PaperlayerView exact revision on arXiv

Source-linked AI summary

Frequency-aware forecasting for short-term typhoon gust prediction

Xuefei Wang, Tingyi Liu, Heng Zhang, Shengjun Zhang

arXiv:2608.25604v1cs.LG

TL;DR

Typhoon gust forecasting remains difficult because extreme fluctuations are non-stationary and multi-scale, while existing approaches may miss rapid gust dynamics. WDANet combines shift-invariant stationary wavelet decomposition with frequency-aware dual-branch modeling to separate atmospheric trends from gust fluctuations. Across the reported experiments, it performs best at short lead times, including higher accuracy than ECMWF-HRES in the short term and more accurate extreme-event gust prediction.

  • Problem

    Gust forecasting under typhoon conditions is difficult because extreme wind fluctuations are non-stationary and multi-scale, and existing models can struggle to capture both long-term trends and rapid local variations.

  • Method

    WDANet integrates shift-invariant stationary wavelet decomposition and frequency-aware learning to separately model smooth atmospheric evolution and high-frequency gust dynamics.

  • Results

    WDANet achieves superior gust forecasting accuracy against other machine learning models and outperforms ECMWF-HRES in short-term forecasting, with more than 50% RMSE and MAE reductions at certain horizons.

  • Takeaways & Limitations

    WDANet shows potential for offshore wind power operation, disaster warning, and risk mitigation through improved short-term gust forecasting.

  • Takeaways & Limitations

    The study uses ERA5 reanalysis at 0.25°×0.25° resolution with approximately 5 days of data latency, limiting direct real-time operational use and finer-scale representation.

Abstract

from arXiv · show

Accurate gust forecasting under typhoon conditions remains challenging due to the highly non-stationary and multi-scale characteristics of extreme wind fluctuations. Existing deep learning models often struggle to simultaneously capture long-term trends and rapid local variations, resulting in degraded performance during extreme events. We propose WDANet, a frequency-aware forecasting framework that integrates stationary wavelet decomposition, a Feature-wise Linear Modulation (FiLM) strategy, and a dual-branch encoder-decoder architecture, enabling separate modeling of trend and fluctuation components. Taking the offshore regions of the Western Pacific in China as an example, we conduct fine-grid wind gust prediction research. The results demonstrate that WDANet shows advantages for short lead times under the experimental setting across a 24-h forecasting horizon and achieves higher prediction accuracy than ECMWF-HRES within the first 6 h. During extreme wind events, WDANet more accurately captures gust peaks and attains the best RMSE and MAE performance. These results highlight its potential for offshore wind power operation, disaster warning, and risk mitigation.

1 Introduction

Typhoon gusts are operationally critical but difficult to forecast because they are intermittent, non-stationary, and span multiple temporal scales. WDANet addresses these challenges with shift-invariant wavelet decomposition and frequency-aware modeling of atmospheric trends and gust fluctuations.

  • Motivation: Gusts impose larger aerodynamic loads than mean winds and can trigger shutdowns, fatigue, or catastrophic turbine failures.Operational safety therefore depends directly on gust characteristics.
  • Motivation: Gust forecasting is difficult because localized turbulence, convection, rapid energy transfer, intermittency, abrupt transitions, and non-stationarity complicate prediction.These dynamics differ from the smoother evolution of mean wind speed.
  • Research gap: Most existing studies target mean wind speed, so models optimized for averaged quantities can miss transient gust peaks relevant to turbine safety.This creates a mismatch between common forecasting targets and operational needs.
  • Research gap: Wavelet-based methods may reduce temporal resolution, become sensitive to time shifts, and overlook cross-scale interactions when components are treated independently.These limitations are especially problematic for aligning abrupt gust events.
  • Proposed approach: WDANet uses non-decimated, shift-invariant SWT to preserve temporal resolution and alignment of transient gust structures across decomposition levels.This design is intended to capture abrupt wind changes more faithfully.
  • Proposed approach: WDANet separates smooth atmospheric evolution from rapid gust-induced fluctuations through frequency-aware learning of low-frequency trends and high-frequency residual dynamics.The framework is presented as a dual-frequency strategy for gust forecasting.

2 Results

WDANet delivers strong short-term gust forecasts and robust performance during extreme typhoon events, while ablations support the value of its frequency-aware architecture. Its advantages are strongest when jointly modeling trends, rapid fluctuations, and extreme gust intensity.

  • Comparison methods analysis: 0.84 m/s at 1 h and 2.79 m/s at 24 h were WDANet’s RMSE values, with 5.6%–6.1% lower error than iTransformer across lead times.Against Autoformer, error reductions reached 18.9%–52.8%.
  • Comparison methods analysis: 1.85 m/s was WDANet’s final-step MAE, improving on iTransformer’s 2.03 m/s and Autoformer’s 2.51 m/s.WDANet maintained the lowest MAE among the data-driven models.
  • Comparison methods analysis: 56.5% lower RMSE than ECMWF-HRES at 1 h demonstrates WDANet’s short-term advantage, although ECMWF-HRES better captures long-term atmospheric evolution as lead time increases.The reported short-term advantage applies particularly within the first 6 h.
  • Ablation Study: 2.28 RMSE with one-level sym4 decomposition was best, whereas two- and three-level sym4 increased RMSE to 2.59 and 2.51.The results indicate that excessive decomposition may reduce effective temporal information for sequence modeling.
  • Ablation Study: 2.92 RMSE without SWT versus 2.28 for the complete model shows that wavelet decomposition and adaptive frequency responses both contribute to performance.The fixed-SWT variant achieved 2.38 RMSE, between the complete model and the no-SWT variant.
  • Ablation Study: 2.33 and 2.32 RMSE for approximation-only and detail-only variants show that accurate gust forecasting requires joint modeling of trends and rapid variations.Global-only and local-only attention also underperformed the complete model, with RMSE values of 2.34 and 2.39.
  • Performance under extreme typhoon conditions: 7.02 m/s was WDANet’s lowest 24-h averaged MAE under extreme typhoon conditions, below iTransformer’s 7.10 m/s and Autoformer’s 9.46 m/s.WDANet reduced error relative to Autoformer by 25.8% and achieved a 3.87 m/s overall MAE across all typhoon-affected cases.

3 Discussion

WDANet provides more accurate short-term typhoon gust forecasts than existing deep learning models and ECMWF-HRES in the reported experiments, while its scope is constrained by reanalysis-data latency and limited spatial information. Future work targets real-time inputs, physical constraints, geographic transfer, and improved extreme-peak prediction.

  • Discussion: WDANet achieves superior RMSE, MAE, and RAcc performance compared with existing deep learning models, with more than 50% RMSE and MAE reductions versus ECMWF-HRES at certain horizons.The reported evaluations cover typhoon cases, extreme events, and Typhoon Yagi.
  • Discussion: WDANet has superior short-term predictive capability, whereas ECMWF-HRES shows stronger long-term stability from physical constraints imposed by atmospheric governing equations.The paper attributes WDANet’s short-term advantage to direct nonlinear learning and notes possible nonlinear error amplification during autoregressive prediction.
  • Discussion: ERA5’s 0.25° × 0.25° resolution may miss finer-scale wind variations, and its approximately 5-day latency prevents direct real-time operational deployment.Operational use would require replacing historical reanalysis inputs with near-realtime data streams.
  • Discussion: Future research may combine numerical weather prediction outputs with WDANet, adapt it across regions through transfer learning, and use data augmentation or cost-sensitive learning.These directions are proposed to improve long-term forecasting, extreme-peak prediction, and geographic applicability.

4.1 Dataset and preprocessing

The study uses hourly, 0.25° × 0.25° ERA5 meteorological data and constructs continuous monthly samples around 400 typhoon events. Chronological organization and missing-value filling preserve usable temporal sequences for the regional forecasting experiment.

  • Dataset: ERA5 data provide hourly meteorological variables on a 0.25° × 0.25° grid, including pressure, temperature, humidity, and 10 m wind gust.Relative humidity is computed as actual vapor pressure divided by saturation vapor pressure.
  • Dataset: 400 typhoon events affecting the study region from 1960 to 2025 are identified for sample construction.Hourly reference gust data are extracted at station-corresponding grid points.
  • Preprocessing: Each event contributes a continuous monthly time series covering pre-typhoon, typhoon, and post-typhoon phases rather than isolated extreme segments.This design captures both stable background conditions and typhoon-driven variations.
  • Preprocessing: Samples are kept in chronological order, while missing values are filled forward, then backward at sequence beginnings, and finally with zeros if needed.The preprocessing preserves the temporal organization of the dataset.

4.2 Task formulation

The wind gust forecasting task is formulated as multi-step prediction: historical multivariate meteorological reanalysis data are mapped to future wind gust values over a specified horizon.

  • Task formulation: Given meteorological reanalysis data from the past M time steps, the task predicts wind gust evolution across a forecasting horizon of length H.This defines wind gust forecasting under typhoon conditions as a multi-step time-series problem.
  • Task formulation: At each time step, the atmospheric state includes pressure, temperature, dew point temperature, relative humidity, and the 10 m wind gust.The variables are represented as a multivariate atmospheric-state vector.
  • Task formulation: The model learns a nonlinear mapping from the historical sequence to the predicted wind gust sequence over the next H time steps.The output is an H-dimensional future gust sequence.

4.3 Learnable stationary wavelet transform module

The learnable SWT module decomposes input sequences into temporally aligned low- and high-frequency components without downsampling, then sends them to dedicated forecasting branches. Adaptive sym4-initialized filters and single-level decomposition are supported by the reported analyses and ablation results.

  • SWT design: SWT decomposes multivariate input sequences into frequency-specific components while preserving all temporal samples and shift invariance.Unlike DWT, SWT eliminates downsampling, avoiding shift sensitivity and preserving transient gust alignment.
  • SWT design: Dilated low-pass and high-pass filters are applied by one-dimensional convolution, preserving sequence length and temporal alignment across frequency components.At level s, 2^s − 1 zeros are inserted between adjacent original filter coefficients.
  • Learnable filters: The method initializes learnable filters with the Symlet-4 basis and optimizes them during training to adapt spectral selectivity to wind gust signals.The initialization supplies a physically meaningful frequency prior before adaptation.
  • Dual-branch architecture: A single decomposition level produces approximation A and detail D components that feed two dedicated forecasting branches for slow atmospheric evolution and rapid gust fluctuations.The architecture directly specializes the branches by frequency content.
  • Representation analysis: The adaptive SWT produces more structured block-like correlation patterns than the original meteorological variables.The comparison is visualized through correlation matrices before and after SWT processing.
  • Ablation: The ablation study finds that m = 1 consistently performs best, while deeper decomposition does not improve forecasting performance.The reported conclusion supports using a single decomposition level.
  • SWT role: The resulting decomposition separates slowly varying atmospheric evolution from transient gust fluctuations for subsequent multi-timescale forecasting.This is the stated role of the adaptive SWT module in the forecasting pipeline.

4.4 Frequency-aware encoder-decoder architecture

WDANet separates low-frequency atmospheric evolution from high-frequency gust fluctuations through dual encoder-decoder branches, with frequency-specific attention and FiLM-based cross-frequency modulation. The branch outputs are then additively reconstructed into the final gust forecast.

  • Frequency decomposition: SWT separates the input into low-frequency component A and high-frequency component D for independent encoder-decoder processing.The low-frequency branch models large-scale atmospheric evolution, while the high-frequency branch models transient gust fluctuations.
  • Dual-branch encoding: Two parameter-independent bidirectional LSTM encoders specialize in the temporal dynamics of the trend and fluctuation branches.Their encoded representations are passed to attention-based decoders for wind gust prediction.
  • Frequency-specific attention: Global attention captures long-range dependencies in the trend branch, whereas local attention with a learnable temporal prior targets transient fluctuations in the high-frequency branch.The local mechanism restricts attention to a dynamically predicted neighborhood while retaining content relevance.
  • Cross-frequency modulation: FiLM modulation lets the low-frequency branch generate adaptive parameters that recalibrate the high-frequency representation according to atmospheric evolution.The modulation uses element-wise multiplication and adjustment of the high-frequency features.
  • Output reconstruction: The final prediction is the sum of the low-frequency and high-frequency branch outputs, reflecting complementary trend and fluctuation modeling.This additive reconstruction combines large-scale atmospheric evolution with transient local variations.

5 Tables

The tables report RMSE, MAE, relative accuracy, wavelet sensitivity, and component ablations across the 24-hour wind-gust forecasting evaluation. Captions and accompanying descriptions identify lower error and higher relative accuracy as preferable, with comparisons spanning model performance and design choices.

  • Model comparison: Table 1 reports RMSE results for different models in the wind gust forecasting task.RMSE is evaluated across forecast horizons, with lower values preferred.
  • Model comparison: Table 2 reports MAE results for different models in the wind gust forecasting task.MAE is evaluated across forecast horizons, with lower values preferred.
  • Model comparison: Table 3 reports relative accuracy results for different models in the wind gust forecasting task.Relative accuracy is evaluated across forecast horizons, with higher values preferred.
  • Component ablation: Table 5 reports ablations of wavelet configurations, frequency branches, attention mechanisms, and FiLM-based cross-frequency modulation.The full model is compared with variants including w/o SWT, w/o FiLM, fixed SWT, Only A, Only D, and single-attention variants.

Declarations

The declarations report no applicable funding, ethics approval, or publication consent requirements, and no competing interests. The study also provides data-availability information for ERA5 and ECMWF-HRES data.

  • Declarations: The authors declare no funding, competing interests, ethics approval, or consent requirements applicable to this study.The declaration lists each item as not applicable or absent.
  • Data availability: ERA5 reanalysis and ECMWF-HRES forecast data were obtained from the Copernicus Climate Data Store and ECMWF, respectively.The declaration includes data-availability links for both sources.
Loading 2608.25604v1…