% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@INPROCEEDINGS{Kleinert:872741,
author = {Kleinert, Felix and Gong, Bing and Götz, Markus and
Schultz, Martin},
title = {{N}ear {S}urface {O}zone {P}redictions {B}ased on
{M}ultiple {A}rtificial {N}eural {N}etwork {A}rchitectures},
reportid = {FZJ-2020-00219},
year = {2019},
abstract = {Artificial neural networks (ANNs) are well suited to solve
complex and highly non-linear problems. Various ma-chine
learning architectures like convolutional neural networks
(CNNs) and Long Short-Term Memory networks (LSTMs), both
subclasses of ANNs, are applied to the prediction of near
surface ozone concentrations (dma8eu) for a lead time of up
to four days at 51 measurement sites in southern Germany.
Only stations with at least 3500 days of valid data between
1997 and 2015 were used, while the first $80\%$ of the data
were used for training and the remaining $20\%$ for testing
and validation. Forecasts were evaluated with respect to
other continuous predictions from climatological,
persistence and ordinary least square (ols) models.
Furthermore, the quality of threshold exceedance predictions
for varying thresholds was analysed based on the joint
distribution of forecasts and ob-ervations. Finally, it was
examined which input variables are most important to
generate skilful predictions. The results of all three
analyses will be presented for all network architectures.
For example, experiments with a CNN architecture show that
those networks outperform ols and persistence pre-dictions
on all four days. The climatological predictions, however,
are outperformed at the first two days, only. On day three
and four the CNN’s skill score with respect to multivalued
climatological forecasts decreased with time,indicating no
added value on those days as the scores no longer differ
substantially. The CNNs performed best for continuous ozone
predictions between the $20\%$ and the $80\%$ percentile,
but cannot generate skilful predictions for higher or lower
concentrations due to imbalanced training data. The quality
of forecasts was determined primarily by the previous
day’s ozone concentration. Of all other input variables,
which were selected based on their importance for ozone
formation and air mass transport, only temperature and the
wind’s v-component had a small impact on forecast
quality.},
month = {Apr},
date = {2019-04-07},
organization = {EGU General Assembly 2019, Wien
(Austria), 7 Apr 2019 - 12 Apr 2019},
cin = {JSC},
cid = {I:(DE-Juel1)JSC-20090406},
pnm = {512 - Data-Intensive Science and Federated Computing
(POF3-512) / PhD no Grant - Doktorand ohne besondere
Förderung (PHD-NO-GRANT-20170405) / Earth System Data
Exploration (ESDE)},
pid = {G:(DE-HGF)POF3-512 / G:(DE-Juel1)PHD-NO-GRANT-20170405 /
G:(DE-Juel-1)ESDE},
typ = {PUB:(DE-HGF)1},
url = {https://juser.fz-juelich.de/record/872741},
}