The predictive skill of neural network models for the large-scale
dynamics of the multi-level Lorenz '96 systems
Abstract
The predictive skill of a neural network model for the prediction of the
highly nonlinear Lorenz ’96 dynamics is examined and a way to improve
the skill is investigated. We train neural networks with pairs of a
large-scale variable and its tendency generated by numerical
integrations of full-level Lorenz ’96 equations. The Neural Network (NN)
models are then used to estimate the tendency given state of the
variable which is then updated without resolving or parameterizing
smaller-scale processes. We also apply ensemble data assimilation to the
predicted background states and examine to which degree NN models
capture the dynamics in a long-term prediction-analysis cycle. It has
been found that NN models are skillful in estimating the tendencies for
dynamics with quasi-periodic characteristics. Moreover, they have strong
potentials in predicting even more chaotic waves when an external
forcing has been increased. We have examined if the performance of NN
models can be enhanced by using ensemble frameworks in the context of
machine learning or training an NN with a dataset generated by ensemble
simulations of full-level Lorenz ’96 equations. In these approaches,
prediction-analysis cycles run stably for long periods and NN models are
skillful in representing the large-scale dynamics. However, NN models
can face difficulties in capturing extreme events occurring rarely and
whose predictability is very low. An interesting aspect of the results
is that efforts need to be focused in finding an effective way to
increase the diversity of a whole ensemble and improve the skill in such
situations.