PatchTCN: Patch-Based Transformer Convolutional Network for Times Series Analysis

This study introduces a novel model for multivariate time series forecasting and imputation, called PatchTCN. This hybrid model combines the global and local processing capabilities of Transformer and Convolutional Neural Network(CNN), respectively. Developed from the PatchTST framework, which converts input time series data into a series of patches, our research addresses PatchTST's limitations in capturing intricate local dynamics and its fixed-length patch constraint. By integrating the innovative InterPatch-Conv and MSPatch-Conv modules, our model enhances local feature modeling and adapts to variable-length patches, providing a deeper understanding of time series data. Experimental results across various datasets have demonstrated that PatchTCN outperforms existing models in both forecasting and imputation tasks, proving its effectiveness in handling global dependencies and local temporal variations. This fusion of Transformer and CNN methodologies not only sets a new benchmark for time series analysis but also paves the way for further research into combining architectural strengths.