Quantification of forecast uncertainty using neural networks

Maximiliano Sacco (1) , Yicun Zhen (2, 3) , Pierre Tandeo (3, 2) , Juan Ruiz (4, 1) , Manuel Pulido (4)

1 CIMA - Centro de Investigaciones del Mar y la Atmósfera
2 IMT Atlantique - SC - Département Signal et Communications
3 Lab-STICC_IMTA_CID_TOMS
4 CONICET - Consejo Nacional de Investigaciones Científicas y Técnicas [Buenos Aires]

Uncertainty quantification in numerical weather and climate prediction is usually achieved using a Monte Carlo estimation (i.e., ensemble forecasting) of the forecast probability distribution function of the state of the system. In this work, we present a method for uncertainty quantification based on neural networks and using a likelihood-based loss function to train the network. This provides state dependent uncertainty estimation, without the need of integrating an ensemble of forecasts. The method is evaluated with a chaotic low-dimensional model in two scenarios: with stochastic errors only (SE) and systematic and stochastic errors (SSE).

Domaines: Océan, Atmosphère; Applications [stat.AP]

Maximiliano Sacco, Yicun Zhen, Pierre Tandeo, Juan Ruiz, Manuel Pulido. Quantification of forecast uncertainty using neural networks. Climate Informatics 2020, Sep 2020, Oxford, United Kingdom. ⟨hal-02942802⟩