Professional Documents
Culture Documents
com
Expert Systems
with Applications
Expert Systems with Applications 35 (2008) 14151421
www.elsevier.com/locate/eswa
a,b,*
a
Faculty of Engineering, Shiraz University, Iran
Department of Water and Environmental Engineering, Sazeh Pardazi Co. Engineering, Iran
c
Department of Civil Engineering, Isfahan University of Technology, Iran
d
Department of Civil Engineering, Faculty of Engineering, Shiraz University, Iran
Abstract
Rainfall forecasting plays many important role in water resources studies such as river training works and design of ood warning
systems. Recent advancement in articial intelligence and in particular techniques aimed at converting input to output for highly nonlinear, non-convex and dimensionalized processes such as rainfall eld, provide an alternative approach for developing rainfall forecasting model. Articial neural networks (ANNs), which perform a nonlinear mapping between inputs and outputs, are such a technique.
Current literatures on articial neural networks show that the selection of network architecture and its ecient training procedure are
major obstacles for their daily usage. In this paper, feed-forward type networks will be developed to simulate the rainfall eld and a socalled back propagation (BP) algorithm coupled with genetic algorithm (GA) will be used to train and optimize the networks. The technique will be implemented to forecast rainfall for a number of times using rainfall hyetograph of recording rain gauges in the Upper
Parramatta catchment in the western suburbs of Sydney, Australia. Results of the study showed the structuring of ANN network with
the input parameter selection, when coupled with GA, performed better compared to similar work of using ANN alone.
2007 Elsevier Ltd. All rights reserved.
Keywords: Rainfall forecasting; Articial neural networks; Genetic algorithms; Input determination
1. Introduction
Rainfall forecasting is one of the most dicult and
important processes of the hydrologic cycle. This is largely
related to the variability it displays over a wide range of
scales both in time and space. Flash ooding, being a product of intense rainfall, is a life-threatening phenomenon.
Developing a rainfall forecasting and ood warning system
for typical catchments is not considered a simple task. Both
internal and external characteristics of rainfall eld depend
on many factors including: pressure, temperature, wind
speed and its direction, meteorological characteristics of
the catchments and so on. Although a physically-based
*
Corresponding author. Address: Faculty of Engineering, Shiraz
University, Iran.
E-mail addresses: mm_nasseri@yahoo.com (M. Nasseri), kasghari@
cc.iut.ac.ir (K. Asghari), abedini@shirazu.ac.ir (M.J. Abedini).
0957-4174/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2007.08.033
1416
1417
4. Application
Recording precipitation gauges from the Upper Parramatta River basin were chosen for this study. The catchment
is located in the western suburbs of Sydney, Australia
(Fig. 2) with a catchment size of about 112 km2. Within
the catchment, the land use is typical of urban environment
with a mix of commercial, agricultural and parkland areas.
In the study area, there are fourteen recording rain gauges.
The data set had a length of four years consisting of eventwise recording precipitation from 1996 to 2000. The temporal resolution of rainfall measurement for all fourteen rain
3. Sensitivity analysis
Sensitivity analysis should be considered an essential
pace to all mathematical-based modeling. The main advantage of performing sensitivity analysis is to identify sensitive parameters or processes associated with model
output (Skaggs & Barry, 1996). In ANN modeling, like
any mathematical-based model, sensitivity analysis provides feedback as to which input parameters are the most
signicant. It has to be emphasized that computation of
sensitivity coecients in a typical inputoutput model is
not a trivial task as we are not faced with a closed form
function, but a complex procedure to convert input to output. For the networks which are developed in this research,
sensitivity coecients are computed by using mean value of
each parameter as a base value, having a bound dened by
standard deviation.
1418
Table 1
Parameter and comparative results of various models
Table 4
Results of sensitivity analysis for models 6 and 7
Models
Pc%
Pm%
MSE
NMSE
R2
1
2
3
4
5
6
7
5
10
10
15
15
15
15
92
96
94
96
94
96
96
1.2
1.4
1.2
1.2
1.2
1.2
1.2
0.007
0.020
0.030
0.009
0.010
0.121
0.018
0.161
0.351
0.400
0.280
0.310
0.002
0.360
0.84
0.66
0.74
0.71
0.7
0.99
0.66
Models
Previous lags
6
7
1.24
0.32
0.23
0.18
0.08
0.1
0
0
gauges is 5-min interval. After doing some sort of preliminary exploratory data analysis, a total of 26 storm events
were selected keeping zeros among data sets for synchronization purposes. Eighteen of the events were used for training, four events for validation and the other four for
testing. Due to chaotic behavior of rainfall eld, some sort
of transformation has been implemented to reduce the variance of variation, in order to achieve better performance
and faster convergence in training procedure. The transformation, y = 0.5Log10(x + 1), was used for the input data
(Luk et al., 2000, 2001); where x is the original rainfall data
and y is the data after transformation.
0.7
0.6
Computed
0.5
0.4
0.3
0.2
0.1
0
0
0.2
0.4
0.6
0.8
Observed
Table 2
Results of sensitivity analysis for network modeling by discrete data
Models
7253
7255
7257
7259
7261
7263
7265
7267
7269
7273
7283
7285
7299
First lag
1
2
3
4
5
0.01
0.0
0.0
0.00
0.00
0.0
0.04
0.0
0.05
0.15
0.19
0.1
0.0
0.01
0.0
0.0
0.07
0.08
0.06
0.12
0.1
0.12
0.15
0.06
0.2
0.19
0.15
0.1
0.01
0.05
0.09
0.0
0.05
0.0
0.11
0.09
0.11
0.01
0.0
0.0
0.01
0.05
0.06
0.18
0.04
0.0
0.03
0.0
0.0
0.03
0.0
0.06
0.09
0.0
0.05
0.09
0.16
0.1
0.09
0.02
0.0
0.0
0.13
0.14
Second lag
1
2
3
4
5
0.02
0.0
0.03
0.06
0.0
0.02
0.0
0.0
0.01
0.0
0.0
0.0
0.05
0.02
0.0
0.0
0.06
0.0
0.0
0.00
0.0
0.0
0.06
0.0
0.02
0.0
0.03
0.00
0.02
0.02
0.02
0.06
0.05
0.0
0.03
0.0
0.0
0.0
0.02
0.1
0.04
0.0
0.03
0.0
0.0
0.0
0.02
0.02
0.0
0.0
0.0
0.0
0.05
0.0
0.00
0.0
0.0
0.0
0.02
0.0
0.02
0.0
7285
7299
Table 3
Results of mean sensitive coecients (SC) for model 4 rank the most eective station
Stations Nos.
7251
7253
7255
Mean of SC
0.002
0.05
7257
7259
7261
7263
7265
0.08
0.14
0.13
0.1
13
7267
0.003
12
7269
7273
7283
0.02
0.02
0.09
10
0.14
11
0.25
Computed
0.2
0.15
0.1
0.05
0
0
0.1
0.2
0.3
0.4
0.5
Observed
derived from optimized ANNs by GA to achieve an ecient input conguration for minimum error. Basically,
by detecting eective inputs, the best input combination(s)
for intelligence prediction will appear. Delineation of optimum lag time(s) at a particular location, investigation of
the extend to which inclusion of spatial information might
improve network performance, and assessment of performance indicator with regard to application of cumulative
versus discrete data were addressed in this paper as well.
Seven cases of simulation models have been selected to
illustrate the performance of the proposed technique. These
runs dier from one another mainly on the genetic algorithm parameters (i.e. Pm and Pc), surrounding rain gauges
used, number of time lags associated with each rain gauge,
temporal resolution of rainfall events, and data type
(cumulative versus discrete). The probability of mutation
(Pm) varied, traditionally, between 1.2% and 1.4%, while
1419
Elevation (m)
0.6
0.5
0.4
0.3
Observed
0.2
Computed
0.1
0
1
51
101
151
201
251
301
351
401
451
501
Fig. 4a. Estimated and observed events in time series form for experimental runs #6.
551
1420
Elevation (m)
0.3
Computed
0.25
0.2
0.15
0.1
0.05
0
1
101
201
301
401
501
Fig. 4b. Estimated and observed events in time series form for experimental runs #7.
unique input vector that best represents the underlying processes. This was proved with an application of PACF analysis, rst applied by Sudheer, Gosain, and Ramasastri
(2002), with the data series used in this problem in delineating a unique input vector that best represents the underlying processes. Fig. 5a and 5b shows the autocorrelation
plot for the recording rain gauge of interest (i.e. 7261 station) along with 95% and 99% condence limits considering data in cumulative and discrete mode. Careful
analysis of Tables 2 and 4 clearly conrm the fact that
the rst lag has a major contribution, particularly in the
model 6, and other lags fall outside the condence interval.
This implies that this tool can be used eectively to delineate optimum time lag and consequently network structure
especially in model with cumulative data type.
1.2
PACF Coefficient
Cofidence Limit 99%
Correlation Coefficient
0.8
0.6
0.4
0.2
0
0
Time Lag
0.7
Confidenc Limit 99%
Correlation Coefficient
0.6
0.5
0.4
0.3
0.2
0.1
0
1
Time Lag
6. Conclusion
The focus of this paper was to apply optimized ANN
using GA for short term rainfall forecasting by determination of suitable input parameters and designing the best
network architecture. The study reported in this article
has led to the conclusion that MLP type network coupled
with GA, consistently performed better compared to
MLP type network alone. Compare to previous study
(Luk et al., 2000), number of eective rain stations as input
parameters have been decreased in order to forecast a rain
target and reduced the model noise. Furthermore, the integration of the GA method with ANN indicates an
improvement in reducing the order of errors compare to
the ANN model alone. Considering the use of dierent
time lags of surrounding gauge stations, through the use
of sensitivity analysis along with the special correlation
function, an eective number of time lags were identied
which are also useful tools in reducing the input parameters. Also, due to chaotic nature of rainfall and short time
memory, it was observed that number of eective input stations decreases as time steps increases. Finally, caused by
nature of integration, cumulative data could lead to highly
better statistical performance in rainfall forecasting compared to discrete data type.
References
Bowden, G. J., Dandy, G. C., & Maier, H. R. (2005a). Input determination for neural network models in water resources applications,
Part 1-background and methodology. Journal of Hydrology, 301,
7592.
Bowden, G. J., Dandy, G. C., & Maier, H. R. (2005b). Input determination for neural network models in water resources applications, Part
2. Case study: Forecasting salinity in a river. Journal of Hydrology,
301, 93107.
1421