INVNET Inverted TimesNet for Multivariate Time Series Forecasting

Recent studies have revealed that TimesNet, a widely used task-generalized time series model, does not exhibit significant advantages when handling multivariate input sequences, and its efficiency is further reduced on certain datasets. The architecture of TimesNet embeds variables representing potential delay events and different physical measurements into the same token within a single time step, which may limit the model's ability to learn variable-centric representations effectively. To address these limitations and fully leverage the potential of the TimesNet framework for multivariate time series inputs, this paper proposes a novel model, InvNet. Rather than discarding the core idea of transforming 1D data into 2D representations and using 2D convolution kernels for feature extraction, InvNet introduces an inverted dimensionality structure and replaces the depth-separable convolution mechanism with the TimesBlocks module, significantly enhancing computational efficiency. To validate the performance of InvNet, extensive experiments were conducted on four real-world datasets (M4, Weather, Traffic, and Electricity), with comprehensive comparisons against both a baseline model and a state-of-the-art model. Experimental results demonstrate that the proposed InvNet model achieves outstanding performance across all datasets.