Abstract:
In this research, neural network–based autoregressive moving average with exogenous inputs
(NNARMAX) and autoregressive moving average with exogenous inputs (ARMAX) models
were used to analyse specific humidity (q) data calculated for ten years (1999-2008) from
daily meteorological data (temperature and relative humidity) obtained from the archives of
Nigeria Meteorological Agency NIMET, Oshodi Lagos, Nigeria. The results showed that the
two models captured and predicted specific humidity (q) accurately at 9 hours and 15 hours
respectively over sixteen (16) selected stations in four (4) regions across Nigeria. The
prediction of the training data and validation data of the two models showed that the
NNARMAX model gave better prediction of q than ARMAX model. For example, over
Coastal region in Lagos the NNARMAX model gave lower MSE values for training data of
(0.0007 & 0.0004) for 9 hours and 15 hours respectively and for validation data of (0.0001 &
0.0003) compared to the ARMAX model MSE values for training data (0.2396 & 0.2492)
and for validation data of (0.0242 & 0.0136) for 9 hours and 15 hours respectively while over
Guinea Savannah region in Ibadan the NNARMAX model gave lower MSE values for
training data of (0.0219 & 0.0001) and for validation data of (0.0008 & 0.0005) compared to
the ARMAX model MSE values for training data (0.3478 & 0.4188) and for validation data
of (0.0154 & 0.0198). Also, over Mid-land region in Abuja the NNARMAX model gave
lower MSE values for training data of (0.0021 & 0.0008) and for validation data of (0.0011 &
0.0001) compared to the ARMAX model MSE values for training data (0.3465 & 0.3502) and
for validation data of (0.0183 & 0.0205) while over Sahel region in Kaduna the NNARMAX
model gave lower MSE values for training data of (0.0001 & 0.0003) and for validation data
of (0.0001 & 0.0003) compared to the ARMAX model MSE values for training data (0.2944
& 0.3164) and for validation data of (0.0859 & 0.0907). This clearly showed that the
NNARMAX model yielded better results than the ARMAX model with lower prediction errors.
KEYWORDS: Weather Forecasting, Artificial Neural Networks, ARMAX model, time
series.
