TVNet: A Novel Time Series Analysis Method Based on Dynamic Convolution and 3D-Variation

TL;DR

TVNet introduces a 3D-Embedding to reshape 1D time series into a 3D representation and applies dynamic convolution via stacked 3D-blocks to capture intra-patch, inter-patch, and cross-variable interactions. The approach preserves CNN efficiency while achieving state-of-the-art performance across long-term/short-term forecasting, imputation, classification, and anomaly detection, and demonstrates strong transferability. The work provides a practical time-series backbone with competitive computational costs and robustness, paving the way for large-scale pretraining and multi-scale patch extensions. Overall, TVNet offers a principled CNN-based alternative to Transformers/MLPs for versatile, efficient time-series analysis.

Abstract

With the recent development and advancement of Transformer and MLP architectures, significant strides have been made in time series analysis. Conversely, the performance of Convolutional Neural Networks (CNNs) in time series analysis has fallen short of expectations, diminishing their potential for future applications. Our research aims to enhance the representational capacity of Convolutional Neural Networks (CNNs) in time series analysis by introducing novel perspectives and design innovations. To be specific, We introduce a novel time series reshaping technique that considers the inter-patch, intra-patch, and cross-variable dimensions. Consequently, we propose TVNet, a dynamic convolutional network leveraging a 3D perspective to employ time series analysis. TVNet retains the computational efficiency of CNNs and achieves state-of-the-art results in five key time series analysis tasks, offering a superior balance of efficiency and performance over the state-of-the-art Transformer-based and MLP-based models. Additionally, our findings suggest that TVNet exhibits enhanced transferability and robustness. Therefore, it provides a new perspective for applying CNN in advanced time series analysis tasks.

Figures (19)

• Figure 1: In the context of time series analysis, the three-dimensional variation encompasses intra-patch, inter-patch, and cross-variable interactions.
• Figure 2: The overarching architecture of TVNet is constructed by stacking 3D-blocks in a residual way, which enables the capture of inter-patch, intra-patch, and cross-variable features from the time series 3D tensor.
• Figure 3: A univariate time series example is segmented into four distinct patches and the odd and even components specifically for Patch1 (ETTh1).
• Figure 4: The Time varying weight generation flow chart
• Figure 5: Model efficiency comparison under the setting of $L(\text{prediction length})$ =192/720 of ETTm2.