Yang et al. (2026) WinG-LSTM: a precipitation nowcasting model integrating swin transformer and LSTM

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-02-14
• Authors: Xu Yang, Changyong Zheng, Yating Wu, Xu Chu, Jun Zhang
• DOI: 10.1007/s00704-026-06079-0

Research Groups

• School of Electronic and Information Engineering, Anhui Jianzhu University, Hefei, 230601, China

Short Summary

This paper introduces WinG-LSTM, a novel deep learning model for precipitation nowcasting that integrates Swin Transformer with a gated Multi-Layer Perceptron into PredRNN's recurrent units. The model addresses limitations of traditional CNN-RNN approaches by effectively capturing long-range spatiotemporal dependencies, demonstrating superior accuracy in predicting precipitation intensity and spatial extent on two real-world radar datasets.

Objective

• To develop a precipitation nowcasting model that overcomes the limitations of existing Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) methods in capturing long-range spatial dependencies and accurately predicting severe precipitation scenarios, by integrating Swin Transformer and a gated mechanism into a recurrent neural network architecture.

Study Configuration

• Spatial Scale: Radar echo images covering a 101 km x 101 km region, with a resolution of 101 x 101 pixels, zero-padded to 128 x 128 pixels for processing.
• Temporal Scale: Precipitation nowcasting for 0-2 hours. The model uses 5 input radar echo images (30 minutes of observation history, 6-minute intervals) to predict 10 future images (1-hour forecast).

Methodology and Data

• Models used: WinG-LSTM (integrating Swin Transformer with a SwiGLU-based Multi-Layer Perceptron and PredRNN's Spatiotemporal LSTM units). Compared against ConvLSTM, TrajGRU, PredRNN, PredRNN++, IDA-LSTM, and MogaNET.
• Data sources:
  • CIKM2017 dataset: 10,000 samples (8,000 training, 2,000 validation, 4,000 test) from a 101 km x 101 km Shenzhen-centered region.
  • Shanghai Radar dataset: 2,060 samples (1,442 training, 206 validation, 412 test) from the Shanghai Meteorological Bureau (2015-2018).
  • Both datasets consist of 15 radar echo images per sample, 101 x 101 pixels, 6-minute temporal resolution, pixel values 0-255.

Main Results

• WinG-LSTM consistently achieved optimal or leading performance across all evaluation metrics (POD, HSS, CSI, MAE, MSE, SSIM) and precipitation thresholds (5 dBZ, 20 dBZ, 40 dBZ) on both CIKM2017 and Shanghai Radar datasets.
• On the CIKM2017 dataset, WinG-LSTM improved POD by 0.14% over IDA-LSTM at 5 dBZ, and by 1.3% (20 dBZ) and 13% (40 dBZ) over MogaNET.
• HSS improved by 3.6% and CSI by 2.9% compared to TrajGRU at 5 dBZ on CIKM2017. At 40 dBZ, WinG-LSTM's HSS was 19.1% higher than MogaNET, and its CSI was 23.9% higher than IDA-LSTM.
• The model demonstrated superior prediction of echo boundary delineation and intensity, particularly for severe precipitation regions, and maintained superior performance effectively over extended lead times (up to 60 minutes).
• Ablation studies confirmed the efficacy of both the Swin Transformer and SwiGLU-based MLP modules in enhancing prediction accuracy.
• Statistical significance of performance improvements was confirmed with p-values far below 0.01 in paired-sample t-tests for CSI at the 40 dBZ threshold against all baseline models.
• While WinG-LSTM incurred higher computational overhead (parameter count and FLOPs per frame), its single-frame prediction time (tens of milliseconds on NVIDIA 4090 GPUs) is well within real-time requirements for precipitation nowcasting.

Contributions

• Proposed the WinG-LSTM model, which deeply integrates Swin Transformer with PredRNN at the state update layers to efficiently extract spatiotemporal precipitation features, overcoming the limitations of local receptive fields in convolutions and capturing global-scale spatial dependencies.
• Introduced a gated mechanism (SwiGLU-based Multi-Layer Perceptron) into the Swin Transformer module to augment feature representation capabilities, highlighting critical precipitation features while suppressing irrelevant interference.
• Rigorously evaluated the model's effectiveness and superiority on the CIKM2017 dataset and the Shanghai Radar dataset, demonstrating a better balance between prediction accuracy and inference efficiency, particularly for severe precipitation events.

Funding

• Anhui Provincial Natural Science Foundation

Citation

@article{Yang2026WinGLSTM,
  author = {Yang, Xu and Zheng, Changyong and Wu, Yating and Chu, Xu and Zhang, Jun},
  title = {WinG-LSTM: a precipitation nowcasting model integrating swin transformer and LSTM},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-026-06079-0},
  url = {https://doi.org/10.1007/s00704-026-06079-0}
}