CNformer: a convolutional transformer with decomposition for long-term multivariate time series forecasting

Improving long-term time series forecasting accuracy and efficiency is of great value for real-world applications. The main challenge in the long-term forecasting of multivariate time series is to accurately capture the local dynamics and long-term dependencies of time series. Currently, most approaches capture temporal dependencies and inter-variable dependencies in intertwined temporal patterns, which are unreliable. Moreover, models based on time series decomposition methods are still unable to capture both short- and long-term dependencies well. In this paper, we propose an efficient multivariate time series forecasting model CNformer with three distinctive features. (1) The CNformer is a fully CNN-based time series forecasting model. (2) In the encoder, the stacked dilated convolution as a built-in block is combined with the time series decomposition to extract the seasonal component of the time series. (3) The convolution-based encoder-decoder attention mechanism refines seasonal patterns in the decoder and captures complex combinations between different related time series. Owing to these features, our CNformer has lower memory and time overhead than models based on self-attention and the Auto-Correlation mechanism. Experimental results show that our model achieves state-of-the-art performance on four real-world datasets, with a relative performance improvement of 20.29%.