Geology Reference
In-Depth Information
Chapter 4
Machine Learning and Artificial
Intelligence-Based Approaches
Abstract The skill of Arti
cial Intelligence (AI)-based computational mechanisms
to model important nonlinear hydrological processes is addressed in this chapter.
Three major themes are illustrated: (1) conventional data-based nonlinear concepts
such as Box and Jenkins Models, ARX, ARIMAX, and intelligent computing tools
such as LLR, ANN, ANFIS, and SVMs; (2) the discrete wavelet transform (DWT),
a powerful signal processing tool and its application in hydrology, and (3) con-
junction models of DWT, namely neuro-wavelet models, Wavelet-ANFIS models,
and Wavelet-SVMs. This chapter gives a detailed description of the training
algorithms used in this topic and points out the conceptual advantages of Leven-
berg
Marquardt (LM) algorithms over Broyden-Fletcher-Goldfarb-Shanno (BFGS)
training algorithms and Conjugate Gradient (CG) training algorithms.
-
Arti
cial Intelligence (AI)-based techniques offered many popular data-driven
models which have been used extensively in the past couple of decades in different
aspects of hydrology, including stream
flow forecasting, evapotranspiration esti-
mation, solar radiation modeling and rainfall-runoff modeling. Rainfall-runoff
dynamics are usually highly nonlinear,
time-dependent, and spatially varying.
Signi
cant advancements in hydrological modeling started with the introduction of
a unit hydrograph model and its related impulse response functions [
83
], and is
considered to be the
first data-driven model in hydrology. In the last four decades,
mathematical modeling of rainfall-runoff series, for reproducing the underlying
stochastic structure of this type of hydrological process, has been widely performed.
Models of AR (autoregressive) and autoregressive moving average (ARMA)
classes [
12
] have played a key role in this kind of approach, producing runoff
prediction for many different time step cases [
6
].
Variant forms of these models such as PAR (periodic AR), PARMA (periodic
ARMA), DARMA (discrete ARMA), etc., were introduced with some considerable
improvements in prediction [
16
,
64
,
86
] introduced a bivariate character to the
conventional way of adopting the ARX (AutoRegressive with eXogenous input)
concept in hydrological time series modeling. ARX and its variant form ARMAX
(ARMA with exogenous input) were considered as much a success for runoff
predictive tools as other models compared in its generation [
89
]. Furthermore, they
