Description of neural network model (version 3.5)

A neural network (NN) model has been used by our group to predict the SST anomalies in the Nino3.4 region in the equatorial Pacific (Tang et al. 2000). By version 3.1, the NN model has been extended to forecast the SST anomalies over the whole tropical Pacific. In version 3.2, the subsurface temperature anomalies in the tropical Pacific ocean were included as predictors. In version 3.3, we have removed the subsurface temperature anomalies as predictors, since their record is relatively short (available only since 1980), and they are not always available on time for our monthly forecasts. We have, however, introduced a Bayesian NN in version 3.3 (see Wu et al. 2006 for details).

The data used in version 3.3 came from two datasets: (a) the monthly sea level pressure (SLP) on 2.5° × 2.5° grids from the NCEP/NCAR reanalysis (Kalnay et al. 1996; downloadable from ftp.cdc.noaa.gov/Datasets/ncep.reanalysis.derived/surface); (b) the monthly extended reconstructed sea surface temperature on 2° × 2° grids (ERSST version 2; Smith and Reynolds 2004).

By Jan. 2010, ERSST version 2 was no longer available. We therefore had to retrain our Bayesian NN model (version 3.4) using ERSST version 3 (downloadable from http://www.ncdc.noaa.gov/oa/climate/research/sst/ersstv3.php; Smith et al. 2008).

Relative to version 3.3 (Wu et al. 2006), the most current version (version 3.5) has several changes: (1) The data used to train and validate models has been extended (Jan. 1948- Dec. 2009). (2) A more rigorous cross-validation method has been implemented to verify forecast skills -- namely, data from five consecutive calendar years were held for validation/verification, while data from the remaining years were used as training data. Only the 3rd year in the 5-year window was forecasted by the trained model. The 5-year window was moved forward by one year each time until independent forecast verification was done for all years. (3) Among the training data, only 2/3 of the data were randomly selected (in one-year blocks) and used to train the NN models. The remaining 1/3 data were used as testing data for model selection. NNs with 1, 2 and 3 hidden neurons were trained separately and validated on the testing data, and the one model with highest correlation skill was selected. This procedure was repeated 100 times and the ensemble average from these 100 models was used to give the final forecast. Cross-validation forecast correlation skills by Bayesian NN model (with corresponding value from multiple linear regression given in parenthesis) for the Nino4, Nino3.4, Nino3 and Nino1+2 regions and for lead times of 3, 6, 9, and 12 months (with definition of lead time same as in Wu et al. 2006) are given in the table below:
Nino4 Nino3.4 Nino3 Nino1+2
3 month 0.9109 (0.9100) 0.9051 (0.9048) 0.8734 (0.8754) 0.7997 (0.8022)
6 month 0.7657 (0.7615) 0.7387 (0.7366) 0.6810 (0.6799) 0.5524 (0.5457)
9 month 0.6320 (0.6283) 0.5960 (0.5921) 0.5364 (0.5322) 0.3500 (0.3377)
12 month 0.4899 (0.4882) 0.4737 (0.4703) 0.4413 (0.4367) 0.2583 (0.2486)


References:

Kalnay, E., et al., 1996. The NCEP/NCAR 40 year reanalysis project. Bull. Amer. Meteorol. Soc., 77, 437-471.

Smith, T.M., and R.W. Reynolds, 2004. Improved extended reconstruction of SST [1854-1997]. Journal of Climate, 17, 2466-2477.

Smith, T.M., R.W. Reynolds, Thomas C. Peterson, and Jay Lawrimore, 2008. Improvements to NOAA's historical merged land-ocean surface temperature analysis (1880-2006). Journal of Climate, 21, 2283-2296.

Tang, B., W.W. Hsieh, A.H. Monahan and F.T. Tangang, 2000. Skill comparisons between neural networks and canonical correlation analysis in predicting the equatorial Pacific sea surface temperatures. J.Climate, 13: 287-293.

Wu, A., W.W. Hsieh and B. Tang, 2006. Neural network forecasts of the tropical Pacific sea surface temperatures. Neural Networks. 19: 145-154. doi:10.1016/j.neunet.2006.01.004. (preprint in PDF).