Page 195 -
P. 195
5.5 Multi-Layer Perceptrons 183
period of January 1, 1999 until August 23, 2000. With the help of Statistics's
intelligent problem solver, an MLP3:3:1 solution was found with features H
(relative humidity) and NS (North-South component of the wind) as external
variables. Forecast one day ahead was achieved with average RMS error of 8%,
and more than half of the cases had absolute deviations between true values and
predicted values of the temperature lower than 2" C.
Figure 5.29. Temperature ("C) forecast (Oporto, 12H00), two days ahead, using a
multi-layer perceptron with two external variables (relative humidity and North-
South wind component).
The following two days ahead prediction of the temperature, T(r), was also
tested:
The multi-layer perceptron implementing this approach achieved, as expected, a
better forecast performance with more than half of the cases having absolute
deviations lower than 1.5". This is the solution corresponding to Figure 5.29.
Notice that the higher deviations occur when data losses occur; after that the MLP
tracks in the incoming data with high performance. Data losses can substantially
decrease the prediction performance in time series forecast.
