A radar echo extrapolation prediction method and system fusing multi-source meteorological data

Abstract

This invention discloses a radar echo extrapolation prediction method and system that integrates multi-source meteorological data to solve short-term precipitation prediction within 0-2 hours, involving the interdisciplinary fields of short-term meteorological forecasting and artificial intelligence. The embodiments of this invention improve upon the ST-LSTM model to create the MRMDST-LSTM model. Through multi-source data fusion, a teacher-student model training strategy using knowledge distillation is employed to enhance generalization ability. During training, a custom triple loss function (distillation loss, time-step and intensity-based weighted loss, and physical constraint loss from Z-R relationship embedding) is used to optimize prediction accuracy for long-term and strong echo regions. Finally, with a time step of 6 minutes, the first 10 frames of processed data are input, and the subsequent 20 frames of radar echo prediction results are decoded. This invention provides a radar echo extrapolation prediction method and system that integrates multi-source meteorological data, achieving accurate short-term precipitation prediction within 0-2 hours, improving the accuracy of locating heavy precipitation areas and the stability of long-term predictions, and enhancing the model's generalization ability.

Images