Improving multi-scale short-term precipitation forecasting through frequency domain analysis and attention mechanisms

Area of Science:

• Meteorology and Atmospheric Science
• Artificial Intelligence and Machine Learning
• Data Science and Analytics

Background:

• Existing data-driven models struggle with short-duration precipitation events, lacking multi-scale characteristics and accurate intensity prediction.
• Improving forecasting accuracy often demands significant computational resources, posing a practical challenge.

Purpose of the Study:

• To develop a novel short-term precipitation forecasting model that addresses limitations in existing data-driven approaches.
• To quantify multi-scale precipitation distribution characteristics in the frequency domain.
• To enable rapid and precise prediction of hourly precipitation distributions over a 0-24 hour horizon.

Main Methods:

• Integration of a cross-channel multi-scale attention mechanism (EMA) with the TransUNet hybrid architecture.
• Quantification of multi-scale precipitation distribution characteristics in the frequency domain using wavelet transform.
• Development of a novel short-term precipitation forecasting model with reduced parameters.

Main Results:

• WE_TransUNet achieved optimal forecasting performance, outperforming four common deep learning models.
• Average Threat Score (TS) improved by 12% to 0.447, with an average Root Mean Square Error (RMSE) of 0.281.
• The EMA module demonstrated superior capability in capturing nonlinear time-series relationships across scales compared to conventional CBAM modules.

Conclusions:

• The proposed model effectively extracts multi-scale information from historical precipitation data with fewer parameters, mitigating deep learning model ambiguity.
• Wavelet transform aids in preserving high-frequency, localized severe convective phenomena in forecasts.
• This approach offers a novel optimization strategy for short-term precipitation forecasting, enhancing accuracy and efficiency.