A Study of Upper Tropospheric Circulations over the Northern Hemisphere Prediction Using Multivariate Features by ConvLSTM

Spatiotemporal prediction on climate data is aiming to predict future spatial data by learning from prior spatial sequence data. In this paper, we are interested in a prediction of upper tropospheric circulations over the Northern Hemisphere by predicting a geopotential height at 300 hPa (Z300) variable. We proposed a predictive model by constructing an architecture with convolutional layers and deconvolutional layers and applied to convolutional long short-term memory (ConvLSTM) network. The results show that our model obtained root mean square error (RMSE) of 77.36 m (0.84% comparing to average Z300 value) in short-term prediction. While, a convolutional neural network (CNN) and a linear regression (LR) model obtained RMSE of 109.35 (1.19%) and 153.61 (1.67%), respectively. The ConvLSTM maintains RMSE even in long-term prediction. Furthermore, the prediction features’ investigation result shows that temperature at 300 hPa (T300) and self prior Z300 features are important for Z300 prediction.