 |
| Titel |
An experiment on the evolution of an ensemble of neural networks for streamflow forecasting |
| VerfasserIn |
M.-A. Boucher, J.-P. Laliberté, F. Anctil |
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| ISSN |
1027-5606
|
| Digitales Dokument |
URL |
| Erschienen |
In: Hydrology and Earth System Sciences ; 14, no. 3 ; Nr. 14, no. 3 (2010-03-30), S.603-612 |
| Datensatznummer |
250012232
|
| Publikation (Nr.) |
 copernicus.org/hess-14-603-2010.pdf |
|
|
|
|
|
| Zusammenfassung |
| We present an experiment on fifty multilayer perceptrons trained for
streamflow forecasting on three watersheds using bootstrapped input series. This type of neural network
is common in hydrology and using multiple training repetitions
(ensembling) is a popular practice: the information issued by the
ensemble is then aggregated and considered to be the final
output. Some authors proposed that the ensemble could serve the
calculation of confidence intervals around the ensemble mean. In the
following, we are interested in the reliability of confidence
intervals obtained in such fashion and in tracking the evolution of
the ensemble of neural networks during the training process. For each
iteration of this process, the mean of the ensemble is computed along
with various confidence intervals. The performance of the ensemble
mean is evaluated based on the mean absolute error. Since the ensemble
of neural networks resemble an ensemble streamflow forecast, we also
use ensemble-specific quality assessment tools such as the Continuous
Ranked Probability Score to quantify the forecasting performance of
the ensemble formed by the neural networks repetitions. We show that
while the performance of the single predictor formed by the ensemble
mean improves throughout the training process, the reliability of the
associated confidence intervals starts to decrease shortly after the
initiation of this process. While there is no moment during the
training where the reliability of the confidence intervals is
perfect, we show that it is best after approximately 5 to 10
iterations, depending on the basin. We also show that the Continuous
Ranked Probability Score and the logarithmic score do not evolve in
the same fashion during the training, due to a particularity of the
logarithmic score. |
| |
|
| Teil von |
|
|
|
|
|
|
|
|