A Short‐Term Regional Precipitation Prediction Model Based on Wind‐Improved Spatiotemporal Convolutional Network. Issue 9 (7th September 2022)

• Record Type: Journal Article
• Authors: Qiu, Yunan; Lu, Zhenyu; Tang, Haibo
• Is Part Of: Earth and space science. Volume 9:Issue 9(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-09-07
• Subjects: short‐term precipitation prediction -- improved GCN -- GRU; Space sciences -- Periodicals; Geophysics -- Periodicals; 500.5
• DOI: 10.1029/2022EA002411
• Languages: English
• ISSNs: 2333-5084

Abstract

Abstract: Accurate precipitation prediction can help decision makers judge the trend of climate change and formulate more effective measures, and prevent flood and drought disasters. In this paper, we propose a short‐term regional precipitation prediction model based on wind‐improved spatiotemporal convolutional network. Among them, the improved graph convolution network integrates the effects of wind direction and geographic location at past moments to capture the spatial dependence, whilst the gated recurrent unit captures the temporal dependence by learning the dynamic changes of data. The spatio‐temporal memory flow module and attention module are added to capture spatial deformation and temporal variation more accurately, thereby better matching the physical properties of precipitation. The proposed model achieves better prediction results on real data sets. Experiments show that our method is better at extracting the spatio‐temporal information of precipitation data and capturing its time dependence and spatial correlation. Plain Language Summary: Deep‐learning technology has not been fully explored in regional short‐term precipitation prediction. The traditional graph convolution neural network does not consider the practical significance of wind direction in precipitation. Therefore, we introduce a novel short‐term regional precipitation prediction model based on wind improved spatiotemporal convolution network (ASS‐TGCN). Measured data of automatic meteorological station in Jiangsu Province, China have been utilized. Compared with the comparison model, our proposed model has achieved better performance in various indicators. Key Points: Meteorological data are processed to build a data set for automatic stations in Jiangsu province The improved graph convolution network takes considers effects of the wind direction at past moments and geographical location to capture the spatial correlation The proposed model is more in line with the physical characteristics of precipitation and suitable for precipitation prediction tasks