[ACG50-P02] Climate Prediction by Using Periodic Convolutional Recurrent Network
キーワード：Climate prediction, Periodicity, Periodic-CRN, Convolutional recurrent network
Machine learning based prediction methods have been investigated to many climate predictions such as precipitation forecasting, and temperature prediction. However, most machine learning researches on this domain do not consider periodic patterns, which are important for climate data that has strong periodicity. For example, the positive and negative phases of North Atlantic Oscillation (NAO) are repeatedly every year that can indicate seasonal over the Northern Hemisphere. Features from such periodicity may improve prediction accuracy.
In order to consider periodicity, we adopt the periodic convolutional recurrent network (Periodic-CRN) to our research. The idea of this model is to store and load the hidden states that learned by convolutional recurrent network (CRN) as periodic representation. The stored periodic representation is loaded and combined with the current prediction output. The model also has the attention mechanism that is used to weigh multiple periodic representations (e.g., monthly and yearly representation). In this work, we extend the idea of the periodic attention mechanism component to includes nearby time interval (month) representation. The group of nearby periodic representation is considered instead of using only a single periodic time interval. Then we use convolutional long short-term memory (ConvLSTM) as CRN.
We applied the proposed method to upper tropospheric circulations over the Northern Hemisphere and the land surface variable, which are geopotential height at 300 hPa (Z300) and temperature at 2 meter height from the surface (T2M), respectively. These data have a dominant periodic pattern that repeats every year. The proposed method obtained the average root mean square error (RMSE) of Z300 short-term prediction from 2016 to 2018 of 62.48 meters (0.68% compared to average Z300 value). On the other hand, ConvLSTM, convolutional neural network (CNN), and linear regression (LR) obtained average RMSE of 79.24, 113.76, and 125.25 meters, respectively. The proposed method also outperforms in T2M prediction that obtained average RMSE of 1.79 degrees Celsius (°C) while ConvLSTM, CNN, and LR obtained average RMSE of 2.54, 4.84, and 3.87 °C, respectively. The result has shown that the proposed method that considers data’s periodicity can improve prediction accuracy, which is a significant improvement from conventional methods.
In order to consider periodicity, we adopt the periodic convolutional recurrent network (Periodic-CRN) to our research. The idea of this model is to store and load the hidden states that learned by convolutional recurrent network (CRN) as periodic representation. The stored periodic representation is loaded and combined with the current prediction output. The model also has the attention mechanism that is used to weigh multiple periodic representations (e.g., monthly and yearly representation). In this work, we extend the idea of the periodic attention mechanism component to includes nearby time interval (month) representation. The group of nearby periodic representation is considered instead of using only a single periodic time interval. Then we use convolutional long short-term memory (ConvLSTM) as CRN.
We applied the proposed method to upper tropospheric circulations over the Northern Hemisphere and the land surface variable, which are geopotential height at 300 hPa (Z300) and temperature at 2 meter height from the surface (T2M), respectively. These data have a dominant periodic pattern that repeats every year. The proposed method obtained the average root mean square error (RMSE) of Z300 short-term prediction from 2016 to 2018 of 62.48 meters (0.68% compared to average Z300 value). On the other hand, ConvLSTM, convolutional neural network (CNN), and linear regression (LR) obtained average RMSE of 79.24, 113.76, and 125.25 meters, respectively. The proposed method also outperforms in T2M prediction that obtained average RMSE of 1.79 degrees Celsius (°C) while ConvLSTM, CNN, and LR obtained average RMSE of 2.54, 4.84, and 3.87 °C, respectively. The result has shown that the proposed method that considers data’s periodicity can improve prediction accuracy, which is a significant improvement from conventional methods.