TimeCNN: Refining inscross-variable interaction on time point for time series forecasting

Time series forecasting is widely applied across various domains. Transformer-based models have shown considerable potential in modeling interactions across time and variables. However, we observe that the cross-variable correlations in multivariate time series exhibit multifaceted characteristics, including both positive and negative correlations, and evolve dynamically over time. These aspects are not adequately captured by existing Transformer-based models. To address this issue, we propose the TimeCNN model, which refines cross-variable interactions to improve time series forecasting. The key innovation of TimeCNN is its timepoint-independent approach, where each time point is assigned an independent convolution kernel. This allows each time point to have its own model for capturing relationships among variables. Building upon this, TimeCNN effectively manages both positive and negative correlations and adapts to the dynamic evolution of variable relationships over time. Extensive experiments on 12 real-world datasets demonstrate that TimeCNN consistently outperforms state-of-the-art models. Notably, our model significantly reduces computational requirements by approximately 60.46% and parameter count by about 57.50% while achieving inference speeds 3 to 4 times faster than the benchmark iTransformer model. The code will be publicly released on GitHub upon acceptance of the paper.