UBC Theses and Dissertations
Enso simulation and prediction using hybrid coupled models with data assimilation Tang, Youmin
The possibility of using a nonlinear empirical atmospheric model for hybrid coupled atmosphere-ocean modelling has been examined by using a neural network (NN) model for predicting the contemporaneous wind stress field from the upper ocean state in the tropical Pacific. Upper ocean heat content (HC) from a 6-layer dynamic ocean model was a better predictor of the wind stress than the (observed or modelled) sea surface temperature (SST). The results showed that the NN model generally had slightly better skills in predicting the contemporaneous wind stress than the linear regression (LR) model in the off-equatorial tropical Pacific and in the eastern equatorial Pacific. Next, the NN and LR atmospheric models were separately coupled to the dynamic ocean model, yielding respectively a hybrid coupled model with a nonlinear atmosphere (NHCM) and one with a linear atmosphere (LHCM). The POP (Principal Oscillation Pattern) analysis shows that the NHCM can produce more realistic ENSO oscillatory behaviour than the LHCM. The phase-locking between the peak of the warm events and the seasonal cycle is more realistically distributed in the NHCM than the very narrow phase-locking in the LHCM. Sensitivity experiments show that in the absence of external forcing, neither NHCM nor LHCM displays the irregular behaviour of ENSO oscillations, suggesting that stochastic forcing (instead of chaos) is likely to cause the irregularity of ENSO. ENSO prediction skills in the two hybrid coupled models have also been investigated under two initialization schemes. The stress-from-ocean initialization scheme, which considers the ocean-feedback in the initial conditions, generally has better predictive skills than the stress-only scheme, where the ocean is simply forced by the observed wind stress. The stress-from-ocean scheme also manifests less decadal variability in the forecast skills than the stress-only scheme. From 1964-1998, with the stress-from-ocean initialization scheme, the forecast correlation skill at 12-month lead time for the NIN03 SST anomalies (SSTA) is about 0.55 for the NHCM and 0.50-0.55 for the LHCM. The main difference in forecast skills between the NHCM and the LHCM occurs in the 1990s, where the NHCM has better skills. A nonlinear canonical correlation analysis of the wind stress and the SSTA shows that the relation between the two was indeed more nonlinear in the 1990s than in the 1980s, which would give the NHCM an advantage over the LHCM in the 1990s. The impact of assimilating different types of data on ENSO simulations and predictions was investigated by separately assimilating the SST, sea surface height anomalies (SSHA), upper ocean heat content anomalies (HCA), and wind stress, with the 3-D Var (variational assimilation) technique. For ENSO prediction, HCA assimilation is the best in improving the correlation skills of the prediction for lead times greater than 5 months. The improvement from SST assimilation is the best for lead times of 4 months or shorter; at longer lead times, SST assimilation degrades the predictions. Wind-stress assimilation is generally the second best for lead times 6 months or longer, while SSHA assimilation generally produces prediction skills only slightly better than without data assimilation. In general, data assimilation yielded less significant improvements to ENSO prediction in the 1990s than in 1980s, which was explained upon examining the HCA in the western and in the eastern equatorial Pacific. In summary, this thesis has initiated the fusion of neural network techniques into dynamical models. Using an NN model for the atmosphere, it has produced the first HCM with a nonlinear empirical atmospheric component, and showed that the nonlinear atmosphere could have advantages over a linear atmosphere in modelling ENSO variability and in ENSO prediction. This study has introduced NN for the assimilation of non-prognostic variables (e.g. HCA) by using NN to relate the non-prognostic variable to prognostic variables, which are then assimilated into the model. While the full 4-D Var HCM is beyond the scope of this thesis, a neural-dynamical hybrid approach under 4-D Var has been developed to study the simple Lorenz system.
Item Citations and Data