[Previous Article] [Next Article]
Neural Network to Predict "Short Rains" at the Coast of East Africa
for Boreal Autumn 1997
contributed by Larry Greischar and Stefan Hastenrath
University of Wisconsin, Madison, Wisconsin
The coastal region of East Africa has its "short rains" season in boreal autumn, and its "long rains"
period in boreal spring. The diagnostics of East African rainfall anomalies have been discussed in
Hastenrath et al. (1993). In particular, the autumnal rains were shown to have a strong
relationship with concurrent equatorial zonal winds. The predictand here is a 7-station average of
the normalized departures from the 1931-60 mean for Sep-Oct-Nov-Dec (SOND) and for
Oct-Nov (ON). The stations are Lamu, Malindi, Mombasa and Voi in Kenya; and Dar Es Salaam,
Tanga and Bagamoyo in Tanzania. The predictors are simply the values of the predictand itself for
the 7 years immediately preceding the forecast year, averaged over the respective target seasons
(SOND or ON). That the most recent predictor data is nearly a year prior to the beginning of the
target period provides the opportunity for multiseason lead times for these forecasts. Thus, once
the data for December 1996 became available, this forecast for fall 1997 "short rains" can be
made.
The methodology used to relate the 7-station mean rainfall anomaly over the prior 7 years to that
of the following year is a neural network, as described in Hastenrath et al. (1995) in the context of
South African summer rainfall prediction. Neural nets have been found more successful in
predicting East Africa's "short rains" than multiple regression, especially when the predictor is a
series of values of the predictand itself for previous years. This success was found in forecasts for
South Africa as well, although multiple regression also produced a reasonably skillful forecast
(see the presentations on p. 24 of the December 1994, p. 27 of the December 1995, and p. 31 of
the December 1996 issues of this Bulletin).
The neural network models were developed over the training period of 1928-78, and tested on the
independent verification period of 1979-94 for realistic skill estimation. Resulting skills are
expressed in a correlation framework: For the SOND period 45% of the predictand variance is
explained, and for ON 38% is explained.
The standardized rainfall anomalies over the last 7 years have been as follows:
1990 | 1991 | 1992 | 1993 | 1994 | 1995 | 1996 | |
SOND | 0.54 | -0.01 | 0.65 | 0.08 | 2.04 | 0.34 | -0.52 |
ON | 0.46 | -0.09 | 0.44 | 0.16 | 1.49 | 0.72 | 0.21 |
Because the neural net models have an intermediate (hidden) layer of 3 nodes, the weighting formula applied to the past 7 years is not readily apparent from the system weights. The neural-based forecast for boreal autumn 1996 is shown in the following table:
1997 Forecast (Stand Dev) | Mean/SD (mm), 1931-60 | Best Analog Years | |
SOND | -0.40 | 238/126 | 1964, 73, 76, 79 |
ON | -0.52 | 124/ 95 | 1974, 76, 79, 83 |
The 1996 prediction of 0.04 for SOND turned out to be somewhat too high, while 0.06 for ON
was quite accurate. For boreal autumn 1997, rainfall is expected to be modestly below the
long-term mean.
Hastenrath, S., A. Nicklis and L. Greischar, 1993: Atmospheric-hydrospheric mechanisms of
climate anomalies in the western equatorial Indian Ocean. J. Geophys. Res. (oceans), 98 (C11),
20219-20235.
Hastenrath, S., L. Greischar and J. Van Heerden, 1995: Prediction of the summer rainfall over
South Africa.J. Climate, 8, 1511-1518.