Open AccessCascading Models of CNN and GRU with Autoencoder Loss for Precipitation Forecast in Thailand
Author(s) -
Fuenglada Manokij,
Peerapon Vateekul,
Kanoksri Sarinnapakorn
Publication year - 2021
Publication title -
ecti transactions on computer and information technology
Language(s) - English
Resource type - Journals
ISSN - 2286-9131
DOI - 10.37936/ecti-cit.2021153.240957
Subject(s) - precipitation , autoencoder , autoregressive integrated moving average , meteorology , flood myth , climatology , environmental science , quantitative precipitation forecast , deep learning , computer science , artificial intelligence , geography , machine learning , time series , geology , archaeology
It is a crucial task to accurately forecast precipitation, especially rainfall in Thailand, since it relates to flood prevention and agricultural planning. In our prior work, we have presented a model based on deep learning approach; however, its performance is still limited due to two main issues. First, there is an imbalance issue, where most rainfall is zero or no rain because Thailand has short rainy season. Second, predicted rainfall is still underestimated since moderate and heavy rainfall cases barely occurs. In this paper, we propose an enhanced deep learning model to forecast rainfall in Thailand. Our model is a cascading of CNN and GRU along with exogenous variables, i.e., temperature, pressure, and humidity. There are two stages in our model. First, CNN is specialized for classifying rain and non-rain events. In this stage, an imbalanced issue is alleviated by applying “focal loss”. Second, GRU is responsible for forecasting rainfall. Its predicted range is lifted using “autoencoder loss”. The experiment was conducted on hourly rainfall dataset between 2012 and 2018 obtained from a public government sector in Thailand. The results show that our enhanced model outperforms ARIMA and CNN-GRU in terms on RMSE of most regions in Thailand.