一种雷达回波外推的注意力融合和信息回忆的LSTM方法

AN ATTENTION FUSION AND INFORMATION RECALL LSTM METHOD FOR RADAR ECHO EXTRAPOLATION

关键词:  雷达回波外推  深度学习  注意力机制  时空长短期记忆网络

Key words:  radar echo extrapolation  deep learning  attention mechanism  space-time long short-term memory network

摘要:

临近天气预报是气象研究中的热点问题，雷达回波外推技术作为处理临近天气预报的有效手段，具有重要的应用价值。近年来，深度学习技术被应用于处理这一任务，但提高雷达回波外推的预报准确率仍然是一个具有挑战性的问题。在ST-LSTM网络基础上，本文提出一种AFR-LSTM网络，以进一步提高雷达回波外推的预报准确率。首先提出一种注意力融合的时空长短期记忆网络的方法，以关联更多的历史信息，保证信息在传递过程中能够充分关联，减少信息丢失。同时，考虑编码过程中信息丢失问题，在编码器与解码器之间构建信息回忆模块，进一步保存雷达回波预测细节。通过在真实的雷达回波数据集（2019—2021江苏气象雷达数据）上进行消融实验，AFR-LSTM整体效果较好。此外，对该雷达回波数据集进行对比实验，结果表明AFR-LSTM在雷达回波预测中评分函数临界成功指数(CSI)值为0.520 9、Heidke Skill Score (HSS)值为0.532 4，并且能较好地保留强回波和位置准确度，优于现有方法，证明了该方法能够获得更准确的预测准确度。

Abstract:

Nowcasting is a prominent area of research in meteorology, and radar echo extrapolation is an effective technique for generating nowcasts. In recent years, deep learning technology has been applied to this task, but improving the accuracy of radar echo extrapolation forecasting remains a challenge. Based on the ST-LSTM network, this paper proposes an AFR-LSTM network to enhance the accuracy of radar echo extrapolation forecasting. Firstly, an attention fusion method for a spatiotemporal long-short-term memory network is proposed to integrate more historical information, ensuring that information can be fully integrated during the transmission process and reducing information loss. Moreover, we address the issue of information loss in the encoding process by incorporating an information reminiscence module between the encoder and decoder, which helps preserve the details of radar echo prediction. Through ablation experiments conducted on a real radar echo dataset (2019—2021 Jiangsu Meteorological Radar Data), AFR-LSTM demonstrates strong overall performance. Comparative experiments on this radar echo dataset also reveal that AFR-LSTM achieves a critical success index (CSI) value of 0.520 9 and a Heidke skill score (HSS) value of 0.532 4 in radar echo prediction, effectively preserving strong echoes and ensuring accurate location prediction. These results outperform existing methods, demonstrating that our proposed method can achieve more accurate image prediction.