PREDICTION OF THE SUMMER RAINFALL OVER SOUTH-AFRICA

This study develops methods for the extended-range forecasting of summ er rainfall in the eastern part of southern Africa. The predictand is an index (TVR) of the December-January-February precipitation in the T ransvaal. The predictors are based on empirical-diagnostic analyses an d include the July-August-September values of Tahiti minus Darwin pres sure difference as an index (SOI) of the Southern Oscillation; the pre ceding January-February-March value of the 50-mb zonal wind over Singa pore (U50); an index of the October-November surface westerlies along the Indian Ocean equator (UEQ); and an index of November sea surface t emperature in the southwestern Indian Ocean (UKT). These predictors se rve as input to stepwise multiple regression, linear discriminant anal ysis, and neural networking. The training period is 1954-78, and the v erification period 1979-93. Regression models, using as predictors U50 , UEQ, and UKT, account for more than 30% of the variance in the indep endent dataset. The linear discriminant analysis does not perform well . Most powerful is a neural networking model having as input informati on through the end of September, namely U50 and SOI, and explaining 62 % of the variance in the verification period. The predictors used here could, in principle, be compiled in quasi-real time, so that the meth od lends itself to operational application.