摘要
为提高宁夏回族自治区多流域范围内长期降水预报的准确度,本文以宁夏回族自治区10个3级流域为研究区域,提出含有自更新记忆细胞的自注意记忆-窥孔连接-卷积长短时序网络,采用整层水汽通量为降水预报因子,月降水为预测对象,选用回归与分类共9项评价指标,探究新方法的适用性。结果表明:相对于其他3种方法,NCMSAM-Peephole-ConvLSTM方法预测精度最高,MAE下降12.0%,RMSE下降10.0%,CORR提升6.2%,MSLE降低61.8%,NSE提升12.8%;阈值在25 mm以内时,NCMSAM-Peephole-ConvLSTM方法、SAM-Peephole-ConvLSTM方法的POD、CSI、HSS中位数均高于其他2种方法。4种方法中NCMSAM-Peephole-ConvLSTM方法预报准确度最高,其次为SAM-Peephole-ConvLSTM方法,ConvLSTM与ConvGRU方法效果较差。
There are 10 local watersheds at level 3 in Ningxia taken as research objects to improve the accuracy of long-term precipitation forecasting within multiple watersheds in the region.A self-attention memory-peephole connection-convolution long-short time series network containing self-updating memory cells is proposed in this paper.Taking the whole layer water vapor flux as the precipitation prediction factor and monthly precipitation as the prediction object,a total of 9 evaluation indicators including regression and classification are selected to explore the applicability of the new method.The results show that compared to the other three methods,NCMSAM Grapole ConvLSTM has the highest prediction accuracy,with a decrease of 12.0%in MAE,10.0%in RMSE,61.8%in MSLE as well as an increase of 6.2%in CORR and 12.8%in NSE.When the threshold is within 25 mm,the median values of POD,CSI and HSS for NCMSAM-Peephole-ConvLSTM and SAM-Peephole-ConvLSTM are higher than those of the other two methods.Among the four methods,NCMSAM-Peephole-ConvLSTM has the highest prediction accuracy,followed by SAM-Peephole-ConvLSTM.ConvLSTM and ConvGRU have poor performance.
作者
高嘉辉
钟德钰
GAO Jiahui;ZHONG Deyu(School of Civil Engineering and Water Resources,Qinghai University,Xining 810016,China;State Key Laboratory of Plateau Ecology and Agriculture,Qinghai University,Xining 810016,China;State Key Laboratory of Hydroscience and Engineering,Tsinghua University,Beijing 100084,China;Key Laboratory of Hydrosphere Sciences of the Ministry of Water Resources,Tsinghua University,Beijing 100084,China;Department of Hydraulic Engineering,Tsinghua University,Beijing 100084,China)
出处
《青海大学学报》
2023年第5期93-101,共9页
Journal of Qinghai University
基金
国家自然科学基金项目(91547204)。
关键词
降水预报
深度学习
自注意机制
卷积长短时序网络
precipitation forecasting
deep learning
self-attention mechanism
convolutional long-short time series network
