Zhang et al. (2025) A novel spatiotemporal transformer network with multivariate fusion for short-term precipitation forecasting

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-11-27
• Authors: Kai Zhang, Guojing Zhang, Xiaoying Wang, Yu Zhu, Wu Li
• DOI: 10.1038/s41598-025-29415-2

Research Groups

• School of Computer Technology and Application, Qinghai University, Xining, China
• Qinghai Provincial Laboratory for Intelligent Computing and Application, Qinghai University, Xining, China
• School of Computer and Information Science, Qinghai Institute of Technology, Xining, China

Short Summary

This study proposes ST-MFTransNet, a novel spatiotemporal transformer network with multivariate fusion, to improve short-term precipitation forecasting by integrating diverse meteorological variables. The model significantly outperforms existing deep learning methods, achieving notable enhancements in detection probability and critical success index for 12-hour and 24-hour accumulated precipitation forecasts.

Objective

• To develop a novel spatiotemporal transformer network with multivariate fusion (ST-MFTransNet) that addresses the limitations of single-source data models in short-term precipitation forecasting by efficiently combining information from diverse meteorological variables.

Study Configuration

• Spatial Scale: Global, using a 5.625° resolution grid.
• Temporal Scale: Forecasts for subsequent 12-hour and 24-hour accumulated precipitation, using 12-hour and 24-hour cumulative input data. Training data from 1980–2011, validation from 2012–2013, and testing from 2014–2018.

Methodology and Data

• Models used: ST-MFTransNet (novel spatiotemporal transformer network with multivariate fusion), comprising a Multivariate Fusion Module (MFM) based on Omni-dimensional Dynamic Convolution (ODConv) and StarReLU, an encoder-decoder framework with a Transformer, a multi-scale convolution module (MCM), an Attention-Convolution Fusion module (ACF), and an Intersectional MultiLayer Perceptron (IMP). Comparative analysis against ViT, SimVP, TAU, ConvLSTM, and PredRNN.
• Data sources: WeatherBench dataset (regridded ERA5 data) at 5.625° resolution, including temperature, relative humidity, specific humidity, and the u- and v-components of wind speed. Precipitation data units converted to millimeters.

Main Results

• ST-MFTransNet consistently achieved superior performance across all evaluation metrics (MSE, MAE, FAR, POD, CSI) compared to baseline models (ViT, SimVP, TAU, ConvLSTM, PredRNN) for both 12-hour and 24-hour accumulated precipitation forecasts.
• For 12-hour accumulated precipitation forecast (threshold 15 mm), ST-MFTransNet showed enhancements in POD and CSI of 21.2% and 18.4%, respectively, relative to ViT.
• For 24-hour accumulated precipitation forecast (threshold 25 mm), ST-MFTransNet demonstrated improvements in POD and CSI of 10.3% and 10.7%, respectively, relative to ViT.
• Compared to the second-best attention-based model (TAU), ST-MFTransNet reduced MSE by 0.6052 and MAE by 0.0128 for 12-hour forecasts, and MSE by 1.2299 and MAE by 0.0121 for 24-hour forecasts.
• Ablation studies confirmed that both the Multivariate Fusion Module (MFM) and the Multi-scale Convolution Module (MCM) significantly contribute to the model's improved performance.

Contributions

• Proposed ST-MFTransNet, a novel spatiotemporal transformer network that integrates multivariate meteorological data for enhanced short-term precipitation forecasting.
• Introduced a Multivariate Fusion Module (MFM) utilizing Omni-dimensional Dynamic Convolution and attention mechanisms to effectively combine diverse meteorological variables, capturing complex inter-variable relationships.
• Designed an encoder-decoder framework incorporating a Transformer and a multi-scale convolution module (MCM) to extract both global and fine-grained local spatiotemporal features from the fused data.
• Demonstrated state-of-the-art performance in 12-hour and 24-hour accumulated precipitation forecasting, significantly outperforming existing deep learning models.

Funding

• Natural Science Foundation of Qinghai Province (No. 2023-ZJ-906M)
• National Natural Science Foundation of China (No. 62162053)
• High-performance computing center of Qinghai University

Citation

@article{Zhang2025novel,
  author = {Zhang, Kai and Zhang, Guojing and Wang, Xiaoying and Zhu, Yu and Li, Wu},
  title = {A novel spatiotemporal transformer network with multivariate fusion for short-term precipitation forecasting},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-29415-2},
  url = {https://doi.org/10.1038/s41598-025-29415-2}
}