融合时空特征的GCN-LSTM西北地区沙尘天气预测模型研究
Spatiotemporal feature-based GCN-LSTM model for predicting sand-dust weather in northwest China
苏佳 1李高雅 1张新生1
作者信息
- 1. 西安建筑科技大学管理学院,西安 710055
- 折叠
摘要
针对以往沙尘天气预测算法中时空特征提取不足的问题,文中提出一种基于融合时空特征的图卷积和长短期记忆网络(GCN-LSTM)的沙尘天气预测模型.以西北地区为研究对象,利用城市之间的植被指数和距离构建邻接矩阵,通过图卷积网络(GCN)对空间特征、长短期记忆网络(LSTM)对时间特征进行提取,将特征融合后用于预测各个城市的沙尘天气.与GCN、LSTM、时空因果卷积神经网络(STCN)模型相比,文中提出的GCN-LSTM模型的准确率分别提高6％、8％、2％,且其接收者操作特征曲线(ROC)、ROC曲线下的面积(AUC)、精确度-召回度曲线(P-R)评价指标表现更优.文中研究为沙尘天气发生采取防范措施、减少损失提供借鉴意义.
Abstract
Aiming at the issue of insufficient spatiotemporal feature extraction in previous sand and dust weather prediction algorithms,this paper proposes a spatiotemporal feature prediction model based on graph convolution and long short-term memory network(GCN-LSTM)fused with spatiotemporal features.Taking Northwest China as the research object,the vegetation index and distance between cities are used to construct an adjacency matrix,spatial features are extracted by graph convolutional network(GCN),and long short-term memory net-work(LSTM)are used to extract temporal features,and the features are fused to predict the dust weather of each city.Compared with the GCN,LSTM,and spatiotemporal causal convolutional neural network(STCN)models,the accuracy of the GCN-LSTM model proposed in this paper is improved by 6％,8％and 2％,respectively,and its receiver operation characteristic curve(ROC),area under ROC curve(AUC),and accuracy-recall curve(P-R)evaluation indicators are better.It provides some references for taking preventive measures and reducing losses in sand and dust weather.
关键词
沙尘天气预测/图卷积网络/长短期记忆网络/时空特征Key words
sand and dust weather forecast/graph convolutional networks/long short-term memory networks/space-time characteristics引用本文复制引用
基金项目
陕西省自然科学基础研究计划(2021JQ-517)
中国博士后科学基金(2020M683433)
出版年
2024
国家科技期刊平台
NSTL
万方数据