深度学习方法在强降水临近预报技术中的应用及其改进思考

Thoughts on application and improvement of deep learning in severe precipitation nowcasting technology

如何有效将基于预报员经验的、不受物理公式约束的强降水天气特征作为先验知识融入到深度学习模型,是提升强降水临近预报准确率面临的重要挑战之一。首先,简述了目前主流的深度学习模型以及一些已开展的针对深度学习模型可解释性的探索成果;然后,对目前常用的几种强降水临近预报方法进行了简单介绍;最后,通过对深度学习方法在强降水临近预报技术中改进的思考,提出了将先验知识融入到深度学习模型的可能途径,具体包括:将一些经验特征进行量化或自动识别后作为模型输入,将预先提取的特征作为模型标签,在模型架构中嵌入先验知识的量化编码,结合可求导的检验指标设计损失函数对模型进行优化;同时指出,需要提高对强降水临近预报产品的主客观评估的一致性,以便能更好地确定强降水临近预报技术的改进方向。

How to integrate effectively the synoptic characteristics of severe precipitation events based on the experience of forecasters that is not constrained by physical formulas into the deep learning model as a priori information is one of the important challenges to improve the nowcasting accuracy of severe precipitation. First, we describe briefly the out-of-the-state deep learning models and some studies about the interpretability of deep learning models. Second, several widely-used methods of nowcasting for severe precipitation are briefly introduced.Finally, based on the consideration of the improvement of deep learning algorithm in the nowcasting technology of severe precipitation, we propose the possible way to integrate the prior information into the deep learning models. It includes quantifying or automatically identifying some empirical features as model inputs, taking the pre-extracted features as model labels, embedding the quantitative code of the prior knowledge in the model architectures, and optimizing the model by designing loss functions based on the differentiable test indicators. At the same time, improving the consistency of the subjective and objective evaluation of nowcasting products of severe precipitation can better determine the direction of the improvement of nowcasting technology of severe precipitation.