| Summary: | Neural network models were used to seasonally forecast the tropical Pacific sea surface
temperature anomalies (SSTA) in the Nino 3.4 region (6°S - 6°N, 120°W - 170°W).
The inputs to the neural networks (i.e. the predictors) were the first seven wind stress
empirical orthogonal function (EOF) modes of the tropical Pacific (20°S - 20°N, 120°E -
70°W) for four seasons and the Nino 3.4 SSTA persistence for the final season. The period
1952-1981 was used for training the neural network models, and the period 1982-1992
for forecast validation. At 6-month lead time, neural networks attained forecast skills
comparable to the other El Nino - Southern Oscillation (ENSO) models. The results
suggested that neural network models were viable for ENSO forecasting even at longer
lead times of 9 to 12 months. It appeared that at these longer leads, the underlying
relationship between the wind stress and Nino 3.4 SSTA became increasingly nonlinear.
Two types of neural network models were further compared for forecasting the SSTA
over several standard equatorial Pacific regions (Nino 3, 3.4, 3.5, 4, P2, P4 and P5).
The first type used the sea level pressure (SLP) as predictor, while the second one used
the wind stress. By ensemble averaging over 20 forecasts with random initial weights,
the resulting forecasts were much less noisy than those in the earlier models. The wind
stress models had better forecast skill at short lead times, while the SLP models generally
had better skill at lead times of 6 months or longer. The western-central regions of the
equatorial Pacific Ocean were best forecasted, with the Nino 4 region (6°S - 6°N, 160°E
- 150°W) having the highest skiU, foUowed by Nino 3.4 and 3.5 (10°S - 6°N, 120°W -
180°), then Nino 3 (6°S - 6°N, 90° W - 150°W). The eastern boundary regions P4 (0 -
10°N, 80°W - 100°W) and P5 (10°S - 0, 80°W - 100° W) had much lower skill, while the western boundary region P2 (10°S - 10°N, 140°E - 180°) had no forecast skill.
In an attempt to understand the inner working of the models, smaller networks were
constructed with the extended empirical orthogonal functions (EEOF) of the SLP field
as inputs. These smaller networks delivered forecasting skills similar to those of earlier
models. By network pruning and spectral analyses, four important inputs were identified:
modes 1, 2 and 6 of the SLP EEOFs and the SSTA persistence. Mode 1 characterized
the low frequency oscillation (LFO, with 4 - 5 years period), and was interpreted as the
typical ENSO signal, while mode 2, with a period of 2 - 5 years, appeared to characterize
the quasi-biennial oscillation (QBO) plus the LFO. Mode 6 was dominated by decadal
and interdecadal variations. Thus, forecasting ENSO seems to require information from
the QBO, and the decadal-interdecadal oscillations. The nonlinearity of the networks
tended to increase with lead time, and to become stronger for the eastern regions of the
equatorial Pacific Ocean. === Science, Faculty of === Earth, Ocean and Atmospheric Sciences, Department of === Graduate
|
|---|
