MLPT: MLP-Enhanced Patch Transformer for Multivariate Time Series Forecasting

Multivariate time series forecasting has a wide range of applications in the fields of computer supported cooperative work. In recent years, deep learning models like transformers, have been widely used in time series forecasting. However, traditional transformers mainly capture short-term dependencies (point to point) within single-variable sequences and struggle with capturing cross-variable and multi-scale dependencies for long-term forecasting. To address this issue, we propose a novel transformer model called MLPT, an MLP-Enhanced Patch Transformer for Multivariate Time Series Forecasting. Our model extracts global features, reconnects them with patches and then calculates both inter and intra attenions, being capable of capturing dependencies in time series from both inter and intra variables, and models these dependencies globally, which makes our model outperform in long-term forecasting. We evaluate our method using seven real-world public datasets, and the experimental results demonstrate that our approach has improved over existing methods in the vast majority of scenarios, with an overall improvement of above 5% on the accuracy. At the same time, the time cost is reduced to one-fourth compared with MTGNN.