5:15 PM - 6:30 PM
[AHW21-P02] Prediction process in the seasonal forecast model of monthly precipitation by global climate data using deep learning
Keywords:Seasonal forecast, Deep learning, Monthly precipitation, Climate factor, Thailand
Thailand, a country in Southeast Asia, has a clear distinction between the dry season (November-April) and the rainy season (May-October). As a result, many natural disasters such as droughts occur during the dry season and tropical storms and floods occur during the rainy season, and their frequency and intensity are increasing. The GCM model using physical equations is often used as a long-term prediction method for precipitation. However, the prediction by the GCM model is highly uncertain and cannot be quantitatively predicted. The purpose of this study is to develop a lunar precipitation prediction model after 2 months in the upper Chao Phraya River basin, Thailand, from global climate value data using deep learning.
The used data were five types of global monthly average climate values (surface temperature, atmospheric pressure, specific humidity, soil moisture content, precipitation), latitude, longitude, and land mask obtained from JRA55. The study period was 1958 / 1-1988 / 12, the verification period was 1989 / 1-2009 / 12, and the examination period was 2010 / 1-2020 / 9.
In this study, we developed a model for predicting monthly precipitation two months later using the Deep Convolutional Neural Network (DCNN), which is one of the deep learning algorithms. The input data is a total of 4 types of global climate data (surface temperature, pressure, specific humidity, soil moisture) and 3 types of geographic data (Lat, Lon, Land mask) 2 to 9 months before the forecast month. Output data is monthly precipitation after 2 months. For the purpose of reducing the calculation cost and suppressing overfitting, the input data was average pooled using a 3 × 3 window. The DCNN model consists of one convolutional layer and two fully connected layers. In the convolution layer, the global data at each point is compressed into one feature (neuron) by the filter. The hyperparameters of the deep learning method were optimized using Bayesian optimization.
Next, as a result of predicting the monthly precipitation after 2 months under the above-mentioned conditions, the MAE was 41 mm and the RMSE was 53 mm during the test period. When the amount of precipitation by JRA55 was large, the predicted precipitation was underestimated, and when the amount of precipitation by JRA55 was small, the predicted amount of precipitation was overestimated (threshold value of about 300 mm). However, from 2011 to 2016, the fluctuations in the predicted monthly precipitation in June and July showed the same behavior as the fluctuations in the monthly precipitation by JRA55. Therefore, it was suggested that this prediction model has prediction skills. We visualized the first fully connected layer of the developed monthly precipitation prediction model. The positive and negative values of neurons in this layer are related to the occurrence and suppression of precipitation phenomena. Looking at the distribution of the neuron values at each point output from the first fully connected layer of the DCNN model during the test period, the maximum neuron values in the continental regions are a large positive value was confirmed, and a large negative value was confirmed for the minimum neuron value in the waters.
Finally, as a result of developing a monthly precipitation prediction model in the upper reaches of the Chao Phraya River in Thailand from global climate value data using deep learning, the monthly precipitation prediction model is difficult to predict quantitatively. It was confirmed that he had predictive skills. By visualizing this precipitation prediction model, it was found that this prediction model has the characteristic that the precipitation phenomenon is caused by the meteorological conditions of the continental area and the precipitation phenomenon is suppressed by the meteorological conditions of the sea area.
The used data were five types of global monthly average climate values (surface temperature, atmospheric pressure, specific humidity, soil moisture content, precipitation), latitude, longitude, and land mask obtained from JRA55. The study period was 1958 / 1-1988 / 12, the verification period was 1989 / 1-2009 / 12, and the examination period was 2010 / 1-2020 / 9.
In this study, we developed a model for predicting monthly precipitation two months later using the Deep Convolutional Neural Network (DCNN), which is one of the deep learning algorithms. The input data is a total of 4 types of global climate data (surface temperature, pressure, specific humidity, soil moisture) and 3 types of geographic data (Lat, Lon, Land mask) 2 to 9 months before the forecast month. Output data is monthly precipitation after 2 months. For the purpose of reducing the calculation cost and suppressing overfitting, the input data was average pooled using a 3 × 3 window. The DCNN model consists of one convolutional layer and two fully connected layers. In the convolution layer, the global data at each point is compressed into one feature (neuron) by the filter. The hyperparameters of the deep learning method were optimized using Bayesian optimization.
Next, as a result of predicting the monthly precipitation after 2 months under the above-mentioned conditions, the MAE was 41 mm and the RMSE was 53 mm during the test period. When the amount of precipitation by JRA55 was large, the predicted precipitation was underestimated, and when the amount of precipitation by JRA55 was small, the predicted amount of precipitation was overestimated (threshold value of about 300 mm). However, from 2011 to 2016, the fluctuations in the predicted monthly precipitation in June and July showed the same behavior as the fluctuations in the monthly precipitation by JRA55. Therefore, it was suggested that this prediction model has prediction skills. We visualized the first fully connected layer of the developed monthly precipitation prediction model. The positive and negative values of neurons in this layer are related to the occurrence and suppression of precipitation phenomena. Looking at the distribution of the neuron values at each point output from the first fully connected layer of the DCNN model during the test period, the maximum neuron values in the continental regions are a large positive value was confirmed, and a large negative value was confirmed for the minimum neuron value in the waters.
Finally, as a result of developing a monthly precipitation prediction model in the upper reaches of the Chao Phraya River in Thailand from global climate value data using deep learning, the monthly precipitation prediction model is difficult to predict quantitatively. It was confirmed that he had predictive skills. By visualizing this precipitation prediction model, it was found that this prediction model has the characteristic that the precipitation phenomenon is caused by the meteorological conditions of the continental area and the precipitation phenomenon is suppressed by the meteorological conditions of the sea area.