DLENSO: A Deep Learning ENSO Forecasting Model

Abstract

El Niño-Southern Oscillation (ENSO) phenomenon is the strongest signal in the interannual time scale of global climate, and has a significant impact on the global short-term climate (temperature, precipitation, etc.). Every year, researchers around the world would predict ENSO for the coming year, and have been studied new forecasting methods all the time, including numerical methods, statistical methods and deep learning methods. The existing deep learning methods are only for the ENSO index or single-point meteorological elements forecasting and rarely involve forecasting of specific regions. In this paper, we formulate a deep learning ENSO forecasting model (DLENSO) to predict ENSO through predicting Sea Surface Temperature (SST) in the tropical Pacific region directly. DLENSO is a sequence to sequence model whose encoder and decoder are both multilayered Convolutional Long Short-Term Memory (ConvLSTM), the input and prediction target of DLENSO are both spatiotemporal sequences. We explore the optimal setting of this model by experiments and report the accuracy on Niño3.4 region to confirm the effectiveness of the proposed method. Moreover, it can be concluded that DLENSO is superior to the LSTM model and deterministic forecast model, and almost equivalent to the ensemble-mean forecast model in the medium and long-term (4–12 months ahead) forecast. This model will pave a new way of predicting ENSO using deep learning technology.