TCLformer: Enhancing Multi-Scale Time Series Forecasting with Temporal Decomposition and Sparse Convolutional Attention

Real-world multivariate time series (MTS) often contain complex multi-scale temporal dependencies, and accurately capturing their short-term and long-term patterns remains a challenging task. Although Transformer-based models have made significant progress in this field in recent years, they still suffer from high computational overhead and limited ability to model long-term dependencies. Convolutional-based models, while computationally efficient, struggle to capture global dependencies and tend to overlook periodic and trend information in time series. To address this, we propose a predictive model that combines multi-scale adaptive convolutions with sparse convolutional attention mechanisms, named TCLformer. The model first performs trend-seasonal decomposition on the input time series to extract key intertwined temporal patterns; then, it uses parallel multi-scale convolutional modules to adaptively capture local features at different time scales. Finally, it introduces a sparse attention module based on causal convolutions, to efficiently model long-term dependencies. Experimental results on six public datasets demonstrate that TCLformer significantly outperforms existing state-of-the-art methods in terms of prediction accuracy, validating its effectiveness and superiority in multi-variable time series prediction tasks.