Pred-SF: A Precipitation Prediction Model Based on Deep Neural Networks

Area of Science:

• Meteorology
• Artificial Intelligence
• Data Science

Background:

• Accurate and efficient precipitation prediction remains a significant challenge in weather forecasting.
• Existing numerical weather forecasting and radar echo extrapolation methods have inherent limitations.
• High-precision meteorological data is available but requires advanced processing for effective precipitation prediction.

Purpose of the Study:

• To propose a novel precipitation prediction model, Pred-SF, that addresses the limitations of current methods.
• To leverage multi-modal meteorological data and a step-by-step prediction structure for improved accuracy.
• To enhance precipitation forecasting in specific target areas.

Main Methods:

• Developed the Pred-SF model featuring a self-cyclic and step-by-step prediction structure.
• Utilized a two-step approach: autoregressive spatio-temporal prediction using PredRNN-V2 and spatial information fusion.
• Integrated multiple meteorological data sources (ERA5) and precipitation measurement data (GPM).

Main Results:

• The Pred-SF model demonstrated strong precipitation prediction capabilities for continuous precipitation over 4 hours.
• Experimental results validated the effectiveness of combining multi-modal data for prediction.
• The stepwise prediction methodology within Pred-SF proved advantageous over comparative methods.

Conclusions:

• The Pred-SF model offers a promising advancement in precipitation prediction accuracy and efficiency.
• Combining multi-modal data and employing a stepwise prediction strategy are key to the model's success.
• Pred-SF provides a robust framework for enhancing precipitation forecasting in meteorological applications.