A Deep Learning Architecture for Long-Range Forecasting of Sea Surface Temperature Anomalies

Sea surface temperature (SST) is an important state variable in the climate system due to its strong influence on weather and seasonal climate. Accurate prediction of SST is essential for forecasting tropical cyclogenesis as well as the ENSO cycles. In this work, we propose to improve the forecast skills of SST anomalies using a novel machine learning approach to combine the monthly SST forecasts generated from the North American Multi-Model Ensemble (NMME) experiment. The ensemble consists of seasonal forecasts from multiple coupled climate models with lead time up to 12 months. Although the multi-model ensemble can produce reasonably accurate short-term probabilistic forecasts, the uncertainty increases significantly with the lead time. As different models have their own biases, our goal is to learn a nonlinear transfer function that will combine the ensemble model outputs in order to improve the point estimates of long-range forecasts of SST anomalies. Specifically, we extend a type of deep learning architecture commonly employed in time series prediction, the long short-term memory network, to effectively combine outputs from multiple models using a novel hierarchical architecture. We compared the proposed framework against several state-of-the-art baselines and and the results showed the effectiveness in improving longer-range forecasting accuracy. The approach has the potential to be used for post-processing other variables from multi-model ensemble predictions.