|
Titel |
Identifying Patterns in the Weather of Europe for Source Term Estimation |
VerfasserIn |
Iraklis Klampanos, Charalambos Pappas, Spyros Andronopoulos, Athanasios Davvetas, Andreas Ikonomopoulos, Vangelis Karkaletsis |
Konferenz |
EGU General Assembly 2017
|
Medientyp |
Artikel
|
Sprache |
en
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 19 (2017) |
Datensatznummer |
250147158
|
Publikation (Nr.) |
EGU/EGU2017-11273.pdf |
|
|
|
Zusammenfassung |
During emergencies that involve the release of hazardous substances into the atmosphere the
potential health effects on the human population and the environment are of primary concern.
Such events have occurred in the past, most notably involving radioactive and toxic
substances. Examples of radioactive release events include the Chernobyl accident in 1986, as
well as the more recent Fukushima Daiichi accident in 2011. Often, the release of dangerous
substances in the atmosphere is detected at locations different from the release origin. The
objective of this work is the rapid estimation of such unknown sources shortly after the
detection of dangerous substances in the atmosphere, with an initial focus on nuclear or
radiological releases.
Typically, after the detection of a radioactive substance in the atmosphere indicating the
occurrence of an unknown release, the source location is estimated via inverse modelling.
However, depending on factors such as the spatial resolution desired, traditional inverse
modelling can be computationally time-consuming. This is especially true for cases where
complex topography and weather conditions are involved and can therefore be problematic
when timing is critical. Making use of machine learning techniques and the Big Data Europe
platform1,
our approach moves the bulk of the computation before any such event taking place,
therefore allowing for rapid initial, albeit rougher, estimations regarding the source
location.
Our proposed approach is based on the automatic identification of weather patterns within
the European continent. Identifying weather patterns has long been an active research field.
Our case is differentiated by the fact that it focuses on plume dispersion patterns and these
meteorological variables that affect dispersion the most. For a small set of recurrent weather
patterns, we simulate hypothetical radioactive releases from a pre-known set of nuclear
reactor locations and for different substance and temporal parameters, using the Java flavour
of the Euratom-supported funded RODOS (Real-time On-line DecisiOn Support)
system2
for off-site emergency management after nuclear accidents. Once dispersions have been
pre-computed, and immediately after a detected release, the currently observed
weather can be matched to the derived weather classes. Since each weather class
corresponds to a different plume dispersion pattern, the closest classes to an unseen
weather sample, say the current weather, are the most likely to lead us to the release
origin.
In addressing the above problem, we make use of multiple years of weather reanalysis data from NCAR’s
version3 of ECMWF’s
ERA-Interim4.
To derive useful weather classes, we evaluate several algorithms, ranging from
straightforward unsupervised clustering to more complex methods, including relevant
neural-network algorithms, on multiple variables. Variables and feature sets, clustering
algorithms and evaluation approaches are all dealt with and presented experimentally. The
Big Data Europe platform allows for the implementation and execution of the above tasks in
the cloud, in a scalable, robust and efficient way. |
|
|
|
|
|