A Generalized Canonical Mixed Regression Model for ENSO Prediction with Its Experiment

A scheme is proposed for predicting NINO-region SST in terms of a generalized canonical mixed regression model based on principal component canonical correlation analysis (PC-CCA), and into the scheme are introduced such techniques as EEOF, PRESS criterion and consensus prediction. By optimizing physical factors and selecting optimal model parameters, experiments were made successful in predicting the NINO SST index for 1 to 4 seasons to follow. The scheme is shown to be stable in operation and its total technical level compares well with that of the model published in NOAA/NWS/NCEP CPC Climate Diagnostics Bulletin but the number of factors needed in our scheme is much fewer than that for the CPC°¯s model in dealing with the same problems. This makes it possible to establish an operational ENSO monitoring system in China. The main results are as follows:

The independent sample predictions, 1981-98 back forecasting, and the 1997-98 ENSO prediction all demonstrate high applicability of our model and scheme that are quite steady during operation. Take NINOs 3.4 and 4 predictions for example. Their 1-4 season leading predictions gave the correlations of greater than 0.85, 0.70, 0.50 and 0.35, respectively.

Our model compares advantageously to the CCA statistical model developed in the US CPC (Climate Prediction Center, Barnston et al., 1992) as regards predictive skill, and part of our predictions has exceeded those of the CPC model. In reference to 2-season leading predictions, for instance, the CCA model gives maximum scoring ranging over 0.85-0.89 just for wintertime months compared to the maximum of 0.85-0.87 for all seasons from our model, with the minimum of 0.30-0.35 (from the CCA) versus 0.37 (from our model). Particularly, it should be pointed out that the CCA model requires a large volume of gridded data, consisting of global sea level pressure (SLP), tropical Pacific SST and sea level height (SLH), and the 20 C isothermal depth in contrast to 20 predictors for our model, leading to the much higher efficiency compared to the CCA model.

The paper presents a canonical mixed regression model which is actually the generalized simple type with scattered coefficients. The predictive scheme includes EEOF technique for the information on previous SST evolution (suggestive of the CCA establishment of a certain similarity relation between the previous SST evolution and the SST index at the prognostic time interval), PRESS sorting out model parameters and factors for independent sample experiments with 1-4 season leading predictions (suggesting a consensus prediction technique). All these have led to successful prediction of, say, the 1997/98 ENSO episode for longer than 6 months in advance and quite good prognosis of the termination of 1998 warm phase by the end of June based on measured data prior to January 1998. Preliminary prediction has been made of a strong La Nina event to occur in October 1998 whose precursors were beginning to emerge in the context of data before May 1998.