency modeling. ApolloConv incorporates (i) a multi-scale embedding stem that captures local-to-global patterns while emphasizing recent context, (ii) a lightweight spectral gating mechanism that modulates periodic components in the frequency domain while preserving phase coherence, and (iii) an adaptive dilated convolution block that prioritizes recent time steps through logarithmically scaled receptive fields. Together, these components enable effective handling of multi-scale seasonality, trend structures, and cross-variable dependencies with near-linear complexity. Extensive experiments on benchmark datasets demonstrate that ApolloConv consistently outperforms state-of-the-art CNN-based models such as TimesNet, TVNet, and ModernTCN across both short- and long-term forecasting settings, while matching or exceeding Transformer-based counterparts with significantly lower computational cost. ApolloConv provides a robust and efficient convolutional alternative for practical time series forecasting.

Primary Area: learning on time series and dynamical systems

Submission Number: 22766