Diagram showing the three barriers in seasonal forecast and the Conditional Generative Forecasting methodology developed to tackle these three barriers.
Three barriers in seasonal forecast (a), and the Conditional Generative Forecasting (CGF) methodology developed to tackle these three barriers (b). The three barriers are 1) difficulty for initializing the forecasting model due to limited observation or state shocks, 2) model formulation error due to imperfect representation of unresolved climate processes, and 3) insufficient sampling of forecast spread due to chaos. The CGF methodology is illustrated via a case example of seasonal forecasting for global precipitation or 2 m air temperature (Y), using predictor of previous season upper ocean thermal state (X). This CGF methodology explicitly models the internal climate variability using stochastic gradient variational Bayes (b center), and simulates the climate model formulation differences using the entity embedding technique (b top). Convolutional and transposed convolutional neural networks serve as building blocks of the CGF methodology. Credit: Pan et al. [2022]
Editors’ Highlights are summaries of recent papers by AGU’s journal editors.
Source: Journal of Advances in Modeling Earth Systems

Skillful seasonal forecast benefits a broad range of societal sectors. However, current dynamical seasonal forecast systems face three stubborn challenges. First, we may start from a “wrong” place: it is difficult to align a model’s state to match observations at forecast starting time, due to limited coverage of observations, or the state shock problem. Second, we may follow “wrong” paths: the forecasting models cannot perfectly simulate the climate state evolution trajectory, due to poor representation of unresolved climate processes. Third, we may have not tried all possibilities: using limited (~10) forecasting ensemble members, we may not have sufficiently sampled the plausible future climate states. The huge observational, computational, and intellectual cost for resolving these challenges have significantly slowed down our progress toward better seasonal forecast.        

Targeting toward these three challenges, Pan et al. [2022] developed a deep learning model that learns from climate simulation big data for seasonal forecast. Their Conditional Generative Forecasting (CGF) methodology applies a deep generative modeling technique to sample possible climate states a season ahead and adopts a transfer learning technique to account for the data-generating climate model formulation differences. These treatments allow bypassing the aforementioned barriers in dynamical forecast, offering a top-down viewpoint to examine how complicated climate models encode the seasonal predictability information.

Experiments for global seasonal forecast of precipitation and 2 m air temperature show that the CGF methodology achieves competitive performance compared to dynamical forecasts. Using this CGF as benchmark, the authors reveal the impact of insufficient forecast spread sampling that limits the skill of the considered dynamical forecast system. Furthermore, the authors introduced different strategies for composing forecasting ensembles using the CGF methodology, highlighting the potential for leveraging the strengths of multiple climate models to achieve advantageous seasonal forecast.

Citation: Pan, B., Anderson, G. J., Goncalves, A., Lucas, D. D., Bonfils, C. J. W., & Lee, J. (2022). Improving seasonal forecast using probabilistic deep learning. Journal of Advances in Modeling Earth Systems, 14, e2021MS002766. https://doi.org/10.1029/2021MS002766

 —Jiwen Fan, Editor, Journal of Advances in Modeling Earth Systems

Text © 2022. The authors. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.