Segmented Frequency-Domain Correlation Prediction Model for Long-Term Time Series Forecasting Using Transformer

Abstract

Long-term time series forecasting has received significant attention from researchers in recent years. Transformer model-based approaches have emerged as promising solutions in this domain. Nevertheless, most existing methods rely on point-by-point self-attention mechanisms or employ transformations, decompositions, and reconstructions of the entire sequence to capture dependencies. The point-by-point self-attention mechanism becomes impractical for long-term time series forecasting due to its quadratic complexity with respect to the time series length. Decomposition and reconstruction methods may introduce information loss, leading to performance bottlenecks in the models. In this paper, we propose a Transformer-based forecasting model called NPformer. Our method introduces a novel multiscale segmented Fourier attention mechanism. By segmenting the long-term time series and performing discrete Fourier transforms on different segments, we aim to identify frequency-domain correlations between these segments. This allows us to capture dependencies more effectively. In addition, we incorporate a normalization module and a desmoothing factor into the model. These components address the problem of oversmoothing that arises in sequence decomposition methods. Furthermore, we introduce an isometry convolution method to enhance the prediction accuracy of the model. The experimental results demonstrate that NPformer outperforms other Transformer-based methods in long-term time series forecasting.