Improving Seasonal Forecast Using Probabilistic Deep Learning

The path toward realizing the potential of seasonal forecasting and its socioeconomic benefits relies on improving general circulation model (GCM) based dynamical forecast systems. To improve forecasts, it is crucial to set up benchmarks, clarify limitations posed by initialization errors, formulation deficiencies, internal climate variability. With huge costs in generating large ensembles, limited observations for verification, benchmarking diagnosing task proves challenging. Here, we develop a probabilistic deep learning-based statistical methodology, drawing wealth simulations enhance capability diagnosis. By explicitly modeling variability GCM differences, proposed Conditional Generative Forecasting (CGF) methodology enables bypassing barriers forecast, offers top-down viewpoint examine how complicated GCMs encode predictability information. We apply CGF global precipitation 2 m air temperature, unique data consisting 52,201 years simulation. Results show that can faithfully represent information encoded GCMs. successfully this learned relationship real-world achieving competitive performance compared forecasts. Using as benchmark, reveal impact insufficient spread sampling limits skill considered system. Finally, introduce different strategies composing ensembles using highlighting leveraging strengths multiple achieve advantgeous forecast.