What question did this study set out to answer?

This research aims to improve weather forecasting accuracy for environmental applications by addressing non-stationarity in climate data.

April 26, 2026Open Access

Frequency-adaptive deep learning for multi-horizon weather forecasting in environmental monitoring applications

Key Points

This research aims to improve weather forecasting accuracy for environmental applications by addressing non-stationarity in climate data.
Proposed the Weather-oriented Frequency Adaptive Network (WFAN) using frequency-domain normalization.
Decomposed input sequences with Fourier transform to select dominant frequencies for non-stationary pattern identification.
Conducted experiments on benchmark datasets like Weather, ETTh1, and ECL, comparing WFAN to existing normalization methods.
WFAN significantly reduced mean squared error (MSE) in weather forecasts across multiple backbone networks.
Achieved superior performance against existing methods in most evaluation settings.
Statistical tests confirmed reliability and significant improvements in data stationarity.

Abstract

Accurate weather forecasting is essential for environmental monitoring systems, supporting critical applications in air quality prediction, ecosystem management, and disaster early warning. The inherent non-stationarity of climate time series, characterized by trend drift and seasonal evolution, poses significant challenges to forecasting accuracy. Existing instance normalization methods rely on time-domain statistics and assume unchanged non-stationary patterns between inputs and outputs, limiting their effectiveness in capturing dynamic environmental variations. We propose the Weather-oriented Frequency Adaptive Network (WFAN), a frequency-domain normalization framework that decomposes input sequences via instance-level Fourier transform, adaptively selects dominant frequency components as non-stationary patterns, and employs frequency residual learning to enhance data stationarity. A pattern-adaptive module explicitly predicts output frequency components to model pattern evolution over prediction horizons. Comprehensive experiments on multiple benchmark datasets, including Weather, ETTh1, and ECL, demonstrate that WFAN achieves substantial MSE improvements across four backbone networks (DLinear, FEDformer, Informer, SCINet), consistently achieving competitive or superior performance compared with state-of-the-art normalization methods across the majority of evaluation settings. Augmented Dickey-Fuller tests confirm strong stationarity enhancement, while efficiency analysis shows significant parameter reduction compared to existing methods. Statistical significance tests further validate the reliability of the observed improvements. This study provides an effective framework for environmental monitoring applications requiring accurate multi-horizon weather forecasting, with important implications for climate-sensitive decision-making in environmental management, agriculture, and disaster preparedness.

Bookmark

View Full Paper