Modeling flood occurences using soft computing technique in southern strip of Caspian Sea Watershed
Modeling of hydrological process has been increasingly complicated since we need to take into consideration an increasing number of descriptive variables. In recent years soft computing methods like fuzzy logic and genetic algorithm are being used in modeling complex processes of hydrologic ev...
Saved in:
Main Author: | |
---|---|
Format: | Thesis |
Language: | English |
Published: |
2012
|
Subjects: | |
Online Access: | http://psasir.upm.edu.my/id/eprint/67266/1/FPAS%202012%2025%20IR.pdf |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | Modeling of hydrological process has been increasingly complicated since we need to
take into consideration an increasing number of descriptive variables. In recent years
soft computing methods like fuzzy logic and genetic algorithm are being used in
modeling complex processes of hydrologic events. The complex non linear behavior of
flood and short record of observed data in the region makes the study of flood
problematic. This thesis aims to apply soft computing techniques including fuzzy logic,
neural network and genetic algorithm on different aspects of flood modeling including
hydrological homogeneity and flood prediction in southern Caspian Sea Watersheds.
This area with 42400 square kilometers has been affected by severe floods causing
damages to human life and properties. A total of 61 hydrometric stations and 31 weather
stations with 44 years observed data (1961-2005) are available in the study area.
Delineation of homogeneous regions in terms of flood behavior was the initial step
of this thesis which was achieved by several methods. The conventional methods of
homogeneity i.e. hard clustering (hierarchical and non-hierarchical clustering, K-means) and soft clustering (Fuzzy C-means and Kohonen) were studied and
compared by L-moment techniques. Factor analysis using principle component
analysis (PCA) with an orthogonal rotation method, varimax factor rotation have
resulted in 4 out of 15 parameters namely area, mean elevation, Gravelius factor and
shape factor. In conventional hard clustering approach, the number of clusters was
determined by hierarchical clustering and two-step cluster analysis; then the sites were
allocated to the appropriate cluster by k-means clustering method. In soft clustering
approach, Kohonen network was employed to find the number of clusters and then the
allocation of sites to the appropriate cluster was performed by using fuzzy c-means
method. As a conclusion of regionalization, 38, 13 and 10 catchments were allocated to
3 specified regions. Assessment of homogeneity in this region was achieved and
approved by three proposed heterogeneity measures i.e. HLcv, HLck, HLcs with 1.94,
1.13 and 0.71, respectively. In order to fully investigate the homogeneity (h) of
catchments and overcome incompatibility that may happen on boundaries of cluster
groups, a new method was used which utilizes physical and climatic parameters as well
as flood seasonality and geographical location. This approach was based on fuzzy expert
system (FES) using Fuzzy Toolbox of MATLAB software. Genetic algorithm (GA) was
employed to adjust parameters of FES and optimize the system. Results obtained by
this method were compared to the conventional methods. Since 3 L-moment criteria
obtained by this method were significantly lower than previous methods it can be
concluded that FES has better performance. After defining homogeneous region in the
study area, flood models were developed. A total of 24 sites which were eligible in terms
of adequate rainfall and runoff observed data were selected in this region. This area contains 604 pairs of observed data which was grouped into 60%, 20% and 20% for
training, validation and testing, respectively. The most popular network in hydrology i.e.
Multilayer Feedforward Back Propagation (MLFFBP) was used. Among the available
learning algorithms in the Neural Network Toolbox of MATLAB, three algorithms,
gradient descent back propagation (TRAINGD), gradient descent with adaptive learning
rule back propagation (TRAINGDA) and the Levenberg-Marquardt (TRAINLM) were
studied. Three algorithms of Linear (PURELIN), hyperbolic tangent sigmoid (TANSIG)
and logistic sigmoid (LOGSIG) activation functions were selected for output layer. The
hidden layer includes 1 layer with different neurons. Based on the mentioned criteria
several scenarios were defined and compared which resulted to a structure of 8-10-1
with the Levenberg-Marquardt (LM) as the training algorithm and logistic sigmoid
function in the output layer. This study found that FES technique is a powerful solution
to make pooling groups as a promising approach that satisfies the high homogeneity of
the catchments. The application of FES optimized by GA on regionalization creates opportunities for further researches which utilizes different types of optimization like Ant Colony Optimization (ACO), ANN’s, Particle Swarm Optimization (PSO) and Imperialist Competitive Algorithm (ICA). |
---|