چكيده لاتين :
Introduction
The field of prediction using linear statistical methods was considered for a long time.
Nevertheless, in the late 1970s and early 1980s, it was proved that the linear models
are not of appropriate consistency in many of actual applications. The models known
as block-box or data-driven models have been placed as the serious competitors of
classic statistical models in the field of prediction and recovery.
The results, especially in the long-term predictions, are useful in many of water
resource applications such as environmental protection, drought management,
operation of water supply facilities, optimal reservoir operation including multiple
irrigation objectives, power generation, and sustainable development of water
resources. Thus, the prediction of hydrological meteorological time series has been
always a topic of interest in the operational hydrology. This has attracted much
attention in the last few decades and many models have been proposed for
predicting time series in order to improve the hydrological prediction. Most of the
uses of artificial intelligence in the hydrology have been so far on the monthly and
annual data and it has been less used from daily data. In this study, it is dealt with the
abilities of 5 artificial intelligence methods including MLP, SVM, RBF, FIS, and ANFIS,
in the prediction of hydrological time series of precipitation and the maximum and
minimum daily temperatures, through both the station data by making delay in the
main series, and the neighboring station data. It is attempted to identify the best
method to predict and recovery the missing data of the series.
Material and Methods
-Adaptive Neuro-fuzzy Inference System (ANFIS)
In 1993, Jang has introduced a learning method for fuzzy inference system (FIS). In
this method, the neural network learning algorithm is used to create a set of if-then
fuzzy rules with a number of appropriate membership functions (MFs) of inputoutput
pairs. The method, used to create FIS from neural networks framework, is
called ANFIS.
-Support vector regression (SVR)
SV algorithm was developed in Russia in 1960s (Vapnik, 1964; Vapnik and Lerner,
1963). This was a generalization on Generalized Portrait algorithm. The basic idea of
SVM is to use the linear model to implement nonlinear classification boundaries
through a number of non-linear mapping of input vector to a high-dimensional
feature space (Wang et al., 2009).
- Radial basis function (RBF)
RBF is a three-layer neural network, including input layer, hidden layer and output
layer (Guo et al., 2012). High convergence speed, lower reps during training, not
positioning in the local minimum and stronger robustness are the advantages of RBF
over BP (Liu et al., 2006}. RBF first layer neurons release only the inputs features to
the next layer (hidden layer). In the second layer, each neuron associates with a
kernel function with a center and a width. In the last layer, neurons calculate the
weighted simple sum based on the answer of the hidden layer for input pattern (Wen
et al., 2012).
-Multi-layer perceptron (MLP)
Multi-Layer Perceptron (MLP) can be a generalization on perceptron networks with at
least one hidden layer. MLP is a feed forward neural network that has one or more
layers between the input and output layers (Talebizadeh and Moridnejad, 2011).
In MLP, each neuron calculates total weighted inputs based on an activation function
and expresses the response accordingly (there are many activation functions that the
most common of them are sigmoid, hyperbolic, linear, Elliott) (Moghaddamnia et al.,
2009}.
- Fuzzy Inference System (FIS)
The FIS main structure is consisted of three conceptual components:
a) Rules Base: that is consisted of a set of fuzzy rules,
b) A database that define the membership functions (MFs) used in the fuzzy
rules,
c) An argument mechanism that performs the output inference approach based
on derivation rules.
Conclusions
Since the precipitation (rainfall) is of an alternative nature but its time and place
distributions are very inharmonic and the meteorological series are considered as
chaos series (Watts, 2007; Ott, 2002; Lorenz, 1963; Ivancevic and Tijana, 2008L the
use of an input data of the adjacent station or making delay in its series, cannot
provide an accurate prediction, even thought the used method is successful in other
fields. Although the artificial intelligence methods could not be of the necessary and
all-round efficiency in the case of predicting daily precipitation, the use of an
influencing input in the precipitation can be effective for the precipitation prediction.
With this possibility, it can be yet said that it is better to give the prediction of rainfall
in the future days to the general circulation atmospheric models, because these
models have proven their ability in predicting rainfall and other meteorological series
in the next days (Cox, 2010; Holton, 2004; Brown, 2008). In the case of predicting
meteorological series in the future decades, it can be argued that the use of down
scale models and the outputs of general circulation atmospheric models could
operate much better than the artificial intelligence approaches (Khan et al., 2006).
The artificial intelligence methods operate very well in the prediction and recovery of
temperature data and can be used as a tool for recovery the missing data of
meteorological stations. Also, the use of other inputs affecting the temperature can
lead to increase the acceptable performance of these methods. In the case of limit
cases, it could be stated that the artificial intelligence methods are very weak in
providing the limit values, which this weakness was obvious from the error
histogram.
