NS-Fast: A Non-Stationary Transformer Using Flexible Probabilistic Sparse Attention for Aviation Weather Forecasting

Accurate forecasting of non-stationary meteorological data is essential for aviation safety, particularly under complex tropical weather conditions. This study introduces the Non-stationary FlexSparse Transformer (NS-Fast), an incremental architectural advancement designed to address temporal non-stationarity in time series forecasting. NS-Fast integrates a De-Stationary ProbSparse Attention (DSProb) mechanism that combines the non-stationary modeling capability of De-Stationary Attention with the efficiency of adaptive ProbSparse Attention, offering flexible query importance measures beyond Informer’s KL-divergence approach. We conducted a three-stage evaluation on public datasets (Weather, ETTh1, ETTh2, ETTm2) and an Automatic Weather Observation System (AWOS) dataset. Results show NS-Fast consistently outperforms existing Transformer models, achieving up to 39% MSE and 20% MAE improvements over NST and Informer on ETTm2. On AWOS, the proposed approach demonstrated robust capability in forecasting severe wind events and multiple meteorological variables. Additionally, integrating Reversible Instance Normalization (RevIN) into NST primarily affects error normalization rather than overall forecasting accuracy, it introduces an oversmoothing effect that degrades temporal dependency modeling.