ENSO-GTC: ENSO Deep Learning Forecast Model With a Global Spatial-Temporal Teleconnection Coupler

Abstract

Continuing the tropical Pacific multivariate air-sea coupler proposed by us before, we design the Global spatial-temporal Teleconnection Coupler (GTC), which is modeled to discover the latent teleconnections among global sea surface temperature (SST). To this end, Pacific, Indian, and Atlantic oceans are divided into small ocean patches that compose a dynamics graph, in which the adjacent relationships are artificially constructed by prior knowledge and the non-adjacent relationships are learned from the data by deep learning methods. Based on GTC, an El Niño-Southern Oscillation (ENSO) deep learning forecast model (ENSO-GTC) is established, where probability graph convolution layers are designed to learn spatial-temporal teleconnection, that is, non-adjacent relationships in the dynamics graph. A loss function with a graph total variations penalty term is remarkably proposed to maintain physical consistency. We tune ENSO-GTC to the optimal, which has Niño3.4 index correlation skills of 0.79/0.66/0.51 at 6-/12-/18-month forecasts with iterative strategy, and above 0.6 within 20-month forecasts with direct strategy, outperforming the other state-of-the-art models. We explore the forecast skill of ENSO-GTC on the effective forecast lead time, improvements of persistence barrier, and analysis of forecast errors. Moreover, we find that what GTC has learned exactly matches multiple ENSO theories. The oscillations of Kelvin and Rossby waves make a 6-month lagged correlation on Pacific SST. The north and south Pacific meridional modes (NPMM and SPMM) are strongly linked to the evolutions of ENSO and should be monitored more. The teleconnections between equatorial Indian/Atlantic and Pacific are very important and usually have 2-month/8-month lagged correlations before ENSO.