Forecast Uncertainty for Data Assimilation using Neural Networks

SACCO MAXIMILIANO; ZHEN YICUN; PIERRE TANDEO; JUAN RUIZ; MANUEL PULIDO

Simposio; International Symposium on Data Assimilation; 2021

A fundamental aspect of data assimilation techniques is the quantification of forecast error uncertainty, as this has a major impact on the quality of the analysis produced and, consequently, on the forecasts generated from it.Most of the current operational assimilation systems obtain a state-dependent uncertainty quantification based on ensemble forecasts. However, this methodology is computationally expensive. In this work, we use a fully connected two-layer hidden neural network for the quantification of state-dependent forecast uncertainty in the context of data assimilation. The input to the network is a set of two consecutive forecast states, the initial condition and the desired time forecast of the analysis. The output of the network is a corrected forecast state and an estimate of its uncertainty.Two methods are proposed. The first method consists of training a neural network using a loss function that estimates uncertainty in a local and semi-supervised manner. While in the second method we use a state space transformation prior to training in order to use a cost function based on the transformed observation probability. For training, we use a large database of forecasts and their corresponding analysis previously calculated. We performed simulation experiments of observation systems using the Lorenz'96 model as a proof of concept for an evaluation of the techniques, and compared it with classical ensemble-based approaches.The results show that both proposed approaches can produce state-dependent estimates of forecast uncertainty without the need for an ensemble of states (at much lower computational cost), especially in the presence of model errors.