WDANet, the model tested in the paper, had the lowest reported root mean square error (RMSE), an error score for forecasts, among the data-driven baselines. Its RMSE was 0.84 m/s one hour ahead and 2.79 m/s at 24 hours. The reported reductions were approximately 5.6% to 6.1% against iTransformer and 18.9% to 52.8% against Autoformer, depending on the forecasting step.
The advantage was strongest early
When the model was compared with ECMWF-HRES, the pattern depended on how far ahead the forecast reached. WDANet was reported to have higher accuracy within the first six hours. At the one-hour lead, its RMSE was 0.84 m/s, against 1.93 m/s for ECMWF-HRES, a 56.5% reduction. At longer horizons, ECMWF-HRES performed better.
Mean absolute error (MAE), another way of measuring forecast misses, pointed in the same direction among the machine-learning models. At the final forecasting step, WDANet's reported MAE was 1.85 m/s, compared with 2.03 m/s for iTransformer and 2.51 m/s for Autoformer. Those figures represented approximate improvements of 8.9% and 26.3%, respectively. CNN-LSTM's RAcc was about 0.01 higher, so WDANet did not lead on every reported measure.
The robustness analysis used block bootstrap resampling of individual typhoon events, with a fixed seed of 42 and 1,000 iterations. In that summary, WDANet had the lowest RMSE at 2.27 m/s, with a reported 95% confidence interval from 2.17 to 2.37 m/s. Its mean MAE was 1.51 m/s, the smallest among the comparison models.
Performance in severe cases
The reported advantage also appeared in an extreme-typhoon subset, although the errors were larger. Over a 24-hour average, WDANet's MAE was 7.02 m/s versus 9.46 m/s for Autoformer, a reported 25.8% lower error. Across all typhoon-affected cases, WDANet's overall MAE was 3.87 m/s.
One case study focused on Typhoon Yagi. Across the month-long evaluation, WDANet's reported MAE was 1.70 m/s and RMSE was 2.50 m/s, compared with 1.73 m/s and 2.57 m/s for iTransformer. During the most intense 24-hour period, WDANet's MAE was 6.82 m/s and RMSE was 8.14 m/s. The reported reductions versus Autoformer were about 37.6% for MAE and 36.0% for RMSE. WDANet's peak time error was nine hours; TimesNet had a smaller peak error but a higher MAE.
How the model was built and tested
WDANet uses stationary wavelet decomposition, a way of separating a time series into components, together with FiLM and two encoder-decoder branches so that trend and fluctuation are modeled separately.
The evaluation covered 400 typhoon events affecting the study region from 1960 to 2025. The researchers extracted hourly ERA5 reference gust data for the entire month of each event. Models used a 24-hour sliding input window, advancing one hour at a time. Missing values were imputed, and preprocessing settings were calculated only from training data before the validation and test sets were transformed.
The researchers also removed pieces of the architecture to compare reported scores. The full model and the one-level sym4 version each had an RMSE of 2.28. Two-level and three-level sym4 versions scored 2.59 and 2.51. A version without the stationary wavelet step scored 2.92, while one without FiLM scored 2.30.
A result with clear limits
The findings come with a practical caveat. The ERA5 input grid was 0.25 degree by 0.25 degree, which the authors say may miss finer-scale wind variations. The data also have approximately five days of latency, preventing the tested setup from directly supporting real-time operational deployment. The document is an arXiv version 1 preprint dated 26 Aug 2026.
Paper data and sources
Original title: Frequency-aware forecasting for short-term typhoon gust prediction
Authors: Xuefei Wang, Tingyi Liu, Heng Zhang, Shengjun Zhang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text