Document Sample
20320130406022 Powered By Docstoc
International Journal of Advanced Research in Engineering RESEARCH IN ENGINEERING
                              AND Volume 4, Issue 7, November – December (2013), © IAEME
6480(Print), ISSN 0976 – 6499(Online)TECHNOLOGY (IJARET)

ISSN 0976 - 6480 (Print)
ISSN 0976 - 6499 (Online)
Volume 4, Issue 7, November - December 2013, pp. 183-191
Journal Impact Factor (2013): 5.8376 (Calculated by GISI)                    ©IAEME


                            Nizar Ali Charaniya1,      Dr. Sanjay V. Dudul2
                    (Electronics Engg Department, B.N.College of Engg, Pusad, India)
         (Applied Electronics Department, Sant Gadge Baba Amravati University, Amravati, India)


        Forecasting Indian summer monsoon rainfall (ISMR) has been a challenging task for the
research community. The variability of rainfall in time and space, however makes it extremely
difficult to have quantitative forecasting of rainfall. The depth of rainfall and its distribution in the
temporal and spatial dimensions depends on many ecological parameters. Due to the complexity of
the atmospheric processes by which rainfall is generated the accuracy of prediction is very low.
There are two possible methods for rainfall prediction. The first one is based on the study of the
rainfall processes and its dependence on other meteorological parameters such as pressure, humidity,
vapor pressure, temperature etc. However, this method is complex and non- feasible because rainfall
is the result of a number of complex atmospheric parameters which vary both in space and time and
these parameters are limited in both the spatial and temporal dimensions. Another method to forecast
rainfall is based on the pattern recognition methodology. In these method relevant spatial and
temporal features of rainfall series in past are extracted. These features are then utilized to predict the
rainfall in future. In this paper different models have been designed for the prediction of ISMR for
the next year based on the rainfall pattern for past four years. An adaptive neuro fuzzy inference
system (ANFIS), linear and nonlinear regressive model have been designed for prediction of rainfall.

Keywords: Indian Summer Monsoon Rainfall, Neuro-Fuzzy, Time Series Analysis, Pattern


        In India rainfall information is vital for planning crop production, water management and all
activity plans in the nature. The incident of extended dry period or heavy rain at the critical stages of
the crop growth and development may lead to noteworthy reduction in the crop yield and hence this
may affect the economy of the country. India is an agricultural country and its economy is largely
based upon agricultural product. Thus, rainfall prediction becomes a significant factor in agricultural

International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976 –
6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 7, November – December (2013), © IAEME

countries like India [1]. A wide range of rainfall forecast methods are employed in weather
forecasting at regional and national levels. Rainfall is a random process and therefore prediction of
rainfall is very difficult and cumbersome. Accurate and timely prediction of rainfall is a foremost
challenge for the research community [2]. Rainfall prediction modeling involves a combination of
computer models, inspection and information of trend and pattern. Using these methods, reasonably
accurate forecasts can be made up. Several recent research studies have reported rainfall prediction
using different weather and climate forecasting methods [3], [4], [5].
        Because of strong non-linear, high degree of uncertainty, and time-varying characteristics of
the rainfall, it is very difficult to have a single superior model [6] for accurate prediction of rainfall.
Artificial neural networks (ANNs) have been accepted as a potentially useful tool for modeling
complex non-linear systems and widely used for prediction [7]. In the forecasting context, ANNs
have also proven to be an efficient alternative to traditional methods for rainfall forecasting [8], [9].
        Alternate method for rainfall time series prediction is using pattern recognition. In these
method relevant spatial and temporal features of rainfall series in past are extracted. These features
are then utilized to estimate the rainfall pattern. If rainfall pattern can be estimated accurately then it
can help in predicting the rainfall for the next year [10] [11]. The adaptive neuro fuzzy inference
system (ANFIS) is a network which uses neural network learning algorithms and fuzzy reasoning in
order to map an input space to an output space. It has the ability to combine the verbal power of a
fuzzy system with the mathematical power of a neural system adaptive network. ANFIS has
tremendous capability of learning [12]. Therefore in this paper an attempt has been made to estimate
rainfall pattern using ANFIS model. In addition an attempt has also been made to develop linear and
nonlinear regressive model for prediction of rainfall using different nonlinear estimating functions.


2.1     Architecture and Algorithm
        ANFIS has the benefit of allowing the drawing out of fuzzy rules from numerical data or
expert knowledge and adaptively to constructs a rule base. Furthermore, it can tune the complicated
conversion of human intelligence to fuzzy systems.
        The main drawback of the ANFIS predicting model is the time required for training structure
and determining parameters, is very long. For simplicity, we have assumed a fuzzy inference system
with two inputs, x and y, and one output, z. For a first-order Sugeno fuzzy model [13], a typical rule
set with two fuzzy if–then rules can be expressed as

                           Rule 1 : If x is A1 and y is B1 then f1=p1x+q1y+r1

                           Rule 2 : If x is A2 and y is B2 then f2=p2x+q2y+r2

       where p, q and r denotes linear parameters in the then-part (consequent part) of the first-order
Sugeno fuzzy model. The architecture of ANFIS consists of five layers (Fig. 1), and a brief
introduction of the model is as follows.
Layer 1.input nodes. Each node of this layer generates membership grades to which they belong to
each of the appropriate fuzzy sets using membership functions.

                                               O = µ ( x)
                                                i   Ai

        where x denotes the crisp inputs to node i, and Ai, (small, large, etc.) represents the linguistic
labels characterized by appropriate membership functions        Due to softness and concise notation,

International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976 –
6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 7, November – December (2013), © IAEME

the Gaussian and bell-shaped membership functions are very popular for specifying fuzzy sets. If we
choose        to be bell shaped with maximum equal to 1 and minimum equal to 0 then

                                           µ A ( x) =            1
                                              i             x −c 2bi 
                                                        1+       i
                                                                      
                                                            ai  
                                                                       

where {ai,bi, ci} represent the parameter set. As the values of these parameter change, the bell shaped
function vary accordingly, thus exhibiting various forms of membership function on the linguistic
label Ai .

Layer 2: Every node in this layer gets multiplied with the incoming signal and gives the product
                                  = W = µ ( x ) × µ ( y ),
                                     i   A         B
                                                             i = 1, 2..                    (3)
                                                  i        i

Each node represents the firing strength of a rule.

Layer 3: The ith node calculates the ratio of the ith rule’s firing strength to the sum of all rule’s firing
                                             = wi =         ,        i = 1, 2                         (4)
                                                    w1 + w2

For convenience, the outputs of this layer are called normalized firing strengths.

Layer 4: The node function of the fourth layer computes the contribution of each ith rule toward the
total output and the function is defined as:

                                        O    = wi f = w ( pi x + qi y + ri )                          (5)
                                         4,i       i   i

        where      denotes the ith node’s output from the previous layer. As for {pi,qi, ri}, they
represent the coefficients of this linear combination and are also the parameter set in the consequent
part of the Sugeno fuzzy model.

             Fig1. ANFIS architecture for two-input Sugeno fuzzy model with four rules

International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976 –
6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 7, November – December (2013), © IAEME

Layer 5: output nodes. The single node computes the overall output by summing all the incoming
signals. Accordingly, the defuzzification process transforms each fuzzy rule resulting into a crisp
output in this layer:
                                                               ∑ w f
                                               O    = ∑ wi fi = i i i                            (6)
                                                5,i             ∑i wi

        This network is trained based on supervised learning. So our objective is to train adaptive
networks to be able to estimate unknown functions given by training data and then to find the
specific values of the above parameters.
        The distinguishing feature of the approach is that ANFIS applies a hybrid-learning algorithm,
the gradient descent method and the least-squares method, to update parameters. The gradient
descent method is used to tune premise non-linear parameters ({ai,bi, ci}), while the least-squares
method is employed to identify subsequent linear parameters ({pi,qi, ri}). As seen in Fig. 1, the
circular nodes are fixed (i.e., not adaptive) nodes without parameter variables, and the square nodes
have parameter variables (the parameters are changed during training). The task of the learning
procedure has two steps: In the first step, the least square method is used to recognize the consequent
parameters, while the antecedent parameters (membership functions) are assumed to be fixed for the
current cycle through the training set. Then, the error signals are propagated backward. Gradient
descent method is employed to update the premise parameters, through minimizing the overall
quadratic cost function, while the consequent parameters remain fixed. The detailed algorithm and
mathematical background of the hybrid-learning algorithm can be found in [12].

2.2    Non linear model
       Dynamic models have complex functional dependence between the system's inputs u(t) and
outputs y(t). We can use these relationships to compute the current output from previous inputs and
outputs. The general form of a model in discrete time is:

               y(t) = f(u(t - 1), y(t - 1), u(t - 2), y(t - 2), . . .)                              (7)

        Such a model is nonlinear if the function f is a nonlinear function, which might include
nonlinear components representing arbitrary nonlinearities. Nonlinear models might be necessary to
represent systems that operate over a range of operating points. In some cases, We might fit several
linear models, where each model is accurate at specific operating conditions.
        If We know the nonlinear equations describing a system, We can represent this system as a
nonlinear grey-box model and estimate the coefficients from experimental data. In this case, the
coefficients are the parameters of the model.

This block diagram shows the structure of a nonlinear ARX model:

                    Fig 2. Structure of Nonlinear Auto Regressive (ARX) Models
International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976 –
6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 7, November – December (2013), © IAEME

The nonlinear ARX model computes the output y in two stages:

   1. It computes regressors from the current and past input values and past output data.
      In the simplest case, regressors are delayed inputs and outputs, such as u(t-1) and y(t-3)—
      called standard regressors. Custom regressors can also be specified, which are nonlinear
      functions of delayed inputs and outputs. For example, tan(u(t-1)) or u(t-1)*y(t-3). By default,
      all regressors are inputs to both the linear and the nonlinear function blocks of the
      nonlinearity estimator. We can select a subset of regressors as inputs to the nonlinear function

   2. The nonlinearity estimator block maps the regressors to the model output using a
      combination of nonlinear and linear functions. One can select from available nonlinearity
      estimators, such as tree-partition networks, wavelet networks, and multi-layer neural
      networks. One can also exclude either the linear or the nonlinear function block from the
      nonlinearity estimator.

The nonlinearity estimator block can include linear and nonlinear blocks in parallel. For instance:

                                  F (x) = LT ( x − r ) + d + g ( x − r )                          (8)

x denotes a vector of the regressors. LT ( x) + d is the output of the linear function block and is affine
when d≠0. d is a scalar offset. g (Q(x − r)) represents the output of the nonlinear function block. r
denotes the mean of the regressors x. Q is a projection matrix that makes the calculations well
conditioned. The exact form of F(x) depends on our choice of the nonlinearity estimator. Estimating
a nonlinear ARX model computes the model parameter values, such as L, r, d, Q, and other
parameters specifying g.
        Most nonlinearity estimators represent the nonlinear function as a sum of series of nonlinear
units, such as wavelet networks or sigmoid functions. One can configure the number of nonlinear
units n for estimation.

       A sigmoidnet define a nonlinear function y=F(x) where y is scalar and x is an m-dimensional
row vector. The sigmoid network function is based on the following expansion:

                 F(x) =(x−r)PL+aa f ((x−r)Q 1+c1)) f ((x−r)Q n +cn))+d
                                           b                      b                                (9)

where f is the sigmoid function, given by the following equation:

                                                  f (z) =      −z
                                                             e +1

P and Q are m-by-p and m-by-q projection matrices. The projection matrices P and Q are determined
by principal component analysis of estimation data. If the components of x in the estimation data are
linearly dependent, then p<m. The number of columns of Q, q, corresponds to the number of
components of x used in the sigmoid function.

International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976 –
6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 7, November – December (2013), © IAEME

       A wavenet defines a nonlinear function y=f(x), where y is scalar and x is an m-dimensional
row vector. The wavelet network function is based on the following function expansion:

          F (x) = (x − r) PL + a s −1 f (b s _ 1 ((x − r) Q − c s _ 1 )) + .. + a w _ 1 g (b w _ 1 ((x − rQ − c w _ 1 )))

f is a scaling function.
     • g is the wavelet function.
     • P and Q are m-by-p and m-by-q projection matrices, respectively.
         The projection matrices P and Q are determined by principal component analysis of
         estimation data. If the components of x in the estimation data are linearly dependent, then
         p<m. The number of columns of Q, q, corresponds to the number of components of x used in
         the scaling and wavelet function.
         r is a 1-by-m vector and represents the mean value of the regressor vector computed from
         estimation data.
     • as, bs, aw, and bw are scalars. Parameters with the s subscript are scaling parameters, and
         parameters with the w subscript are wavelet parameters.
     • L is a p-by-1 vector.
     • cs and cw are 1-by-q vectors.


       142 years of rainfall data set (1871-2012) was obtained from the Indian Institute of Tropical
Meteorology website (, with the original
source as referred by the department being the Indian Meteorological Department (IMD), is used for
the analysis. Linear transformation is used to normalize rainfall series data. The outputs of the
normalization function are real numbers between 0 and 1. The equation can be described as:

                                                                         Do − Dmin
                                                                    =                                                              (11)
                                                                        Dmax − Dmin

       where D0 is for the observed data, Dmax and Dmin denote maximum and minimum of the
observed data, respectively.


Different prediction models have been designed such as

   •     Adaptive neuro fuzzy model,
   •     Non linear auto regressive model
   •     Linear regressive model.

        A grid based ANFIS model has been developed with a view to predict the rainfall for the next
year based on rainfall pattern for last four year data. Model has been designed and tested with
different membership functions and different number of members. It is found that Gaussian
membership function delivers better performance.
        A linear regressive model has been designed for prediction with different number of delays
and order. To find the best value, performance parameters of the model were evaluated with different
delay and order. It is found that delay of 3 and order of 2 was found to be the best. The rainfall
International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976 –
6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 7, November – December (2013), © IAEME

pattern has a nonlinear relationship therefore a nonlinear auto regressive model with various
nonlinearity estimator such as sigmoidnet and wavenet was designed with different parameters and
was tested with different order and delay.

                         Nonlinearity                                    structure
                                              g (x) = ∑ ak K ( β k ( x − γ k ) ) where K(s) is
                                                       k =1
                      Wavelet network
                                              the wavelet function and k s the number of
                                              g (x) = ∑ ak K ( β k ( x − γ k ) ) where
                      sigmoid network                   k =1
                                              K(s)=(es+1) -1 is the sigmoid function


To assess the models’ performance, following criteria are used as given below.

        Correlation coefficient (CC): It indicates the strength of relationships between observed and
estimated rainfall. The correlation coefficient is a number between 0 and 1, and the higher the
correlation coefficient the better is the accuracy of the model.

                                              (                     )(
                                           ∑ xact (i ) − x act x pre (i ) − x pre           )
                             CC =         i =1
                                        N                    2N                   2                            (12)
                                          (                     )         (
                                        ∑ xact (i ) − x act ∑ x pre (i ) − x pre                )
                                       i =1                   i =1

            Where xact (i) denotes actual rainfall and x pre (i ) denotes predicted rainfall value at point i.

            xact and x pre denotes the mean value of actual and predicted rainfall series. N denotes the
sample size.
        Root mean square error (RMSE): It evaluates the residual between observed and forecasted
rainfall. This index assumes that larger forecast errors are of greater importance than smaller ones;
hence they are given a more than proportionate penalty

                                                   1 N                         2
                                                            (               )
                                                      ∑ xact ( i )− x pre (i )
                                                   N i =1
                                   RMSE =
                                                  ( Max( x pre )−Min( x pre ))                                 (13)

Max x pre   ) indicates maximum value of predicted rainfall ,                     (
                                                                               Min x pre   ) denotes minimum value of
predicted rainfall It is noticed that RMSE equal to zero represents a perfect fit.

International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976 –
6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 7, November – December (2013), © IAEME


        During the development of the prediction model, various configurations were explored in
order to achieve enhanced performance parameter value. ANFIS model was tested for different
membership function such as Gaussian, trapezoidal, etc and different number of members. Gaussian
membership function with three members is found to be optimum. Linear model with different
number of delay and order were designed and tested. Nonlinear model was designed using different
type of non-linear estimating function. Table 1.1 shows the performance parameter for the model
designed .It is observed that ANFIS model has better prediction capability due to combined power of
fuzzy logic and neural network. But the execution time taken is more. As rainfall process is nonlinear
in nature, It is found that nonlinear model has good predicting capabilities. Nonlinear model is found
to be better as compare to linear model. As the order of the model is increased it become more
complex and hence execution time increases. In nonlinear model sigmoidnet estimator is superior in
making prediction as compared to wavelet estimator. The graph of observed rainfall and predicted
rainfall by different model is shown in the fig 3.

       Table 1.1 The performance of parameters of proposed models during different phases
                                                                     Training period          Testing period
       Model type                          parameter                        Correlation       Correlation
                                                               NRMSE                    NRMSE
                                                                            Coefficient       Coefficient
                                        Member function
      ANFIS model                       ‘Gaussian’, three       0.0011         0.998      0.0046        0.95
     Linear regressive                  Delay=2,order=3         0.0080         0.94       0.0088        0.84
        Nonlinear                       Wavenet estimator
                                                                0.0054         0.96       0.0075        0.86
     regressive model                   Delay=2,order=3
        Nonlinear                     Sigmoidnet estimator
                                                                0.0033         0.97       0.0049        0.93
     regressive model                   Delay =2,order=3

                                                               Obserevd RF
                                                               ANFIS Model
                                0.9                            Linear Model
              A ra e R in ll
               ve g a fa



                                 1997    1999    2001       2003   2005      2007      2009   2011

International Journal of Advanced Research in Engineering and Technology (IJARET), ISSN 0976 –
6480(Print), ISSN 0976 – 6499(Online) Volume 4, Issue 7, November – December (2013), © IAEME


 [1]    Sulochana Gadgil and Siddhartha Gadgil, The Indian Monsoon, GDP and agriculture, Economic and
        Political Weekly, XLI, 2006,pp. 4887–4895.
 [2]    Rajeevan M., “Prediction of Indian summer monsoon: Status, problems and prospects”. Curr. Sci.,
        2001, 81, pp.1451–1457.
 [3]    Thapliyal, V. and Rajeevan, M., “Monsoon prediction”. Encyclopedia of Atmospheric Sciences
        (ed. Holton, J.), Academic Press, New York, 2003, pp. 1391–1400.
 [4]    Sikka, D.R., Some aspects of the large scale fluctuations of summer monsoon rainfall over India in
        relation to fluctuations in the planetary and regional scale circulation parameters,
        Proc. Ind.Acad.Sciences (Earth and Planet Sci), 89, 1980,179-195.
 [5]    F. W. Zwiers, J. V. Storch,”On the role of statistics in climate research,” Int. J. Climatology, 2004
        Vol.24, pp. 665-680.
 [6]    Shamseldin AY. Application of a neural network technique to rainfall-runoff modeling. J Hydrol
 [7]    Ham FM, Kostanic I. Principles of neurocomputing for science and engineering New York:
        McGraw Hill; 2001.
 [8]    French MN, Krajewski WF, Cuykendall RR. Rainfall forecasting in space and time using a neural
        network. J Hydrol 1992;137:1–31.
 [9]    Grimes DIF, Coppola E, Verdecchia M, Visconti G. A neural network approach to real-time rainfall
        estimation for Africa using satellite data. J Hydrometeorol 2003;4:1119–33
 [10]   Kin C. Luk, J. E. Ball AND A. Sharma, An Application of Artificial Neural Networks for Rainfall
        Forecasting, Mathematical and Computer Modelling 33, 2001,pp. 883-699.
 [11]   Poggio,T., and F. Girosi, Networks for approximation and learning, Proc. IEEE, vol. 78, 1990,
        pp. 1481–1497.
 [12]   Jang JSR, ANFIS: adaptive-network-based fuzzy inference system, IEEE Trans Syst, Man,
 [13]   Takagi T, Sugeno M., “Fuzzy identification of systems and its applications to modeling and control,”
        IEEE Trans Syst, Man, Cybernet;15:1985,116–32.
 [14]   K. Hornik, M. Stinchcombe and H. White, Multilayer feedforward networks are universal
        approximators, Neural Networks 1989 2, pp.359-366.
 [15]   Sven F. Crone: A Business Forecasting Competition Approach to Modeling Artificial Neural
        Networksfor Time Series Prediction. IC-AI 2004, pp. 207-213.
 [16]   M.French, W. Krajewski and R.R. Cuykendall, Rainfall forecasting in Space and time using a neural
        network, Journal of Hydrology 1992 137, 1-31.
 [17]   K.L. Hsu, V. Gupta and S. Soroshian, Artificial neural network modeling of the rainfall-runoff
        process, Water Resources Research l995 31 (10), pp.2517-2530.
 [18]   Werbos, P. J., Backpropagation through time:What it does and how to do it,Proc. IEEE, vol. 78,
        1990,pp. 1550–1560.
 [19]   Rumelhart, D. E., Hinton G. E., and Williams R. J., Learning internal representations by error
        propagation,” in D. E. Rumelhart and J. L. McCleland, eds. (Cambridge, MA: MIT Press), vol. 1,
        Chapter 8.
 [20]   Jiansheng. Wu, Long. Jin, Forecast Research and Applying of BP Neural Network Based on Genetic
        Algorithms, Mathematics in Practice and Theory, 2005, vol. 35(1), pp.83-88.
 [21]   V.Anandhi and Dr.R.Manicka Chezian, “Comparison of the Forecasting Techniques – Arima, Ann
        and SVM - A Review”, International Journal of Computer Engineering & Technology (IJCET),
        Volume 4, Issue 3, 2013, pp. 370 - 376, ISSN Print: 0976 – 6367, ISSN Online: 0976 – 6375.
 [22]   M. Nirmala and S. M. Sundaram, “Modeling and Predicting the Monthly Rainfall in Tamilnadu as a
        Seasonal Multivariate Arima Process”, International Journal of Computer Engineering &
        Technology (IJCET), Volume 1, Issue 1, 2010, pp. 103 - 111, ISSN Print: 0976 – 6367,
        ISSN Online: 0976 – 6375.


Shared By: