Fixformer: Air Pollution Forecast in Fixed Time Interval without Sequence-to-Sequence Structure

The Transformer-based models have been successfully applied in time series forecasting tasks. However, most of these models retain the sequence-to-sequence structure, which is widely used in Natural Language Processing (NLP) tasks. In this paper, we aim to determine whether the sequence-to-sequence structure can be replaced. Thus, we designed the Fixformer model, which changes the sequence-to-sequence structure to an N-to-T structure. We applied this model to an air pollution forecasting task and selected the forecast time interval T using Fast Fourier Transform (FFT) analysis. The experimental results indicate that our Fixformer achieves an equivalent forecast accuracy to the Transformer, and the Fixformer requires less training time.