Multi-resolution Patch-based Fourier Graph Spectral Network for spatiotemporal time series forecasting

Jiankai Zheng, Liang Xie, Haijiao Xu

Abstract: Highlights•Multi-resolution Analysis in MPFGSN: MPFGSN captures multi-scale features through multi-resolution patches, enabling the model to discern both long-term trends and short-term fluctuations. This approach improves the model’s ability to learn from complex spatiotemporal data, enhancing its forecasting robustness.•Efficient Graph Convolution with FGSN: FGSN utilizes Fourier transformation to perform graph spectral convolution in the spectral domain. This method is efficient and accelerates processing, particularly beneficial for handling large-scale graph data, which is crucial for accurate spatiotemporal forecasting.•Residual Fusion Mechanism in MPFGSN: MPFGSN incorporates a residual fusion mechanism that eliminates information redundancy by integrating features from different resolutions. This mechanism enhances the model’s feature extraction capabilities and improves its generalization across various forecasting scenarios, as confirmed by extensive experiments on public datasets.