Adaptive temporal feature extraction and prediction model optimization for extreme weather early warning

Abstract

Extreme weather events pose significant threats to human life, infrastructure, and socio-economic stability, necessitating accurate and timely forecasting methods. Traditional time-series prediction models, such as recurrent and convolutional architectures, face limitations in handling non-stationary signals and abrupt regime shifts inherent in meteorological data. To address these challenges, we propose an Adaptive Temporal Feature Extraction (ATFE) model that integrates three core innovations: (i) a multi-scale convolutional feature extractor to capture both short-term local fluctuations and long-term seasonal patterns; (ii) an adaptive attention and divergence-based mechanism to detect abrupt climate transitions and highlight critical precursors of extreme events; and (iii) a dynamic weighted optimization strategy to improve robustness against noise, outliers, and incomplete data. Experiments are conducted on four benchmark meteorological datasets— GHCN, ERA5, CMA, and WeatherBench—with comparisons against representative baselines including LSTM, Transformer, TCN, and Informer. Results demonstrate that ATFE consistently outperforms state-of-the-art models, achieving lower RMSE and MAPE while providing superior sensitivity to abrupt changes and stronger robustness under noisy and cross-regional scenarios. These findings suggest that ATFE is not only accurate but also reliable for real-world early-warning systems, offering practical value for operational meteorological forecasting and disaster risk reduction.