Shingo Zenkoji, Taichi Tebakari, Kazutoshi Sakakibara
Received 2022/03/10, Accepted 2022/07/01, Published 2022/09/30

Shingo Zenkoji¹⁾, Taichi Tebakari²⁾, Kazutoshi Sakakibara³⁾

1) Nippon koei Corporation, Japan
2) Department of Civil and Environmental Engineering, Chuo University, Japan
3) Department of Electronics and Information Engineering, Toyama Prefectural University, Toyama, Japan

Using deep learning to identify meteorological factors has enabled optimal predictions of Thailand’s seasonal precipitation two months in advance. A combination of surface temperature and pressure, specific humidity, and wind speed (zonal and meridional components) was tested. Examining each combination of meteorological factor has created optimal input data for seasonal precipitation forecasts. In addition, the hyperparameters of each model were calculated by Bayesian optimization. Predictive model performance tended to be better when the weight for pressure was higher, while a higher weight for specific humidity reduced predictive performance. Finally, visualization of the positive neuron values in all the coupled layers of the first layer showed that the regions with the highest frequency of occurrence were the El Niño monitoring areas such as the “Indian Ocean Basin Wide” (IOBW) and “NINO WEST”.

[Full Text]

Copyright (c) 2022 The Author(s) CC-BY 4.0