Environmental Engineering Reference
In-Depth Information
8
cial Intelligence Techniques
for Real-Time Flood Forecasting
Arti
JONATHAN LAWRY, DANIEL R. M C CULLOCH,
NICHOLAS J. RANDON AND IAN D. CLUCKIE
Abstract
Traditional approaches to flood forecasting in-
volve multi-dimensional mathematical models
based extensively on underlying physical princi-
ples. In contrast, machine learning algorithms are
data-driven methods whereby models are inferred
directly from a database of training examples.
Consequently the incorporation of background
knowledge, in the form of an understanding of the
hydrology of the system being studied, only takes
place indirectly through, for example, the choice
of input variables to the AI algorithm, or through
the identification of an appropriate lead time for
prediction. For this reason data-driven models are
sometimes referred to as being 'black box'.
Typical examples of machine learning
algorithms that have been applied to flood predic-
tion are Neural Networks (NN) and Support
Vector Machines (SVM). Both methods are based
on the generation of separating hyperplanes, for
the former in attribute space and for the latter in
non-linear transformation of attribute space. Ex-
amples of Neural Network applications include
that of Campolo et al. (1999), which applied NNs
to the prediction of water levels for the River
Tagliamento in Italy, and that of Han
et al. (2007a), which compared NN to transfer
function models. Recent research into forecasting
based on SVMs includes Han et al. (2007b) and
Randon et al. (2004) on the Bird Creek catchment
in USA, and Yu et al. (2006) on the Yan-Yang river
basin in Taiwan.
Algorithms such as NNs and SVMs are also
black-box in a different way, in that the models
they generate are very difficult to interpret.
Fuzzy rule-based artificial intelligence techniques
for flood forecasting are introduced. These incor-
porate both fuzziness and probabilistic uncertain-
ty, which are inherent in many hydrological
systems. Two algorithms, Linguistic Decision
Trees and Fuzzy Bayes, are applied in a number
of case studies relating to weather radar image
classification and flow and level time series fore-
casting. The models produced are shown to be
transparent and understandable as well as provid-
ing accurate forecasts.
Introduction
Artificial intelligence (AI) algorithms such as
those developed in the field of machine learning
are a source of powerful new techniques for flood
prediction. By generating models automatically
from data they can provide computationally effi-
cient and accurate prediction models for real-time
forecasting that avoid the costs of solving complex
non-linear equations. As such they can form the
basis of new prediction tools, which can be run
locally on relatively low-specification computers,
complementing more traditional methods.
Search WWH ::




Custom Search