Geology Reference
In-Depth Information
Chapter 8
Application of Statistical Blockade
in Hydrology
Abstract This chapter introduces a novel Monte Carlo (MC) technique called
Statistical Blockade (SB) which focuses on signi
cantly rare values in the tail
distributions of data space. This conjunctive application of machine learning and
extreme value theory can provide useful solutions to address the extreme values of
hydrological series and thus to enhance modeling of value falls in the
of
hydrological distributions. A hydrological case study is included in this chapter and
the capability of Statistical Blockade is compared with adequately trained Arti
'
Tail End
'
cial
Neural Networks (ANN) and Support Vector Machines (SVM) to get an idea of the
accuracy of the Statistical Blockade.
8.1 Introduction: Statistical Blockade
Statistical analysis could be used to investigate the in
uence of one of the
parameter
is variation to a given simulation. Standard Monte Carlo is an accurate
way to simulate a given scenario. Monte Carlo is a computational method, which is
based on repeated simulations with various random parameters to evaluate the
resulting output. It is usually used in simulation analysis and mathematical systems.
However, the drawback of this approach is the cost of time, and there are a large
numbers of simulations required to capture the rare events. Compared to standard
Monte Carlo methodology, the approach called Statistical Blockade (SB) is able to
sample and analyze the rare events more ef
'
ciently.
SB is an extension of extreme value distribution and its related theorem. The
principle of SB is so simple, in that we
the values unlikely to fall in the low
probability tails (hydrological peaks). Usually, in rare event analysis approaches, to
obtain both samples and statistics for rare events, we may need to generate an
enormous number of samples. We normally use Monte Carlo (MC) simulation for
simulating synthetic values based on the distribution ef
'
block
'
ciency. However, the MC
method is inef
flood values (rare events) as MC
still follows the complete distribution in generating synthetic samples. Thus SB uses a
classi
cient when we only consider the peak
filter out candidate MC points
which will not generate values of interest to us in the tail. In short, an SV classi
er (Support Vector classi
er in our case) to
er
