Real-Time Rainfall Forecasts Based on Radar Reflectivity during Typhoons: Case Study in Southeastern Taiwan

Area of Science:

• Meteorology
• Atmospheric Science
• Data Science

Background:

• Accurate short-term rainfall forecasting is crucial for typhoon preparedness.
• Traditional methods often struggle with the spatiotemporal complexity of typhoon rainfall.

Purpose of the Study:

• To develop and evaluate a real-time rainfall forecasting system for typhoons.
• To optimize the use of Doppler radar data and ground-weather observations for prediction.

Main Methods:

• Utilized Doppler radar reflectivity from nine elevation angles and ground-weather data.
• Developed rainfall prediction models using a gated recurrent unit (GRU) neural network.
• Compared GRU performance against extreme gradient boosting and multiple linear regression benchmarks.

Main Results:

• The optimal radar elevation angle and merged data significantly improved short-term rainfall forecasting accuracy.
• The GRU neural network demonstrated superior performance in predicting rainfall 1, 3, and 6 hours before typhoon landfall.
• The system accurately predicted rainfall for typhoons Nepartak, Meranti, and Megi.

Conclusions:

• A real-time rainfall forecasting system integrating Doppler radar and weather station data is effective.
• The GRU neural network is a promising tool for accurate, short-term typhoon rainfall prediction.
• Optimizing radar data input enhances forecasting precision for extreme weather events.