Surface and high-altitude combined rainfall forecasting using convolutional neural network

• Published in: Peer-to-peer networking and applications Vol. 14; no. 3; pp. 1765 - 1777
• Main Authors: Zhang, Pengcheng, Cao, Wennan, Li, Wenrui
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.05.2021; Springer Nature B.V
• Subjects: Accuracy; Altitude; Artificial neural networks; Cluster analysis; Clustering; Communications Engineering; Computer Communication Networks; Economic forecasting; Engineering; Forecasting; High altitude; Information Systems and Communication Service; Machine learning; Meteorological data; Meteorology; Networks; Neural networks; Principal components analysis; Rainfall; Signal,Image and Speech Processing; Special Issue on P2P Computing for Deep Learning; Vector quantization; Weather stations; Convolutional neural network; Rainfall forecasting; Machine learning
• ISSN: 1936-6442; 1936-6450
• DOI: 10.1007/s12083-020-00938-x

Rainfall forecasting can guide human production and life. However, the existing methods usually have a poor prediction accuracy in short-term rainfall forecasting. Machine learning methods ignore the influence of the geographical characteristics of the rainfall area. The regional characteristics of surface and high-altitude make the prediction accuracy always fluctuate in different regions. To improve the prediction accuracy of short-term rainfall forecasting, a surface and high-Altitude Combined Rainfall Forecasting model (ACRF) is proposed. First, the weighted k-means clustering method is used to select the meteorological data of the surrounding stations related to the target station. Second, the high-altitude shear value of the target station is calculated by using the meteorological factors of the surrounding stations. Third, the principal component analysis method is used to reduce dimensions of the high-altitude shear value and the surface factors. Finally, a convolutional neural network is used to forecast rainfall. We use ACRF to test 92 meteorology stations in China. The results show that ACRF is superior to existing methods in threat rating (TS) and mean square error (MSE).