13:45 〜 15:15
[MIS06-P01] Machine Learning Prediction of Precipitation in Metro Manila, Philippines
★Invited Papers
キーワード：降雨、機械学習、短期予測
Accurate prediction of the extreme weather, such as torrential rainfall, linear rainband, downburst, tornado, etc., is difficult even with today's state-of-the-art observational and numerical forecast technologies. Highly localized rainfall area and rapid temporal development are the major factors that make forecasting difficult. In recent years, application of the machine learning methods to the weather forecast is also intensively studied. Lin et al. [2020] proposed SA-ConvLSTM (Self-Attention Convolutional Long-Short Term Memory), which extends ConvLSTM [Shi et al., 2015] with a self-attention mechanism [Vaswani et al., 2017]. Lin et al. [2020] reported that SA-ConvLSTM was able to make a prediction with higher accuracy than other machine learning models. However, there are only a few studies on the prediction of the meteorological parameters using SA-ConvLSTM. Furthermore, there are no studies using the meteorological data at the localized and rapidly-developed rainfall events, that is a characteristic of extreme rainfall events. Thus, the purpose of this study is to evaluate the prediction performance of SA-ConvLSTM training with the localized rainfall and other weather parameter data and to verify how SA-ConvLSTM learns the background meteorological fields when it predicts the rainfall distributions and amounts via the model’s attention maps.
In order to conduct the short-term prediction of rainfall, our research group installed the automated weather and lightning observation system (P-POTEKA) since 2017 in Metro Manila, the Philippines. This P-POTEKA network can provide spatially high-resolution weather and lightning data and is suitable for detecting the highly-localized and rapidly-developed rainfall events in Metro Manila. By using P-POTEKA data obtained in the period from October 2019 to November 2020, a spatiotemporal dataset of more than 300 rainfall events and associated meteorological parameters was created. As a first step, equally gridded data were created using the Gaussian Process Regression interpolation method for the training of SA-ConvLSTM. Secondly, a superposed epoch analysis was conducted, and the feature of the temperature decrease and humidity increase preceding the more than 40 mm/h rainfalls by 1-2 hours was found. Thirdly, SA-ConvLSTM training with different combinations of weather parameters was conducted. It is found that SA-ConvLSTM trained with rainfall, temperature and humidity showed the highest prediction accuracy, which indicates that SA-ConvLSTM learns the feature of temperature and humidity variations towards rainfall. It is also found that SA-ConvLSTM trained with only rainfall data showed the lowest prediction accuracy. From these results, SA-ConvLSTM procedures trained with only rainfall and with rainfall, temperature and humidity were applied for the case studies. SA-ConvLSTM trained with rainfall, temperature and humidity data improved the prediction accuracy by calculating the higher attention scores for the area of the temperature decrease and humidity increase. The SA-ConvLSTM was able to accurately predict the occurrence of the rainfall from 1 hour before in some cases. In contrast, SA-ConvLSTM trained with only rainfall data tended to calculate high attention scores everywhere and failed to accurately predict the rainfalls in many cases. From these case studies, it is concluded that SA-ConvLSTM is able to learn the background meteorological rainfall field and to improve the prediction accuracy. However, SA-ConvLSTM fails to accurately predict in the rare cases that temperature and humidity changes simultaneously occur with rainfall in the same time scale.
In order to conduct the short-term prediction of rainfall, our research group installed the automated weather and lightning observation system (P-POTEKA) since 2017 in Metro Manila, the Philippines. This P-POTEKA network can provide spatially high-resolution weather and lightning data and is suitable for detecting the highly-localized and rapidly-developed rainfall events in Metro Manila. By using P-POTEKA data obtained in the period from October 2019 to November 2020, a spatiotemporal dataset of more than 300 rainfall events and associated meteorological parameters was created. As a first step, equally gridded data were created using the Gaussian Process Regression interpolation method for the training of SA-ConvLSTM. Secondly, a superposed epoch analysis was conducted, and the feature of the temperature decrease and humidity increase preceding the more than 40 mm/h rainfalls by 1-2 hours was found. Thirdly, SA-ConvLSTM training with different combinations of weather parameters was conducted. It is found that SA-ConvLSTM trained with rainfall, temperature and humidity showed the highest prediction accuracy, which indicates that SA-ConvLSTM learns the feature of temperature and humidity variations towards rainfall. It is also found that SA-ConvLSTM trained with only rainfall data showed the lowest prediction accuracy. From these results, SA-ConvLSTM procedures trained with only rainfall and with rainfall, temperature and humidity were applied for the case studies. SA-ConvLSTM trained with rainfall, temperature and humidity data improved the prediction accuracy by calculating the higher attention scores for the area of the temperature decrease and humidity increase. The SA-ConvLSTM was able to accurately predict the occurrence of the rainfall from 1 hour before in some cases. In contrast, SA-ConvLSTM trained with only rainfall data tended to calculate high attention scores everywhere and failed to accurately predict the rainfalls in many cases. From these case studies, it is concluded that SA-ConvLSTM is able to learn the background meteorological rainfall field and to improve the prediction accuracy. However, SA-ConvLSTM fails to accurately predict in the rare cases that temperature and humidity changes simultaneously occur with rainfall in the same time scale.