An Adaptive Neuro-Fuzzy Inference System based on Vorticity and Divergence for Rainfall forecasting by ijcsiseditor


More Info
									                                                             (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                Vol. 9, No. 12, 2011

      An Adaptive Neuro-Fuzzy Inference System
                        based on
    Vorticity and Divergence for Rainfall forecasting

                                                          Kavita Pabreja
                      Research Scholar, Birla Institute of Technology and Science, Pilani, Rajasthan, India
              Assistant Professor, Maharaja Surajmal Institute (an affiliate of GGSIP University), New Delhi, India

Abstract— A new rainfall forecasting model based on Adaptive             provide information and warning of extreme weather events
Neuro-Fuzzy Inference System is proposed in this paper. A                for minimizing losses both to human and property. Such data
neuro-fuzzy model inherits the interpretability of fuzzy models          consists of a sequence of global snapshots of the Earth,
and learning capability of neural networks in a single system. It        typically available at various spatial and temporal intervals
has got wide acceptance for modelling many real world problems           including atmospheric parameters over land and ocean (such
because it provides a systematic and directed approach for model
building and gives the best possible design parameters in
                                                                         as temperature, pressure, wind speed, wind direction, sea
minimum time. The datasets used in this paper for the training of        surface temperature, etc.). The NWP models do not produce
Adaptive Neuro-Fuzzy Inference System (ANFIS) are the                    forecast of rainfall directly. Forecast of weather elements like
European Center for Medium-range Weather Forecasting                     rain/snow, sky conditions etc. at a place are derived through
(ECMWF) model output products and the gridded rainfall                   statistical relation popularly known as Model Output Statistics
datasets, provided by Indian Meteorological Department (IMD).            (MOS) proposed by National Weather Service [1]. General
To determine the characteristics of ANFIS that best suited the           experience is that MOS products show improved skills over
target rainfall forecasting system, several ANFIS models were            the raw model output. Basis of MOS is statistical relationship
trained, tested and compared. Different training and checking            which requires long term consistent series of NWP products.
data, type and number of membership functions and techniques
                                                                         Since NWP models get upgraded regularly[2], the series does
to generate the initial Fuzzy Inference Systems were analyzed.
Comparisons of the different models were performed and the               not remain consistent.
results showed that the model generated by grid partitioning                  In view of above limitation of MOS, it has been proposed
using gbellmf membership functions provided the smallest errors          to explore other Intelligent techniques like ANFIS so as to
for rainfall forecasting.                                                forecast rainfall based on NWP model output products. In
                                                                         past, Artificial Neural Networks (ANN) has been applied [3]
                                                                         to predict the average rainfall over India during summer-
    Keywords- NWP model forecast, ECMWF model, rainfall,                 monsoon i.e. the months of June, July, and August, by
vorticity, divergence, ANFIS
                                                                         exploring the rainfall data corresponding to the summer
                                                                         monsoon months of years 1871-1999. It has been found that
                      I.    INTRODUCTION                                 the prediction error in case of ANN is 10.2% whereas the
     Weather is not just an environmental issue; it is a major           prediction error in the case of persistence forecast is 18.3%.
economic factor. Economic value of weather for Agriculture,                 A neural network, using input from the Eta Model and
Fishery, Energy, Transportation, Aviation and health area is             upper air soundings, has been developed [4] for the probability
immeasurable. With its huge and growing population and low-              of precipitation (PoP) and quantitative precipitation forecast
lying coastline and an economy that is closely tied to its               (QPF) for the Dallas–Fort Worth, Texas, area. Forecasts from
natural resource base, India is considerably sensitive to                two years were verified against a network of 36 rain gauges.
weather and climate. One failure of monsoon can totally upset            The resulting forecasts were remarkably sharp, with over 70%
the economic performance of our country. But timely                      of the PoP forecasts being less than 5% or greater than 95%.
forecasting can help to considerably minimize the adverse                Of the 436 days with forecasts of less than 5% PoP, no rain
effect.                                                                  occurred on 435 days. Of the 111 days with forecasts of
     Analysis and forecast of weather data created through               greater than 95% PoP, rain always occurred. The application
Numerical Weather Prediction (NWP) models offers an                      of ANFIS for forecasting of meteorological parameters is very
unprecedented opportunity for predicting weather events,                 rare and particularly rainfall forecasting has not been

                                                                                                    ISSN 1947-5500
                                                                 (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                    Vol. 9, No. 12, 2011
considered in the studies and hence has been taken up in this             between 0 and 1. This process is known as fuzzification and
paper to look for even better accuracy of forecast.                       takes place in layer 2, the fuzzification layer. Each node in this
                                                                          layer is adaptive.
  II.   ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM                                  Once the locations of the inputs in the fuzzy spaces are
   ANFIS is a hybrid of two intelligent systems: Artificial               identified, then the product of the degrees to which the inputs
Neural Networks (ANNs) and Fuzzy Inference Systems                        satisfy the membership functions is found. This product is
(FISs). ANNs map an input space to an output space through a              called the firing strength of a rule and is represented by layer
collection of layered processing elements called neurons that
are interconnected in parallel by synaptic junctions. ANNs are
developed by continuously passing real world system data
from its input to output layer. For each pass of data, signals
propagate from the input to output layer to produce an output
which is compared to the desired output. The difference
between these values is then used to adjust the synaptic
connections so that the ANN can mimic the system the data
represents. This procedure gives ANNs the capability of
looking for patterns in the information presented to it, thus
providing it with the advantage of learning about systems.
   FISs are based on fuzzy logic (a continuous range of truth
values from 0 to 1), IF-THEN fuzzy rules and fuzzy reasoning
(which can be likened to human reasoning through linguistic
variables such as small, medium, large). These features of FIS
allow it to make inferences using the rules and known facts to            3, the rule layer where each node in this layer is fixed. Each
derive reasonable decisions [5]. Thus the combination of                  fuzzy space is governed by an ANFIS rule where the
ANNs and FISs to form ANFIS, integrates the benefits of the               antecedent of the rule defines a fuzzy space in the input space
individual intelligent systems to form a superior technique that          [5]. For ANFIS, there are Mn fuzzy rules where M is the
can optimally model the dynamics of difficult systems.                    number of membership functions per input and n is the
   An example of ANFIS has been explained which is of a 6                 number of inputs.
layer feedforward neural network and of the Sugeno FIS type.                   In layer 4, the normalization layer, the ratio of each rule’s
To understand the structure and operation of ANFIS in                     firing strength is calculated with respect to the sum of the
forecasting, a 2 input - 1 output ANFIS model is presented and            firing strengths of all the rules. Each node in this layer is fixed.
its structure and operation is related to a generalized model.                 In layer 5, the defuzzification layer, the output of each
Fig. 1 shows the ANFIS structure and Equations 1 to 4 are the             node is the weighted consequent value. Layer 6 is the
rules for this model where the IF part of the rule is referred to         summation layer and its output which is the sum of all the
as the antecedent and the THEN part is the consequent.                    outputs of the layer 5 which gives the overall output for the
                                                                          respective inputs within the fuzzy space. Before the ANFIS
Rule 1: If x is A1 and y is B1, then f1 = p1x + q1y + r1   (1)            system can be used for prediction, the parameters of the rules
Rule 2: If x is A2 and y is B2, then f2 = p2x + q2y + r2   (2)            are determined by first generating an initial FIS where random
Rule 3: If x is A3 and y is B3, then f3 = p3x + q3y + r3   (3)            values are assigned to the parameters and then applying an
Rule 4: If x is A4 and y is B4, then f4 = p4x + q4y + r4   (4)            optimization scheme to determine the best values of the
                                                                          parameters that would provide rules that would idealistically
     In general, an n-input, 1-output ANFIS model is an n + 1             model the target system. After training, the rules remain so
dimensional input-output space. Therefore, a 2 inputs-1 output            that when new input data is presented to the model, the rules
ANFIS model is a 3-dimensional input-output space. In order               provide a corresponding reasonable output.
for ANFIS to be used to model a system, data that is                          The optimization technique is a learning algorithm which
representative of the target system must be presented to                  uses data (training data) from the target system to generate
ANFIS. The entry of raw data or crisp inputs from the target              signals that propagate backwards and forwards and update the
system into ANFIS corresponds to layer 1 – the input layer in             parameters by a process known as training. The learning
Fig. 1.                                                                   algorithm proposed for ANFIS is a hybrid learning algorithm
     Since the Neural Network classifies data and looks for               that minimizes the error between the ANFIS model and the
patterns within it, then when the input data is in the 3-                 real system [5]. ANFIS employs the least squares estimate and
dimensional space, it is classified into groups called fuzzy              the gradient descent method in the hybrid learning algorithm.
spaces. To do this, the crisp inputs are compared with                    Once input-output data is presented to ANFIS, in one epoch
membership functions in the antecedent of the rules of ANFIS,             the data is propagated forwards from one layer to the next
to determine the degree to which the inputs, in this case, X1             until the fourth layer, and the least squares estimate is
and X2 belong to fuzzy sets Ai and Bi respectively. The degree            employed to update the linear or consequent parameters. An
to which the inputs lie within the fuzzy space is given a value           error is calculated and this is propagated backwards and the

                                                                                                       ISSN 1947-5500
                                                             (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                Vol. 9, No. 12, 2011
gradient descent is used to update the non-linear or premise
parameters [5].

                     III.    DATASETS USED
Two different datasets provided by IMD have been used for
the study. First one is the forecasts by NWP model (ECMWF
model) and the other is the observed values of rainfall
datasets. These dataset files and pre-processing applied on
them are explained in section a and b respectively.
(a) ECMWF T-799 model forecasts and its pre-processing

     The datasets produced as forecast by ECMWF model are
in GRIB format which is a mathematically concise data format
commonly used in meteorology to store historical and forecast
weather data. It is standardized by the World Meteorological
Organization's Commission for Basic Systems. The forecast
datasets of T-799 model includes values for 87variables
(including all atmospheric pressure levels), for latitude -10° to
50° and longitude 50° to 110° at a grid spacing of 0.25°,
making it equal to 241X241 grid points i.e. forecast of 87
variables at 58081 grid points. Finally it becomes a huge
datasets of 50,53,047 (approx. 5million) values for just one
forecast of a particular time.
     The GRIB files have been converted to (.csv) format by
using National Digital Forecast Database - NDFD GRIB2
decoder program of NOAA downloaded from Internet. The
model does not provide vorticity and divergence directly,
which are important determinant of rainfall, so this has been
derived by using vertical (v) and horizontal (u) component of
wind as forecasted by model, using the formulas given below:-
Divergence formula:         ∂u    ∂v
                           ∂x    ∂y                                          Figure 2. Data pre-processing of forecast by ECMWF model

Vorticity formula:      ∂v      ∂u
                       ∂x       ∂y

where     v denotes meridian wind flow
         u denotes zonal wind flow
         x denotes longitude
         y denotes latitude

     These steps of data pre-processing have been shown in
Fig. 2. For the purpose of this study, we have calculated
vorticity and divergence at atmospheric pressure level of
850hPa, on 0000GMT 29 July 09 with initial conditions of
0000GMT 28 July 09 for the model, as input parameters for
training of ANFIS. A small sample of this datasets has been
shown in table I.

                                                                                                   ISSN 1947-5500
                                                                   (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                      Vol. 9, No. 12, 2011
     ON 0000GMT 28JULY 09 VALID FOR 0000GMT 29JULY09
                                                                                             Latitude          Longitude
      Latitude   Longitude   Vorticity          Divergence                                       (°N)               (°E)    Rainfall in mm
          (°N)        (°E)   (X10-5 per sec )   (X10-5 per sec )
                                                                                                  24.5              94.5              14.1

          24.5        94.5                  2                 -6                                    28              94.5              22.7

            28        94.5                  2               -14                                     24              88.5               6.1

            24        88.5                  4                 -4                                  26.5               89               56.6

          26.5          89                  4                 -4                                    30              78.5              16.4

            30        78.5                  6                 -8                                  26.5               90                8.5

          26.5          90                  6               -18                                   21.5               84                1.2

          21.5          84                  8                 -6                                    26              92.5               5.4

            26        92.5                  8               -14                                   27.5               94               12.8

          27.5          94                  8                 -8                                  27.5              81.5              24.1

          27.5        81.5                10                  -4                                    25              85.5              18.4

            25        85.5                10                -10                                   22.5               86               10.4

          22.5          86                10                -10                                   28.5              80.5              40.7

          28.5        80.5                14                  -8                                  27.5               84                27

          27.5          84                16                -10                                   26.5               86                1.2

          26.5          86                16                -10                                     29              79.5               2.5

            29        79.5                20                -14                                     23              92.5               42
                                                                                (SOURCE: AS A RESULT OF PRE-PROCESSING RF2009.GRD PROVIDED BY IMD)
            23        92.5                20                -10

(b) Rainfall datasets and its pre-processing                                     Finally the two different datasets of model forecast and
     A high resolution (0.5° × 0.5°) daily rainfall (in mm)                 rainfall datasets location-wise have been merged, as shown in
dataset for mesoscale meteorological studies over the Indian                table IV using the Rainfall category as explained in table III,
region has been provided by IMD and described by [6]. The                   so that they can be presented to ANFIS model for training and
dataset is in .grd format, a control file describing the structure          obtaining rules that correlate the vorticity and divergence as
of .grd file provided by IMD.                                               antecedents with rainfall category as consequent.
    The rainfall datasets under study are for year 2009. The
                                                                                    TABLE III - CATEGORY AND CODE FOR RAINFALL
data is for the geographical region from longitude 66.5 ºE to                           CORRESPONDING TO RAINFALL (IN MM)
100.5 ºE and latitude 6.5 ºN to 38.5 ºN for each day of the
year. There are 4485 grid points readings every day and                               Rainfall value (in mm)               Category          Code
rainfall record for 122 days (June to September) per year are                         1-15                                 very low          1
selected for analysis i.e 5,47,170 records out of a total of
                                                                                      15.1 – 40                            low               2
16,37,025 records for one year of rainfall. Steps followed for
pre-processing of the .grd so that an intelligent system can be                       40.1-75                              good              3
applied, are mentioned below:                                                         75.5 - more                          heavy             4
     1. The .grd file has been converted to .dat file using a
          FORTRAN programme. This dataset is very huge in
     2. The .txt files have been exported to Excel worksheet
          and then to Access database. The data looks like as if
          a rectangular grid is filled with values of rainfall in
     3. Using a Visual Basic program to organize data in
          tabular format, as shown in table II.
     4. Finally exporting the dataset into .xls format for
          analysis, by Matlab.

                                                                                                           ISSN 1947-5500
                                                                             (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                                Vol. 9, No. 12, 2011
                                                                                      avoid curse of dimensionality problem [7]. Therefore we have
                                                                                      opted for Grid partitioning as we have just 2 input parameters
                  RAINFALL CATEGORY                                                   viz. vorticity and divergence.
                                                                                           Once the grid partitioning technique is applied at the
                                                                                      beginning of training, a uniformly partitioned grid which is
   Vorticity                 Divergence
   (X10-5 per second )       (X10-5 per second )         Rainfall code                defined by membership functions (MFs) with a random set of
                                                                                      parameters is taken as the initial state of ANFIS. During
                                                                                      training, this grid evolves as the parameters in the MFs
                         2                          -6                   1
                                                                                      change. With the grid partitioning technique, the number of
                         4                         -10                   1            MFs in the premise part of the rules must be determined.
                         6                         -18                   1            Negnevitsky et al. [8] stated that a larger number of MFs
                                                                                      better represents a complex system and therefore should
                         8                         -14                   1
                                                                                      produce better results. However, a large number of inputs or
                    18                             -10                   1            MFs in the premise part of the rules can produce a large
                    20                             -14                   1            number of fuzzy rules which can cause the learning
                                                                                      complexity of ANFIS to suffer an exponential explosion,
                    20                              -4                   1
                                                                                      called the curse of dimensionality which can adversely affect
                         2                         -18                   2            the performance of ANFIS [7, 9, 10].
                         8                          -8                   2                 We have generated 5 different ANFIS models by grid-
                                                                                      partitioning. The idea was to explore the ANFIS generation
                    10                             -10                   2
                                                                                      first with different shapes of membership functions, keeping
                    10                              -4                   2            the dataset for training, checking fixed at original values (i.e.
                    16                             -10                   2            no normalization done) and number of membership functions
                                                                                      fixed. Next it was decided to explore the ANFIS generation
                    18                              -8                   2            with increase in membership functions. Finally, it was decided
                    26                             -18                   2            to normalize datasets [-1 1] for training and checking and
                    44                              -8                   2
                                                                                      observing the FIS outputs after training. All these parameters
                                                                                      for different models are explained in table V.
                         4                          -4                   3                The number of MFs was increased with one of the ANFIS
                    14                              -8                   3            models (number 1 in table V) to get a greater understanding of
                                                                                      the impact on the performance of ANFIS with this change. In
                    16                              -4                   3
                                                                                      generating the Rainfall forecasting FIS, by grid partitioning, the
                    20                             -10                   3            bell-shaped MF was favored over the other types since it
                     -8                             4                    4
                                                                                      offered more parameters which provided a greater number of
                                                                                      degrees of freedom. The generalized bell-shaped MF is
                                                                                      standard for ANFIS because of its smoothness and concise
                 IV. GENERATION OF ANFIS                                              notation [7, 8, 9]. Other function such as Gaussian was used as
                                                                                      well to evaluate the performance with different types of MFs.
                                                                                      For the consequent part of the rules the MFs responsible for
     ANFIS in this study was trained and simulated using                              defuzzification were the Sugeno type of first order. The output
Matlab 7.0 (matrix laboratory) designed and developed by                              MF is chosen to be linear for the rainfall forecasting models
Math Works Inc. The fuzzy inference commonly used in                                  since, the higher the order of output MFs, the greater is the
ANFIS is first order Sugeno fuzzy model because of its                                likelihood of ANFIS fitting the target system [11].
simplicity, high interpretability, and computational efficiency,
built- in optimal and adaptive techniques. A typical                                         TABLE V - DIFFERENT ANFIS MODELS USED IN THE STUDY
architecture of an ANFIS has already been shown in Fig. 1.                             ANFIS       Type of membership function      Number      of   Number    of
Among many FIS models, the Sugeno fuzzy model is the most
                                                                                       Model       for input parameters             membership       membership
widely applied one for its high interpretability and
computational efficiency, and built-in optimal and adaptive                                                                         functions for    functions for
techniques.                                                                                                                         Vorticity        Divergence
     Generation of ANFIS involves selecting a structure for                            I           gbellmf (original data)          5                3
the ANFIS model by determining the number of membership
functions per input, type/shape of the membership functions                            II          gaussmf (original data)          5                3
for the premise part of the rule and the output membership                             III         gbellmf (original data)          7                5
functions for the consequent part of the rule. MATLAB 7.0
offers two methods for generating the initial FIS: Grid                                IV          gaussmf (normalized data)        5                3
Partitioning and Subtractive Clustering. Subtractive                                   V           gbellmf (normalized data)        5                3
partitioning is used if number of inputs is more than 6 so as to

                                                                                                                      ISSN 1947-5500
                                                                     (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                        Vol. 9, No. 12, 2011
                  V TRAINING THE ANFIS
          Training involves the selection of optimization
technique, error tolerance and the number of epochs. The
ANFIS toolbox provided two optimization methods: hybrid
and backpropagation. To develop the ANFIS rainfall
forecasting models, the hybrid technique was used since it is
more popularly used with ANFIS than the backpropagation
because it is a combination of least-squares and back-
propagation gradient descent method [5,8]. In addition, it is
regarded as the faster of the two techniques [5]. We have
trained five models with 80% of the datasets for training and
                                                                                 Figure 4 target output (in red) and ANFIS predicted output (in blue) for the
20% as checking data. The number of epochs has been                              2input (5 bell MF for vorticity input, 3 bell MF for divergence input) , 1
changed from 50 to 100 depending on the shape of training                        output ANFIS model
and checking curves. In accordance to the approach provided
by J.S.R. Jang [5], different models were created by changing
                                                                                 2. The checking error curves for the model with 5 gaussian
some part of its structure or parameters, and each was
                                                                                 MFs for vorticity input and 3 gaussian MFs for divergence
compared to the previous models created to determine if the
                                                                                 input, remains almost constant from the first epoch till last
changed characteristic provided better results. If the model
                                                                                 epoch, as shown in Fig. 5.
produced better results, then these characteristics were kept
and if not, the model was retrained with one of the
characteristics of its structure changed. After which, one
feature of the chosen model: type of input data, size of training
or checking data, type of membership functions or the number
of membership functions per input was changed one at a time.
The chosen structures were trained with datasets mentioned in
section 3, once trained they were evaluated using the
performance metrics: RMSE.

                         VI RESULTS
          The findings from these five models trained using the
grid partitioning technique provided following important                         Figure 5 Training and checking error curves for the 2input (5 gaussian MF for
results:-                                                                        vorticity input, 3 gaussian MF for divergence input) , 1 output ANFIS model
1. The checking error curves for the model with 5 bell-shaped
MFs for vorticity input and 3 bell-shaped MFs for divergence                     This model when compared for the actual checking data verses
input, decreases from the first epoch. This was trained for                      output generated by ANFIS model demonstrated very poor
50epochs which resulted in almost same value for training and                    results as shown in Fig. 6.
testing error, as shown in Fig. 3.

                                                                                 Figure 6 target output (in red) and ANFIS predicted output (in blue) for the
                                                                                 2input (5 gaussian MF for vorticity input, 3 gaussian MF for divergence input)
Figure 3 Training and checking error curves for the 2input (5 bell MF for        , 1 output ANFIS model
vorticity input, 3bell MF for divergence input) , 1 output ANFIS model
                                                                                 3.The testing/checking error curves for the model with 7 bell-
This model when compared for the actual checking data verses                     shaped MFs for vorticity input and 5 bell-shaped MFs for
output generated by ANFIS model demonstrated good results                        divergence input, decreases from the first epoch. This was
as shown in Fig. 4.                                                              trained for 50epochs after which model started overfitting, as
                                                                                 shown in Fig. 7.

                                                                                                                  ISSN 1947-5500
                                                                         (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                            Vol. 9, No. 12, 2011

Figure 7 Training and checking error curves for the 2input (7 bell MF for            Figure 10 target output (in red) and ANFIS predicted output (in blue) for the
vorticity input, 5 bell MF for divergence input) , 1 output ANFIS model              2input (5 gaussian MF for vorticity input, 3 gaussian MF for divergence
                                                                                     input), 1 output ANFIS model with normalized data

This model when compared for the actual checking data verses
output generated by ANFIS model demonstrated results as                              5. With the normalized datasets, again a new ANFIS model
shown in Fig. 8.                                                                     was generated with 5 bell-shaped MFs for vorticity and 3 bell-
                                                                                     shaped MFs for divergence in order to get better results for
                                                                                     checking error. This model was trained for 100epochs after
                                                                                     which model became stable, as shown in Fig. 11.

Figure 8 target output (in red) and ANFIS predicted output (in blue) for the
2input (7 bell MF for vorticity input, 5 bell MF for divergence input) , 1
output ANFIS model

4. It was experimented to train the model with Gaussian                              Figure 11 Training and checking error curves for the 2input (5 bell MF for
membership functions for representation of the inputs but the                        vorticity input, 3 bell MF for divergence input) , 1 output ANFIS model with
response of the model was very poor so the datasets for input –                      normalized data
output were normalized so that they fall in the range [-1 1].
With these datasets, it was observed that the testing/checking
                                                                                     This model when compared for the actual checking data verses
error curves with 5 gaussian MFs for vorticity input and 3
                                                                                     output generated by ANFIS model did not produce better
gaussian MFs for divergence input, trained on 80% of rainfall                        results than when trained with actual original datasets, as
data produced good results. This was trained for 50epochs                            shown in Fig. 12.
after which model started overfitting, as shown in Fig. 9.

                                                                                     Figure 12 target output (in red) and ANFIS predicted output (in blue) for the
Figure 9 Training and checking error curves for the 2input (5 gaussian MF for        2input (5 bell MF for vorticity input, 3 bell MF for divergence input) , 1
vorticity input, 3 gaussian MF for divergence input) , 1 output ANFIS model          output ANFIS model with normalized data
with normalized data                                                                 The values of root mean square errors for training and
This model when compared for the actual checking data verses                         checking datasets for all these five ANFIS models are
output generated by ANFIS model demonstrated the results as                          tabulated in table VI.
shown in Fig. 10.

                                                                                                                      ISSN 1947-5500
                                                                        (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                           Vol. 9, No. 12, 2011
TABLE VI - ROOT MEAN SQUARE ERROR FOR THE ANFIS MODELS                               [4]       Hall T., Brooks H.E., Doswell C.A. Precipitation Forecasting
               FOR RAINFALL FORECASTING                                                        Using a Neural Network, Weather and Forecasting. 1999, 14 :
                                                                                     [5]       Jang J.S.R., , ANFIS: adaptive network-based fuzzy inference
       Model                      RMSE                                                         systems, IEEE Transactions on Systems, Man and Cybernetics,
                                                                                               May/June 1993, vol. 23, no. 3, pp. 665 – 685.
                      Training         Checking                                      [6]       Rajeevan M., Bhate J. A high resolution daily gridded rainfall
                                                                                               dataset (1971–2005) for mesoscale meteorological studies, Current
       I              0.6953           0.6973                                                  Science, vol. 96, no. 4, 25, 2009 Feb.
       II             0.7012           1.0291                                        [7]       MATHWORKS, Fuzzy Logic Toolbox – anfis and the ANFIS
       III            0.6455           0.9708                                                  Editor GUI, MATLAB 7.0.1.
       IV             0.4585           1.4004                                        [8]       M. Negnevitsky, C. W. Potter and M. Ringrose, Short Term Wind
       V              0.4494           5.6856                                                  Forecasting Techniques for Power Generation, in Australasian
                                                                                               Universities Power Engineering Conference, September 2004.
                                                                                     [9]       J.S.R. Jang, C.T. Sun and E. Mizutani, Neuro-Fuzzy and Soft
                                                                                               Computing, A Computational Approach to Learning and Machine
                                                                                               Intelligence, New Jersey: Prentice Hall, 1997, pp. 73,74, 86, 95-
                      VI CONCLUSION                                                            97,86-87, 26-28, 74-85.
Of all the five models that have been trained and tested with                        [10]      C. M. Bishop, Neural Networks for Pattern Recognition, Oxford:
different number, shape of membership functions for input                                      Oxford University Press, 1995, pp. 5-10.
parameters and with actual data and normalized data, it has                          [11]      J. Abonyi, R. Babuska and F. Szeifert, Fuzzy Modeling with
                                                                                               Multivariate Membership Functions:Gray Box Identification and
been concluded that the bell shaped membership function is                                     Control Design, IEEE Transactions on Systems, Man, and
the best to map rules for relating input values of vorticity and                               Cybernetics –Part B: Cybernetics, vol. 31, no.5, October 2001, pp.
divergence to the output value of rainfall category. Also even                                 755 – 767.
if we increase number of membership functions or normalize
the antecedents and consequent data variables, it does not                                                       AUTHORS PROFILE
cause any improvement in the RMSE and hence predicting the                                      Ms. Kavita holds more than 17 years of experience with
                                                                                     Educational institution and Industry. She is currently Assistant Professor -
value of rainfall. The rules diagram of the best ANFIS model                         Computer Society, Maharaja Surajmal Institute, an affiliate of GGS
has been shown in Fig. 13.                                                           Indraprastha University. She has teaching experience of over a decade and she
                                                                                     has worked for more than 5 years with Indian as well as USA MNC. These
                  ACKNOWLEDGEMENT                                                    companies include Rockwell International Overseas Corp., Parekh
                                                                                     Microelectronics (I) Ltd., HCL Hewlett Packard Ltd. and Shyam Telecom
This study is based on the datasets made available by courtesy                       Ltd.
of Indian Meteorological Department, India. The author                                          She is M.S.(Software Systems) from BITS, Pilani; AMIETE (eq.
would also like to deeply acknowledge the support and                                B.E. (Electronics and Telecommunication Engg.)) from IETE. She holds
guidance of Dr. Rattan K. Datta, Former Advisor – Deptt. of                          membership of many professional bodies viz. Senior Member of Computer
                                                                                     Society of India, Member of Institute of Electronics and Telecommunication
Science & Technology, Former President - Indian                                      Engineers, Member of Indian Meteorological Society and Member of IACSIT,
Meteorological Society and Computer Society of India.                                Singapore.
                                                                                                She has designed and developed Workbooks and textbooks for the
                           REFERENCES                                                ICT Project, Punjab undertaken by Educational Consultants India Ltd. She
[1]      Hughes H. Model output statistics forecast guidance. United States          has contributed fifteen papers in Journals / Book/ International
         Air Force Environmental Technical Applications Center. pp. 1–16.            conferences. Her paper “Mapping of spatio-temporal relational databases onto
[2]      Uppala S., Dee D., Kobayashi S. Simmons A. Evolution of                     a multidimensional data hypercube” presented at Einblick – Research Paper
         reanalysis at ECMWF, Proceedings of the Workshop by World                   Competition held during Confluence 2010 organized by Amity University in
         Climate Research Programme, France, 2008                                    association with EMC data storage systems (India) Pvt. Ltd. on January 22-
[3]      Chattopadhyay S., Chattopadhyay M., A soft computing technique              23, 2010 was selected as the Best paper and awarded the FIRST prize.
         in rainfall forecasting, Proceedings of the International conference
         on IT, HIT, March 2007, 523-526

                                                                                                                      ISSN 1947-5500
                                                 (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                    Vol. 9, No. 12, 2011

Figure 13 Graphical illustration of a set of rules and their contribution to the final results in case of Model I

                                                                                               ISSN 1947-5500

To top